11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 6674ae819SStefano Zampini 7a13144ffSStefano Zampini /* returns B s.t. range(B) _|_ range(A) */ 8a13144ffSStefano Zampini #undef __FUNCT__ 9a13144ffSStefano Zampini #define __FUNCT__ "MatDense_OrthogonalComplement" 10a13144ffSStefano Zampini PetscErrorCode MatDense_OrthogonalComplement(Mat A, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B) 11a13144ffSStefano Zampini { 12a13144ffSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 13a13144ffSStefano Zampini PetscScalar *uwork,*data,*U, ds = 0.; 14a13144ffSStefano Zampini PetscReal *sing; 15a13144ffSStefano Zampini PetscBLASInt bM,bN,lwork,lierr,di = 1; 16a13144ffSStefano Zampini PetscInt ulw,i,nr,nc,n; 17a13144ffSStefano Zampini PetscErrorCode ierr; 18a13144ffSStefano Zampini 19a13144ffSStefano Zampini PetscFunctionBegin; 20a13144ffSStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 21a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"LAPACK _GESVD not available"); 22a13144ffSStefano Zampini #endif 23a13144ffSStefano Zampini ierr = MatGetSize(A,&nr,&nc);CHKERRQ(ierr); 24a13144ffSStefano Zampini if (!nr || !nc) PetscFunctionReturn(0); 25a13144ffSStefano Zampini 26a13144ffSStefano Zampini /* workspace */ 27a13144ffSStefano Zampini if (!work) { 28a13144ffSStefano Zampini ulw = PetscMax(PetscMax(1,5*PetscMin(nr,nc)),3*PetscMin(nr,nc)+PetscMax(nr,nc)); 29f913dca9SStefano Zampini ierr = PetscMalloc1(ulw,&uwork);CHKERRQ(ierr); 30a13144ffSStefano Zampini } else { 31a13144ffSStefano Zampini ulw = lw; 32a13144ffSStefano Zampini uwork = work; 33a13144ffSStefano Zampini } 34a13144ffSStefano Zampini n = PetscMin(nr,nc); 35a13144ffSStefano Zampini if (!rwork) { 36a13144ffSStefano Zampini ierr = PetscMalloc1(n,&sing);CHKERRQ(ierr); 37a13144ffSStefano Zampini } else { 38a13144ffSStefano Zampini sing = rwork; 39a13144ffSStefano Zampini } 40a13144ffSStefano Zampini 41a13144ffSStefano Zampini /* SVD */ 42a13144ffSStefano Zampini ierr = PetscMalloc1(nr*nr,&U);CHKERRQ(ierr); 43a13144ffSStefano Zampini ierr = PetscBLASIntCast(nr,&bM);CHKERRQ(ierr); 44a13144ffSStefano Zampini ierr = PetscBLASIntCast(nc,&bN);CHKERRQ(ierr); 45a13144ffSStefano Zampini ierr = PetscBLASIntCast(ulw,&lwork);CHKERRQ(ierr); 46a13144ffSStefano Zampini ierr = MatDenseGetArray(A,&data);CHKERRQ(ierr); 47a13144ffSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 48a13144ffSStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,&lierr)); 49a13144ffSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 50a13144ffSStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 51a13144ffSStefano Zampini ierr = MatDenseRestoreArray(A,&data);CHKERRQ(ierr); 52a13144ffSStefano Zampini for (i=0;i<n;i++) if (sing[i] < PETSC_SMALL) break; 53a13144ffSStefano Zampini if (!rwork) { 54a13144ffSStefano Zampini ierr = PetscFree(sing);CHKERRQ(ierr); 55a13144ffSStefano Zampini } 56a13144ffSStefano Zampini if (!work) { 57a13144ffSStefano Zampini ierr = PetscFree(uwork);CHKERRQ(ierr); 58a13144ffSStefano Zampini } 59a13144ffSStefano Zampini /* create B */ 60a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,nr-i,NULL,B);CHKERRQ(ierr); 61a13144ffSStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 62a13144ffSStefano Zampini ierr = PetscMemcpy(data,U+nr*i,(nr-i)*nr*sizeof(PetscScalar));CHKERRQ(ierr); 63a13144ffSStefano Zampini ierr = MatDenseRestoreArray(*B,&data);CHKERRQ(ierr); 64a13144ffSStefano Zampini ierr = PetscFree(U);CHKERRQ(ierr); 65a13144ffSStefano Zampini #else 66a13144ffSStefano Zampini PetscFunctionBegin; 67a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for complexes"); 68a13144ffSStefano Zampini #endif 69a13144ffSStefano Zampini PetscFunctionReturn(0); 70a13144ffSStefano Zampini } 71a13144ffSStefano Zampini 72a13144ffSStefano Zampini #undef __FUNCT__ 73a13144ffSStefano Zampini #define __FUNCT__ "PCBDDCComputeNedelecChangeEdge" 74a13144ffSStefano Zampini PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, Mat* Gins, Mat* GKins, PetscScalar *work, PetscReal *rwork) 75a13144ffSStefano Zampini { 76a13144ffSStefano Zampini PetscErrorCode ierr; 77a13144ffSStefano Zampini Mat GE,GEd; 78a13144ffSStefano Zampini PetscInt rsize,csize,esize; 79a13144ffSStefano Zampini PetscScalar *ptr; 80a13144ffSStefano Zampini 81a13144ffSStefano Zampini PetscFunctionBegin; 82a13144ffSStefano Zampini ierr = ISGetSize(edge,&esize);CHKERRQ(ierr); 83c3c0e390SStefano Zampini if (!esize) PetscFunctionReturn(0); 84a13144ffSStefano Zampini ierr = ISGetSize(extrow,&rsize);CHKERRQ(ierr); 85a13144ffSStefano Zampini ierr = ISGetSize(extcol,&csize);CHKERRQ(ierr); 86a13144ffSStefano Zampini 87a13144ffSStefano Zampini /* gradients */ 88a13144ffSStefano Zampini ptr = work + 5*esize; 89a13144ffSStefano Zampini ierr = MatGetSubMatrix(lG,extrow,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 90a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,rsize,csize,ptr,Gins);CHKERRQ(ierr); 91a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,Gins);CHKERRQ(ierr); 92a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 93a13144ffSStefano Zampini 94a13144ffSStefano Zampini /* constants */ 95a13144ffSStefano Zampini ptr += rsize*csize; 96a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,esize,csize,ptr,&GEd);CHKERRQ(ierr); 97a13144ffSStefano Zampini ierr = MatGetSubMatrix(lG,edge,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 98a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,&GEd);CHKERRQ(ierr); 99a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 100a13144ffSStefano Zampini ierr = MatDense_OrthogonalComplement(GEd,5*esize,work,rwork,GKins);CHKERRQ(ierr); 101a13144ffSStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 102a13144ffSStefano Zampini PetscFunctionReturn(0); 103a13144ffSStefano Zampini } 104a13144ffSStefano Zampini 105a13144ffSStefano Zampini #undef __FUNCT__ 106a13144ffSStefano Zampini #define __FUNCT__ "PCBDDCNedelecSupport" 107a13144ffSStefano Zampini PetscErrorCode PCBDDCNedelecSupport(PC pc) 108a13144ffSStefano Zampini { 109a13144ffSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 110a13144ffSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1110569b399SStefano Zampini Mat G,T,conn,lG,lGt,lGis,lGall,lGe,lGinit; 112a13144ffSStefano Zampini MatNullSpace nnsp; 113a13144ffSStefano Zampini Vec tvec,*quads; 114a13144ffSStefano Zampini PetscSF sfv; 115c2151214SStefano Zampini ISLocalToGlobalMapping el2g,vl2g,fl2g; 116a13144ffSStefano Zampini MPI_Comm comm; 117c2151214SStefano Zampini IS lned,primals,allprimals,nedfieldlocal; 118c2151214SStefano Zampini IS *eedges,*extrows,*extcols,*alleedges; 119c2151214SStefano Zampini PetscBT btv,bte,btvc,btb,btvcand,btvi,btee; 120a13144ffSStefano Zampini PetscScalar *vals,*work; 121a13144ffSStefano Zampini PetscReal *rwork; 122a13144ffSStefano Zampini const PetscInt *idxs,*ii,*jj,*iit,*jjt; 123c2151214SStefano Zampini PetscInt ne,nv,Lv,order,n,field; 124a13144ffSStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 125a13144ffSStefano Zampini PetscInt i,j,extmem,cum,maxsize,rst,nee,nquads=2; 126a13144ffSStefano Zampini PetscInt *extrow,*extrowcum,*marks,*emarks,*vmarks,*gidxs; 127a13144ffSStefano Zampini PetscInt *sfvleaves,*sfvroots; 128c2151214SStefano Zampini PetscBool print,eerr,done,lrc[2],conforming; 129a13144ffSStefano Zampini PetscErrorCode ierr; 130a13144ffSStefano Zampini 131a13144ffSStefano Zampini PetscFunctionBegin; 132a13144ffSStefano Zampini /* test variable order code and print debug info TODO: to be removed */ 133a13144ffSStefano Zampini print = PETSC_FALSE; 134a13144ffSStefano Zampini ierr = PetscOptionsGetBool(NULL,NULL,"-pc_bddc_print_nedelec",&print,NULL);CHKERRQ(ierr); 135a13144ffSStefano Zampini ierr = PetscOptionsGetInt(NULL,NULL,"-pc_bddc_nedelec_order",&pcbddc->nedorder,NULL);CHKERRQ(ierr); 136a13144ffSStefano Zampini 137a13144ffSStefano Zampini /* Return to caller if there are no edges in the decomposition */ 138a13144ffSStefano Zampini ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 139a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&el2g,NULL);CHKERRQ(ierr); 140c2151214SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(el2g,&n);CHKERRQ(ierr); 141a13144ffSStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 142a13144ffSStefano Zampini lrc[0] = PETSC_FALSE; 143c2151214SStefano Zampini for (i=0;i<n;i++) { 144a13144ffSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 145a13144ffSStefano Zampini lrc[0] = PETSC_TRUE; 146a13144ffSStefano Zampini break; 147a13144ffSStefano Zampini } 148a13144ffSStefano Zampini } 149a13144ffSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 150a13144ffSStefano Zampini ierr = MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 151a13144ffSStefano Zampini if (!lrc[1]) PetscFunctionReturn(0); 152a13144ffSStefano Zampini 153c2151214SStefano Zampini /* Get discrete gradient 154c2151214SStefano Zampini If it is defined for a subset of dofs, assumes G is given in global ordering for all the dofs */ 155a13144ffSStefano Zampini G = pcbddc->discretegradient; 156a13144ffSStefano Zampini order = pcbddc->nedorder; 157a13144ffSStefano Zampini conforming = pcbddc->conforming; 158c2151214SStefano Zampini field = pcbddc->nedfield; 159c2151214SStefano Zampini if (pcbddc->n_ISForDofsLocal && field >= pcbddc->n_ISForDofsLocal) SETERRQ2(comm,PETSC_ERR_USER,"Invalid field for Nedelec %d: number of fields is %d",field,pcbddc->n_ISForDofsLocal); 160c2151214SStefano Zampini if (pcbddc->n_ISForDofsLocal && field > -1) { 161c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);CHKERRQ(ierr); 162c2151214SStefano Zampini nedfieldlocal = pcbddc->ISForDofsLocal[field]; 163c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 164c2151214SStefano Zampini } else { 165c2151214SStefano Zampini PetscBool testnedfield = PETSC_FALSE; 166c2151214SStefano Zampini ierr = PetscOptionsGetBool(NULL,NULL,"-pc_bddc_nedelec_field",&testnedfield,NULL);CHKERRQ(ierr); 167c2151214SStefano Zampini if (!testnedfield) { 168c2151214SStefano Zampini ne = n; 169c2151214SStefano Zampini nedfieldlocal = NULL; 170c2151214SStefano Zampini } else { 171c2151214SStefano Zampini /* ierr = ISCreateStride(comm,n,0,1,&nedfieldlocal);CHKERRQ(ierr); */ 172c2151214SStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 173c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 174c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 175c2151214SStefano Zampini for (i=0;i<n;i++) matis->sf_leafdata[i] = 1; 176c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 177c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 178c2151214SStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 179c2151214SStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 180c2151214SStefano Zampini for (i=0,cum=0;i<n;i++) { 181c2151214SStefano Zampini if (matis->sf_leafdata[i] > 1) { 182c2151214SStefano Zampini matis->sf_leafdata[cum++] = i; 183c2151214SStefano Zampini } 184c2151214SStefano Zampini } 185c2151214SStefano Zampini ierr = ISCreateGeneral(comm,cum,matis->sf_leafdata,PETSC_COPY_VALUES,&nedfieldlocal);CHKERRQ(ierr); 186c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 187c2151214SStefano Zampini } 188c2151214SStefano Zampini } 189a13144ffSStefano Zampini 190c2151214SStefano Zampini if (nedfieldlocal) { /* merge with previous code when testing is done */ 191c2151214SStefano Zampini IS is; 192c2151214SStefano Zampini 193c2151214SStefano Zampini /* need to map from the local Nedelec field to local numbering */ 194c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);CHKERRQ(ierr); 195c2151214SStefano Zampini /* need to map from the local Nedelec field to global numbering */ 196c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(el2g,nedfieldlocal,&is);CHKERRQ(ierr); 197c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is,&el2g);CHKERRQ(ierr); 198c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 199c2151214SStefano Zampini } else { 200c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)el2g);CHKERRQ(ierr); 201c2151214SStefano Zampini fl2g = NULL; 202c2151214SStefano Zampini } 203a13144ffSStefano Zampini 204a13144ffSStefano Zampini /* Sanity checks */ 205a13144ffSStefano Zampini if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time"); 206a13144ffSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis"); 207a13144ffSStefano Zampini if (order && ne%order) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"The number of local edge dofs %d it's not a multiple of the order %d",ne,order); 208a13144ffSStefano Zampini 209a13144ffSStefano Zampini /* Drop connections for interior edges (this modifies G) */ 210c2151214SStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 211c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 212c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 213c2151214SStefano Zampini if (nedfieldlocal) { 214c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 215c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1; 216c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 217c2151214SStefano Zampini } else { 218c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1; 219c2151214SStefano Zampini } 220c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 221c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 222c2151214SStefano Zampini ierr = MatGetOwnershipRange(G,&rst,NULL);CHKERRQ(ierr); 223c2151214SStefano Zampini for (i=0,cum=0;i<pc->pmat->rmap->n;i++) { 224c2151214SStefano Zampini if (matis->sf_rootdata[i] < 2) { 225c2151214SStefano Zampini matis->sf_rootdata[cum++] = i + rst; 226c2151214SStefano Zampini } 227c2151214SStefano Zampini } 228a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);CHKERRQ(ierr); 229a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 230c2151214SStefano Zampini ierr = MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);CHKERRQ(ierr); 231a13144ffSStefano Zampini 232a13144ffSStefano Zampini /* Extract subdomain relevant rows of G */ 233a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(el2g,&idxs);CHKERRQ(ierr); 234a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);CHKERRQ(ierr); 235a13144ffSStefano Zampini ierr = MatGetSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);CHKERRQ(ierr); 236a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);CHKERRQ(ierr); 237a13144ffSStefano Zampini ierr = ISDestroy(&lned);CHKERRQ(ierr); 238a13144ffSStefano Zampini ierr = MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);CHKERRQ(ierr); 239a13144ffSStefano Zampini ierr = MatDestroy(&lGall);CHKERRQ(ierr); 240a13144ffSStefano Zampini ierr = MatISGetLocalMat(lGis,&lG);CHKERRQ(ierr); 241a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)lG);CHKERRQ(ierr); 242a13144ffSStefano Zampini if (print) { 243a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"initial_lG");CHKERRQ(ierr); 244a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 245a13144ffSStefano Zampini } 246a13144ffSStefano Zampini 247a13144ffSStefano Zampini /* SF for nodal communications */ 248c2151214SStefano Zampini ierr = MatGetLocalSize(G,NULL,&Lv);CHKERRQ(ierr); 249a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);CHKERRQ(ierr); 250a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)vl2g);CHKERRQ(ierr); 251a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(vl2g,&nv);CHKERRQ(ierr); 252a13144ffSStefano Zampini ierr = PetscSFCreate(comm,&sfv);CHKERRQ(ierr); 253a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(vl2g,&idxs);CHKERRQ(ierr); 254a13144ffSStefano Zampini ierr = PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);CHKERRQ(ierr); 255a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);CHKERRQ(ierr); 256a13144ffSStefano Zampini ierr = PetscMalloc2(nv,&sfvleaves,Lv,&sfvroots);CHKERRQ(ierr); 257a13144ffSStefano Zampini 258a13144ffSStefano Zampini /* Destroy temporary G created in MATIS format */ 259a13144ffSStefano Zampini ierr = MatDestroy(&lGis);CHKERRQ(ierr); 260a13144ffSStefano Zampini 2610569b399SStefano Zampini /* Save lG */ 2620569b399SStefano Zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);CHKERRQ(ierr); 2630569b399SStefano Zampini 264a13144ffSStefano Zampini /* Analyze the edge-nodes connections (duplicate lG) */ 2654e64d54eSstefano_zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGe);CHKERRQ(ierr); 2664e64d54eSstefano_zampini ierr = MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 267a13144ffSStefano Zampini ierr = PetscBTCreate(nv,&btv);CHKERRQ(ierr); 268a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&bte);CHKERRQ(ierr); 269a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&btb);CHKERRQ(ierr); 270c2151214SStefano Zampini ierr = PetscBTCreate(nv,&btvcand);CHKERRQ(ierr); 271a13144ffSStefano Zampini /* need to import the boundary specification to ensure the 272a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 273a13144ffSStefano Zampini if (pcbddc->DirichletBoundariesLocal) { 274c2151214SStefano Zampini IS is; 275c2151214SStefano Zampini 276c2151214SStefano Zampini if (fl2g) { 277c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);CHKERRQ(ierr); 278c2151214SStefano Zampini } else { 279c2151214SStefano Zampini is = pcbddc->DirichletBoundariesLocal; 280c2151214SStefano Zampini } 281c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 282c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 283a13144ffSStefano Zampini for (i=0;i<cum;i++) { 284a13144ffSStefano Zampini if (idxs[i] >= 0) { 285a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 286a13144ffSStefano Zampini } 287a13144ffSStefano Zampini } 288c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 289c2151214SStefano Zampini if (fl2g) { 290c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 291c2151214SStefano Zampini } 292a13144ffSStefano Zampini } 293a13144ffSStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 294c2151214SStefano Zampini IS is; 295c2151214SStefano Zampini 296c2151214SStefano Zampini if (fl2g) { 297c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);CHKERRQ(ierr); 298c2151214SStefano Zampini } else { 299c2151214SStefano Zampini is = pcbddc->NeumannBoundariesLocal; 300c2151214SStefano Zampini } 301c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 302c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 303a13144ffSStefano Zampini for (i=0;i<cum;i++) { 304a13144ffSStefano Zampini if (idxs[i] >= 0) { 305a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 306a13144ffSStefano Zampini } 307a13144ffSStefano Zampini } 308c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 309c2151214SStefano Zampini if (fl2g) { 310c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 311a13144ffSStefano Zampini } 312c2151214SStefano Zampini } 313c2151214SStefano Zampini 314a13144ffSStefano Zampini /* need to remove coarse faces' dofs to ensure the 315a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 316a13144ffSStefano Zampini ierr = PetscCalloc1(ne,&marks);CHKERRQ(ierr); 31762b0c6f7SStefano Zampini ierr = PetscBTCreate(ne,&btee);CHKERRQ(ierr); 318a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 31962b0c6f7SStefano Zampini for (i=1;i<n_neigh;i++) 32062b0c6f7SStefano Zampini for (j=0;j<n_shared[i];j++) 32162b0c6f7SStefano Zampini marks[shared[i][j]]++; 322a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 32362b0c6f7SStefano Zampini for (i=0;i<ne;i++) { 32462b0c6f7SStefano Zampini if (marks[i] > 1 || (marks[i] == 1 && PetscBTLookup(btb,i))) { 32562b0c6f7SStefano Zampini ierr = PetscBTSet(btee,i);CHKERRQ(ierr); 32662b0c6f7SStefano Zampini } 32762b0c6f7SStefano Zampini } 32862b0c6f7SStefano Zampini 32962b0c6f7SStefano Zampini if (!conforming) { 33062b0c6f7SStefano Zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 33162b0c6f7SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 33262b0c6f7SStefano Zampini } 3334e64d54eSstefano_zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 334dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGe,&vals);CHKERRQ(ierr); 33562b0c6f7SStefano Zampini cum = 0; 336a13144ffSStefano Zampini for (i=0;i<ne;i++) { 337dec27d64SStefano Zampini /* eliminate rows corresponding to edge dofs belonging to coarse faces */ 33862b0c6f7SStefano Zampini if (!PetscBTLookup(btee,i)) { 339a13144ffSStefano Zampini marks[cum++] = i; 340dec27d64SStefano Zampini continue; 341dec27d64SStefano Zampini } 342dec27d64SStefano Zampini /* set badly connected edge dofs as primal */ 34362b0c6f7SStefano Zampini if (!conforming) { 34462b0c6f7SStefano Zampini if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */ 345a13144ffSStefano Zampini marks[cum++] = i; 346a13144ffSStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 347a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 348a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 349a13144ffSStefano Zampini } 35062b0c6f7SStefano Zampini } else { 35162b0c6f7SStefano Zampini /* every edge dofs should be connected trough a certain number of nodal dofs 35262b0c6f7SStefano Zampini to other edge dofs belonging to coarse edges 35362b0c6f7SStefano Zampini - at most 2 endpoints 35462b0c6f7SStefano Zampini - order-1 interior nodal dofs 35562b0c6f7SStefano Zampini - no undefined nodal dofs (nconn < order) 35662b0c6f7SStefano Zampini */ 35762b0c6f7SStefano Zampini PetscInt ends = 0,ints = 0, undef = 0; 35862b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 35962b0c6f7SStefano Zampini PetscInt v = jj[j],k; 36062b0c6f7SStefano Zampini PetscInt nconn = iit[v+1]-iit[v]; 36162b0c6f7SStefano Zampini for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--; 36262b0c6f7SStefano Zampini if (nconn > order) ends++; 36362b0c6f7SStefano Zampini else if (nconn == order) ints++; 36462b0c6f7SStefano Zampini else undef++; 36562b0c6f7SStefano Zampini } 36662b0c6f7SStefano Zampini if (undef || ends > 2 || ints != order -1) { 36762b0c6f7SStefano Zampini marks[cum++] = i; 36862b0c6f7SStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 36962b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 37062b0c6f7SStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 37162b0c6f7SStefano Zampini } 37262b0c6f7SStefano Zampini } 37362b0c6f7SStefano Zampini } 374a13144ffSStefano Zampini } 375dec27d64SStefano Zampini /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */ 376dec27d64SStefano Zampini if (!order && ii[i+1] != ii[i]) { 377dec27d64SStefano Zampini PetscScalar val = 1./(ii[i+1]-ii[i]-1); 378dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vals[j] = val; 379a13144ffSStefano Zampini } 380dec27d64SStefano Zampini } 38162b0c6f7SStefano Zampini ierr = PetscBTDestroy(&btee);CHKERRQ(ierr); 382dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGe,&vals);CHKERRQ(ierr); 3834e64d54eSstefano_zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 38462b0c6f7SStefano Zampini if (!conforming) { 38562b0c6f7SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 38662b0c6f7SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 38762b0c6f7SStefano Zampini } 3884e64d54eSstefano_zampini ierr = MatZeroRows(lGe,cum,marks,0.,NULL,NULL);CHKERRQ(ierr); 389a13144ffSStefano Zampini /* identify splitpoints and corner candidates: TODO variable order */ 3904e64d54eSstefano_zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 391a13144ffSStefano Zampini if (print) { 3924e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");CHKERRQ(ierr); 3934e64d54eSstefano_zampini ierr = MatView(lGe,NULL);CHKERRQ(ierr); 3944e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");CHKERRQ(ierr); 395a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 396a13144ffSStefano Zampini } 3970569b399SStefano Zampini ierr = MatDestroy(&lGe);CHKERRQ(ierr); 398a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 399dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGt,&vals);CHKERRQ(ierr); 400a13144ffSStefano Zampini for (i=0;i<nv;i++) { 401dec27d64SStefano Zampini PetscInt ord = order, test = ii[i+1]-ii[i]; 402dec27d64SStefano Zampini if (!order) { 403dec27d64SStefano Zampini PetscReal vorder = 0.; 404dec27d64SStefano Zampini 405dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]); 406dec27d64SStefano Zampini test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON); 407dec27d64SStefano Zampini if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%d)",vorder,test); 408dec27d64SStefano Zampini ord = 1; 409dec27d64SStefano Zampini } 410a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 411dec27d64SStefano Zampini if (test%ord) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected number of edge dofs %d connected with nodal dof %d with order %d",test,i,ord); 412a13144ffSStefano Zampini #endif 413dec27d64SStefano Zampini if (test >= 3*ord) { /* splitpoints */ 414a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %d\n",i); 415a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 416dec27d64SStefano Zampini } else if (test == ord) { 417a13144ffSStefano Zampini if (order == 1) { 418a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %d\n",i); 419a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 420a13144ffSStefano Zampini } else { 421a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %d\n",i); 422a13144ffSStefano Zampini ierr = PetscBTSet(btvcand,i);CHKERRQ(ierr); 423a13144ffSStefano Zampini } 424a13144ffSStefano Zampini } 425a13144ffSStefano Zampini } 426dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGt,&vals);CHKERRQ(ierr); 427a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 428a13144ffSStefano Zampini 429a13144ffSStefano Zampini /* Get the local G^T explicitly */ 4300569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 431a13144ffSStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 4324e64d54eSstefano_zampini ierr = MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 433a13144ffSStefano Zampini 4344e64d54eSstefano_zampini /* Mark interior nodal dofs */ 435a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 4364e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvi);CHKERRQ(ierr); 437a13144ffSStefano Zampini for (i=1;i<n_neigh;i++) { 438a13144ffSStefano Zampini for (j=0;j<n_shared[i];j++) { 4394e64d54eSstefano_zampini ierr = PetscBTSet(btvi,shared[i][j]);CHKERRQ(ierr); 440a13144ffSStefano Zampini } 441a13144ffSStefano Zampini } 442a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 443a13144ffSStefano Zampini 444a13144ffSStefano Zampini /* communicate corners and splitpoints */ 445a13144ffSStefano Zampini ierr = PetscMalloc1(nv,&vmarks);CHKERRQ(ierr); 446a13144ffSStefano Zampini ierr = PetscMemzero(sfvleaves,nv*sizeof(PetscInt));CHKERRQ(ierr); 447a13144ffSStefano Zampini ierr = PetscMemzero(sfvroots,Lv*sizeof(PetscInt));CHKERRQ(ierr); 448a13144ffSStefano Zampini for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1; 449a13144ffSStefano Zampini 450a13144ffSStefano Zampini if (print) { 451a13144ffSStefano Zampini IS tbz; 452a13144ffSStefano Zampini 453a13144ffSStefano Zampini cum = 0; 454a13144ffSStefano Zampini for (i=0;i<nv;i++) 455a13144ffSStefano Zampini if (sfvleaves[i]) 456a13144ffSStefano Zampini vmarks[cum++] = i; 457a13144ffSStefano Zampini 458a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 459a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");CHKERRQ(ierr); 460a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 461a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 462a13144ffSStefano Zampini } 463a13144ffSStefano Zampini 464a13144ffSStefano Zampini ierr = PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 465a13144ffSStefano Zampini ierr = PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 466a13144ffSStefano Zampini ierr = PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 467a13144ffSStefano Zampini ierr = PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 468a13144ffSStefano Zampini 4694e64d54eSstefano_zampini /* Zero rows of lGt corresponding to identified corners 4704e64d54eSstefano_zampini and interior nodal dofs */ 471a13144ffSStefano Zampini cum = 0; 472a13144ffSStefano Zampini for (i=0;i<nv;i++) { 473a13144ffSStefano Zampini if (sfvleaves[i]) { 474a13144ffSStefano Zampini vmarks[cum++] = i; 475a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 476a13144ffSStefano Zampini } 4774e64d54eSstefano_zampini if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i; 478a13144ffSStefano Zampini } 4794e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvi);CHKERRQ(ierr); 480a13144ffSStefano Zampini if (print) { 481a13144ffSStefano Zampini IS tbz; 482a13144ffSStefano Zampini 483a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 4844e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");CHKERRQ(ierr); 485a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 486a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 487a13144ffSStefano Zampini } 488a13144ffSStefano Zampini ierr = MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);CHKERRQ(ierr); 489a13144ffSStefano Zampini ierr = PetscFree(vmarks);CHKERRQ(ierr); 490a13144ffSStefano Zampini ierr = PetscSFDestroy(&sfv);CHKERRQ(ierr); 491a13144ffSStefano Zampini ierr = PetscFree2(sfvleaves,sfvroots);CHKERRQ(ierr); 492a13144ffSStefano Zampini 493a13144ffSStefano Zampini /* Recompute G */ 494a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 495a13144ffSStefano Zampini ierr = MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);CHKERRQ(ierr); 496a13144ffSStefano Zampini if (print) { 497a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"used_lG");CHKERRQ(ierr); 498a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 499a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lGt,"used_lGt");CHKERRQ(ierr); 500a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 501a13144ffSStefano Zampini } 502a13144ffSStefano Zampini 503a13144ffSStefano Zampini /* Get primal dofs (if any) */ 504a13144ffSStefano Zampini cum = 0; 505a13144ffSStefano Zampini for (i=0;i<ne;i++) { 506a13144ffSStefano Zampini if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i; 507a13144ffSStefano Zampini } 508c2151214SStefano Zampini if (fl2g) { 509c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);CHKERRQ(ierr); 510c2151214SStefano Zampini } 511a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 512a13144ffSStefano Zampini if (print) { 513a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");CHKERRQ(ierr); 514a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 515a13144ffSStefano Zampini } 516a13144ffSStefano Zampini ierr = PetscBTDestroy(&bte);CHKERRQ(ierr); 517c2151214SStefano Zampini /* TODO: what if the user passed in some of them ? */ 518a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 519a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 520a13144ffSStefano Zampini 521a13144ffSStefano Zampini /* Compute edge connectivity */ 522a13144ffSStefano Zampini ierr = PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");CHKERRQ(ierr); 523a13144ffSStefano Zampini ierr = MatMatMultSymbolic(lG,lGt,PETSC_DEFAULT,&conn);CHKERRQ(ierr); 524a13144ffSStefano Zampini ierr = MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 525d904f53bSStefano Zampini if (done && i && ii && jj) { /* when lG is empty, don't pass pointers */ 526c2151214SStefano Zampini if (fl2g) { 527c2151214SStefano Zampini PetscBT btf; 528c2151214SStefano Zampini PetscInt *iia,*jja,*iiu,*jju; 529c2151214SStefano Zampini PetscBool rest = PETSC_FALSE,free = PETSC_FALSE; 530c2151214SStefano Zampini 531c2151214SStefano Zampini /* create CSR for all local dofs */ 532c2151214SStefano Zampini ierr = PetscMalloc1(n+1,&iia);CHKERRQ(ierr); 533c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */ 534c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid size of CSR graph %d. Should be %d\n",pcbddc->mat_graph->nvtxs_csr,n); 535c2151214SStefano Zampini iiu = pcbddc->mat_graph->xadj; 536c2151214SStefano Zampini jju = pcbddc->mat_graph->adjncy; 537c2151214SStefano Zampini } else if (pcbddc->use_local_adj) { 538c2151214SStefano Zampini rest = PETSC_TRUE; 539c2151214SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 540c2151214SStefano Zampini } else { 541c2151214SStefano Zampini free = PETSC_TRUE; 542c2151214SStefano Zampini ierr = PetscMalloc2(n+1,&iiu,n,&jju);CHKERRQ(ierr); 543c2151214SStefano Zampini iiu[0] = 0; 544c2151214SStefano Zampini for (i=0;i<n;i++) { 545c2151214SStefano Zampini iiu[i+1] = i+1; 546c2151214SStefano Zampini jju[i] = -1; 547d904f53bSStefano Zampini } 548c2151214SStefano Zampini } 549c2151214SStefano Zampini 550c2151214SStefano Zampini /* import sizes of CSR */ 551c2151214SStefano Zampini iia[0] = 0; 552c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i]; 553c2151214SStefano Zampini 554c2151214SStefano Zampini /* overwrite entries corresponding to the Nedelec field */ 555c2151214SStefano Zampini ierr = PetscBTCreate(n,&btf);CHKERRQ(ierr); 556c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 557c2151214SStefano Zampini for (i=0;i<ne;i++) { 558c2151214SStefano Zampini ierr = PetscBTSet(btf,idxs[i]);CHKERRQ(ierr); 559c2151214SStefano Zampini iia[idxs[i]+1] = ii[i+1]-ii[i]; 560c2151214SStefano Zampini } 561c2151214SStefano Zampini 562c2151214SStefano Zampini /* iia in CSR */ 563c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] += iia[i]; 564c2151214SStefano Zampini 565c2151214SStefano Zampini /* jja in CSR */ 566c2151214SStefano Zampini ierr = PetscMalloc1(iia[n],&jja);CHKERRQ(ierr); 567c2151214SStefano Zampini for (i=0;i<n;i++) 568c2151214SStefano Zampini if (!PetscBTLookup(btf,i)) 569c2151214SStefano Zampini for (j=0;j<iiu[i+1]-iiu[i];j++) 570c2151214SStefano Zampini jja[iia[i]+j] = jju[iiu[i]+j]; 571c2151214SStefano Zampini 572c2151214SStefano Zampini /* map edge dofs connectivity */ 573c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);CHKERRQ(ierr); 574c2151214SStefano Zampini for (i=0;i<ne;i++) { 575c2151214SStefano Zampini PetscInt e = idxs[i]; 576c2151214SStefano Zampini for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j]; 577c2151214SStefano Zampini } 578c2151214SStefano Zampini 579c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 580c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);CHKERRQ(ierr); 581c2151214SStefano Zampini if (rest) { 582c2151214SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 583c2151214SStefano Zampini } 584c2151214SStefano Zampini if (free) { 585c2151214SStefano Zampini ierr = PetscFree2(iiu,jju);CHKERRQ(ierr); 586c2151214SStefano Zampini } 587c2151214SStefano Zampini ierr = PetscBTDestroy(&btf);CHKERRQ(ierr); 588c2151214SStefano Zampini } else { 589c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);CHKERRQ(ierr); 590c2151214SStefano Zampini } 591c2151214SStefano Zampini } 592c2151214SStefano Zampini 593a13144ffSStefano Zampini /* Analyze interface for edge dofs */ 594a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 595a13144ffSStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 596a13144ffSStefano Zampini 597a13144ffSStefano Zampini /* Get coarse edges in the edge space */ 598c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 599a13144ffSStefano Zampini ierr = MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 600a13144ffSStefano Zampini 601c2151214SStefano Zampini if (fl2g) { 602c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 603c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 604c2151214SStefano Zampini for (i=0;i<nee;i++) { 605c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 606c2151214SStefano Zampini } 607c2151214SStefano Zampini } else { 608c2151214SStefano Zampini eedges = alleedges; 609c2151214SStefano Zampini primals = allprimals; 610c2151214SStefano Zampini } 611c2151214SStefano Zampini 612a13144ffSStefano Zampini /* Mark fine edge dofs with their coarse edge id */ 613a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 614c2151214SStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 615c2151214SStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 616c2151214SStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = nee+1; 617c2151214SStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 618c2151214SStefano Zampini if (print) { 619c2151214SStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");CHKERRQ(ierr); 620c2151214SStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 621c2151214SStefano Zampini } 622c2151214SStefano Zampini 623c2151214SStefano Zampini maxsize = 0; 624a13144ffSStefano Zampini for (i=0;i<nee;i++) { 625a13144ffSStefano Zampini PetscInt size,mark = i+1; 626a13144ffSStefano Zampini 627a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 628a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 629a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 630a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 631a13144ffSStefano Zampini maxsize = PetscMax(maxsize,size); 632a13144ffSStefano Zampini } 633a13144ffSStefano Zampini 634a13144ffSStefano Zampini /* Find coarse edge endpoints */ 635a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 636a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 637a13144ffSStefano Zampini for (i=0;i<nee;i++) { 638a13144ffSStefano Zampini PetscInt mark = i+1,size; 639a13144ffSStefano Zampini 640a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 641a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 642a13144ffSStefano Zampini if (print) { 643a13144ffSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %d\n",i); 644a13144ffSStefano Zampini ISView(eedges[i],NULL); 645a13144ffSStefano Zampini } 646a13144ffSStefano Zampini for (j=0;j<size;j++) { 647a13144ffSStefano Zampini PetscInt k, ee = idxs[j]; 648a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," idx %d\n",ee); 649a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 650a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," inspect %d\n",jj[k]); 651a13144ffSStefano Zampini if (PetscBTLookup(btv,jj[k])) { 652a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %d\n",jj[k]); 653a13144ffSStefano Zampini } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */ 654a13144ffSStefano Zampini PetscInt k2; 655a13144ffSStefano Zampini PetscBool corner = PETSC_FALSE; 656a13144ffSStefano Zampini for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) { 657c2151214SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," INSPECTING %d: mark %d (ref mark %d), boundary %d\n",jjt[k2],marks[jjt[k2]],mark,!!PetscBTLookup(btb,jjt[k2])); 658c2151214SStefano Zampini /* it's a corner if either is connected with an edge dof belonging to a different cc or 659c2151214SStefano Zampini if the edge dof lie on the natural part of the boundary */ 660c2151214SStefano Zampini if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) { 661a13144ffSStefano Zampini corner = PETSC_TRUE; 662a13144ffSStefano Zampini break; 663a13144ffSStefano Zampini } 664a13144ffSStefano Zampini } 665a13144ffSStefano Zampini if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */ 666a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found %d\n",jj[k]); 667a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 668a13144ffSStefano Zampini } else { 669a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n"); 670a13144ffSStefano Zampini } 671a13144ffSStefano Zampini } 672a13144ffSStefano Zampini } 673a13144ffSStefano Zampini } 674a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 675a13144ffSStefano Zampini } 676a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 677a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 678c2151214SStefano Zampini ierr = PetscBTDestroy(&btb);CHKERRQ(ierr); 679a13144ffSStefano Zampini 680a13144ffSStefano Zampini /* Reset marked primal dofs */ 681a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 682a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 683a13144ffSStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = 0; 684a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 685a13144ffSStefano Zampini 6860569b399SStefano Zampini /* Now use the initial lG */ 6870569b399SStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 6880569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 6890569b399SStefano Zampini lG = lGinit; 6900569b399SStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 6910569b399SStefano Zampini 692a13144ffSStefano Zampini /* Compute extended cols indices */ 693a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 694a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lG,&i);CHKERRQ(ierr); 695a13144ffSStefano Zampini i *= maxsize; 696a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extcols);CHKERRQ(ierr); 697a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&emarks);CHKERRQ(ierr); 698a13144ffSStefano Zampini ierr = PetscMalloc2(i,&extrow,i,&gidxs);CHKERRQ(ierr); 699a13144ffSStefano Zampini eerr = PETSC_FALSE; 700a13144ffSStefano Zampini for (i=0;i<nee;i++) { 701a13144ffSStefano Zampini PetscInt size; 702a13144ffSStefano Zampini 703a13144ffSStefano Zampini cum = 0; 704a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 705a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 706a13144ffSStefano Zampini for (j=0;j<size;j++) { 707a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 708a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) if (!PetscBTLookup(btv,jj[k])) extrow[cum++] = jj[k]; 709a13144ffSStefano Zampini } 710a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 711a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 712a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 713a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 714a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 715a13144ffSStefano Zampini /* it may happen that endpoints are not defined at this point 716a13144ffSStefano Zampini if it is the case, mark this edge for a second pass */ 717a13144ffSStefano Zampini if (cum != size -1) { 718a13144ffSStefano Zampini emarks[i] = 1; 719a13144ffSStefano Zampini if (print) { 720a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge");CHKERRQ(ierr); 721a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 722a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol");CHKERRQ(ierr); 723a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 724a13144ffSStefano Zampini } 725a13144ffSStefano Zampini eerr = PETSC_TRUE; 726a13144ffSStefano Zampini } 727a13144ffSStefano Zampini } 7284e64d54eSstefano_zampini /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */ 729a13144ffSStefano Zampini ierr = MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 730a13144ffSStefano Zampini if (done) { 731a13144ffSStefano Zampini PetscInt *newprimals; 732a13144ffSStefano Zampini 733a13144ffSStefano Zampini ierr = PetscMalloc1(ne,&newprimals);CHKERRQ(ierr); 734a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 735a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 736a13144ffSStefano Zampini ierr = PetscMemcpy(newprimals,idxs,cum*sizeof(PetscInt));CHKERRQ(ierr); 737a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 7380569b399SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 739a13144ffSStefano Zampini for (i=0;i<nee;i++) { 740a13144ffSStefano Zampini if (emarks[i]) { 741a13144ffSStefano Zampini PetscInt size,mark = i+1; 742a13144ffSStefano Zampini 743a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 744a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 745c2151214SStefano Zampini /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */ 746a13144ffSStefano Zampini for (j=0;j<size;j++) { 747a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 748a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 749a13144ffSStefano Zampini /* set all candidates located on the edge as corners */ 750a13144ffSStefano Zampini if (PetscBTLookup(btvcand,jj[k])) { 751a13144ffSStefano Zampini PetscInt k2,vv = jj[k]; 752a13144ffSStefano Zampini ierr = PetscBTSet(btv,vv);CHKERRQ(ierr); 753a13144ffSStefano Zampini /* set all edge dofs connected to candidate as primals */ 754a13144ffSStefano Zampini for (k2=iit[vv];k2<iit[vv+1];k2++) { 755a13144ffSStefano Zampini if (marks[jjt[k2]] == mark) { 756a13144ffSStefano Zampini PetscInt k3,ee2 = jjt[k2]; 757a13144ffSStefano Zampini newprimals[cum++] = ee2; 758a13144ffSStefano Zampini /* finally set the new corners */ 759a13144ffSStefano Zampini for (k3=ii[ee2];k3<ii[ee2+1];k3++) { 760a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k3]);CHKERRQ(ierr); 761a13144ffSStefano Zampini } 762a13144ffSStefano Zampini } 763a13144ffSStefano Zampini } 764a13144ffSStefano Zampini } 765a13144ffSStefano Zampini } 766a13144ffSStefano Zampini } 767a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 768a13144ffSStefano Zampini } 769a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 770a13144ffSStefano Zampini } 7710569b399SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 772a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,newprimals);CHKERRQ(ierr); 773c2151214SStefano Zampini if (fl2g) { 774c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);CHKERRQ(ierr); 775c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 776c2151214SStefano Zampini for (i=0;i<nee;i++) { 777c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 778c2151214SStefano Zampini } 779c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 780c2151214SStefano Zampini } 781c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 782a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 783a13144ffSStefano Zampini ierr = PetscFree(newprimals);CHKERRQ(ierr); 784a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 785a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 786a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 787c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 788c2151214SStefano Zampini if (fl2g) { 789c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 790c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 791c2151214SStefano Zampini for (i=0;i<nee;i++) { 792c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 793c2151214SStefano Zampini } 794c2151214SStefano Zampini } else { 795c2151214SStefano Zampini eedges = alleedges; 796c2151214SStefano Zampini primals = allprimals; 797c2151214SStefano Zampini } 798a13144ffSStefano Zampini 799a13144ffSStefano Zampini /* Mark again */ 800a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 801a13144ffSStefano Zampini for (i=0;i<nee;i++) { 802a13144ffSStefano Zampini PetscInt size,mark = i+1; 803a13144ffSStefano Zampini 804a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 805a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 806a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 807a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 808a13144ffSStefano Zampini } 809a13144ffSStefano Zampini if (print) { 810a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");CHKERRQ(ierr); 811a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 812a13144ffSStefano Zampini } 813a13144ffSStefano Zampini 814a13144ffSStefano Zampini /* Recompute extended cols */ 815a13144ffSStefano Zampini eerr = PETSC_FALSE; 816a13144ffSStefano Zampini for (i=0;i<nee;i++) { 817a13144ffSStefano Zampini PetscInt size; 818a13144ffSStefano Zampini 819a13144ffSStefano Zampini cum = 0; 820a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 821a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 822a13144ffSStefano Zampini for (j=0;j<size;j++) { 823a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 824a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 825a13144ffSStefano Zampini if (!PetscBTLookup(btv,jj[k])) { 826a13144ffSStefano Zampini extrow[cum++] = jj[k]; 827a13144ffSStefano Zampini } 828a13144ffSStefano Zampini } 829a13144ffSStefano Zampini } 830a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 831a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 832a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 833a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 834a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 835a13144ffSStefano Zampini if (cum != size -1) { 836a13144ffSStefano Zampini if (print) { 837a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");CHKERRQ(ierr); 838a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 839a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");CHKERRQ(ierr); 840a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 841a13144ffSStefano Zampini } 842a13144ffSStefano Zampini eerr = PETSC_TRUE; 843a13144ffSStefano Zampini } 844a13144ffSStefano Zampini } 845a13144ffSStefano Zampini } 846a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 847a13144ffSStefano Zampini ierr = PetscFree2(extrow,gidxs);CHKERRQ(ierr); 848a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 849a13144ffSStefano Zampini /* an error should not occur at this point */ 850a13144ffSStefano Zampini if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS"); 851a13144ffSStefano Zampini 8524e64d54eSstefano_zampini /* Check the number of endpoints */ 853c2151214SStefano Zampini /* TODO: fix case for circular edge */ 8544e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvc);CHKERRQ(ierr); 8550569b399SStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 8564e64d54eSstefano_zampini for (i=0;i<nee;i++) { 8574e64d54eSstefano_zampini PetscInt size, found = 0; 8584e64d54eSstefano_zampini 8594e64d54eSstefano_zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 8604e64d54eSstefano_zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 8614e64d54eSstefano_zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 8624e64d54eSstefano_zampini for (j=0;j<size;j++) { 8634e64d54eSstefano_zampini PetscInt k,ee = idxs[j]; 8644e64d54eSstefano_zampini for (k=ii[ee];k<ii[ee+1];k++) { 8654e64d54eSstefano_zampini PetscInt vv = jj[k]; 8664e64d54eSstefano_zampini if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) { 8674e64d54eSstefano_zampini found++; 8684e64d54eSstefano_zampini } 8694e64d54eSstefano_zampini } 8704e64d54eSstefano_zampini } 8714e64d54eSstefano_zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 8724e64d54eSstefano_zampini if (found != 2) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %d corners for edge %d\n",found,i); 8734e64d54eSstefano_zampini } 8740569b399SStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 8754e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvc);CHKERRQ(ierr); 8764e64d54eSstefano_zampini 877a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 878a13144ffSStefano Zampini /* Inspects columns of lG (rows of lGt) and make sure the change of basis will 879a13144ffSStefano Zampini not interfere with neighbouring coarse edges */ 880a13144ffSStefano Zampini ierr = PetscMalloc1(nee+1,&emarks);CHKERRQ(ierr); 881a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 882a13144ffSStefano Zampini for (i=0;i<nv;i++) { 883a13144ffSStefano Zampini PetscInt emax = 0,eemax = 0; 884a13144ffSStefano Zampini 885a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 886a13144ffSStefano Zampini ierr = PetscMemzero(emarks,(nee+1)*sizeof(PetscInt));CHKERRQ(ierr); 887a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++; 888a13144ffSStefano Zampini for (j=1;j<nee+1;j++) { 889a13144ffSStefano Zampini if (emax < emarks[j]) { 890a13144ffSStefano Zampini emax = emarks[j]; 891a13144ffSStefano Zampini eemax = j; 892a13144ffSStefano Zampini } 893a13144ffSStefano Zampini } 894a13144ffSStefano Zampini /* not relevant for edges */ 895a13144ffSStefano Zampini if (!eemax) continue; 896a13144ffSStefano Zampini 897a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 898a13144ffSStefano Zampini if (marks[jj[j]] && marks[jj[j]] != eemax) { 899c2151214SStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_SUP,"Found 2 coarse edges (id %d and %d) connected through the %d nodal dof at edge dof %d\n",marks[jj[j]]-1,eemax,i,jj[j]); 900a13144ffSStefano Zampini } 901a13144ffSStefano Zampini } 902a13144ffSStefano Zampini } 903a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 904a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 905a13144ffSStefano Zampini #endif 906a13144ffSStefano Zampini 907a13144ffSStefano Zampini /* Compute extended rows indices for edge blocks of the change of basis */ 908a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 909a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);CHKERRQ(ierr); 910a13144ffSStefano Zampini extmem *= maxsize; 911a13144ffSStefano Zampini ierr = PetscMalloc1(extmem*nee,&extrow);CHKERRQ(ierr); 912a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extrows);CHKERRQ(ierr); 913a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&extrowcum);CHKERRQ(ierr); 914a13144ffSStefano Zampini for (i=0;i<nv;i++) { 915a13144ffSStefano Zampini PetscInt mark = 0,size,start; 916a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 917a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) 918a13144ffSStefano Zampini if (marks[jj[j]] && !mark) 919a13144ffSStefano Zampini mark = marks[jj[j]]; 920a13144ffSStefano Zampini 921a13144ffSStefano Zampini /* not relevant */ 922a13144ffSStefano Zampini if (!mark) continue; 923a13144ffSStefano Zampini 924a13144ffSStefano Zampini /* import extended row */ 925a13144ffSStefano Zampini mark--; 926a13144ffSStefano Zampini start = mark*extmem+extrowcum[mark]; 927a13144ffSStefano Zampini size = ii[i+1]-ii[i]; 928a13144ffSStefano Zampini if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %d > %d",extrowcum[mark] + size,extmem); 929a13144ffSStefano Zampini ierr = PetscMemcpy(extrow+start,jj+ii[i],size*sizeof(PetscInt));CHKERRQ(ierr); 930a13144ffSStefano Zampini extrowcum[mark] += size; 931a13144ffSStefano Zampini } 932a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 933a13144ffSStefano Zampini cum = 0; 934a13144ffSStefano Zampini for (i=0;i<nee;i++) { 935a13144ffSStefano Zampini PetscInt size = extrowcum[i],*start = extrow + i*extmem; 936a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&size,start);CHKERRQ(ierr); 937a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);CHKERRQ(ierr); 938a13144ffSStefano Zampini cum = PetscMax(cum,size); 939a13144ffSStefano Zampini } 940a13144ffSStefano Zampini ierr = PetscFree(extrowcum);CHKERRQ(ierr); 941a13144ffSStefano Zampini ierr = PetscFree(marks);CHKERRQ(ierr); 942a13144ffSStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 943a13144ffSStefano Zampini ierr = PetscBTDestroy(&btvcand);CHKERRQ(ierr); 944a13144ffSStefano Zampini 945a13144ffSStefano Zampini /* Workspace for lapack inner calls and VecSetValues */ 946a13144ffSStefano Zampini ierr = PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);CHKERRQ(ierr); 947a13144ffSStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 948a13144ffSStefano Zampini for (i=0;i<maxsize;i++) vals[i] = 1.; 949a13144ffSStefano Zampini 950a13144ffSStefano Zampini /* Create vectors for quadrature rules */ 951c2151214SStefano Zampini /* TODO preserve other quadratures */ 952a13144ffSStefano Zampini ierr = PetscMalloc1(nquads,&quads);CHKERRQ(ierr); 953a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 954a13144ffSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quads[i],NULL);CHKERRQ(ierr); 955a13144ffSStefano Zampini ierr = VecSetLocalToGlobalMapping(quads[i],el2g);CHKERRQ(ierr); 956a13144ffSStefano Zampini } 957a13144ffSStefano Zampini ierr = PCBDDCNullSpaceCreate(comm,PETSC_FALSE,nquads,quads,&nnsp);CHKERRQ(ierr); 958a13144ffSStefano Zampini 959a13144ffSStefano Zampini /* Create change of basis matrix (preallocation can be improved) */ 960a13144ffSStefano Zampini ierr = MatCreate(comm,&T);CHKERRQ(ierr); 961c2151214SStefano Zampini ierr = MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n, 962c2151214SStefano Zampini pc->pmat->rmap->N,pc->pmat->rmap->N);CHKERRQ(ierr); 963a13144ffSStefano Zampini ierr = MatSetType(T,MATAIJ);CHKERRQ(ierr); 964a13144ffSStefano Zampini ierr = MatSeqAIJSetPreallocation(T,10,NULL);CHKERRQ(ierr); 965a13144ffSStefano Zampini ierr = MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);CHKERRQ(ierr); 966a13144ffSStefano Zampini ierr = MatSetLocalToGlobalMapping(T,el2g,el2g);CHKERRQ(ierr); 967a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 968a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 969a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 970a13144ffSStefano Zampini 971a13144ffSStefano Zampini /* Defaults to identity */ 972c2151214SStefano Zampini ierr = MatCreateVecs(pc->pmat,&tvec,NULL);CHKERRQ(ierr); 973a13144ffSStefano Zampini ierr = VecSet(tvec,1.0);CHKERRQ(ierr); 974a13144ffSStefano Zampini ierr = MatDiagonalSet(T,tvec,INSERT_VALUES);CHKERRQ(ierr); 975a13144ffSStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 976a13144ffSStefano Zampini 977a13144ffSStefano Zampini for (i=0;i<nee;i++) { 978a13144ffSStefano Zampini Mat Gins = NULL, GKins = NULL; 979a13144ffSStefano Zampini 980a13144ffSStefano Zampini ierr = PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],&Gins,&GKins,work,rwork);CHKERRQ(ierr); 981a13144ffSStefano Zampini if (Gins && GKins) { 982a13144ffSStefano Zampini PetscScalar *data; 983a13144ffSStefano Zampini const PetscInt *rows,*cols; 984a13144ffSStefano Zampini PetscInt nrh,nch,nrc,ncc; 985a13144ffSStefano Zampini 986a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&cols);CHKERRQ(ierr); 987a13144ffSStefano Zampini /* H1 */ 988a13144ffSStefano Zampini ierr = ISGetIndices(extrows[i],&rows);CHKERRQ(ierr); 989a13144ffSStefano Zampini ierr = MatGetSize(Gins,&nrh,&nch);CHKERRQ(ierr); 990a13144ffSStefano Zampini ierr = MatDenseGetArray(Gins,&data);CHKERRQ(ierr); 991a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);CHKERRQ(ierr); 992a13144ffSStefano Zampini ierr = MatDenseRestoreArray(Gins,&data);CHKERRQ(ierr); 993a13144ffSStefano Zampini ierr = ISRestoreIndices(extrows[i],&rows);CHKERRQ(ierr); 994a13144ffSStefano Zampini /* complement */ 995a13144ffSStefano Zampini ierr = MatGetSize(GKins,&nrc,&ncc);CHKERRQ(ierr); 996a13144ffSStefano Zampini if (ncc > nquads-1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet supported ncc %d nquads %d",ncc,nquads); 997a13144ffSStefano Zampini if (ncc + nch != nrc) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"The sum of the number of columns of GKins %d and Gins %d does not match %d",ncc,nch,nrc); 998a13144ffSStefano Zampini ierr = MatDenseGetArray(GKins,&data);CHKERRQ(ierr); 999a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);CHKERRQ(ierr); 1000a13144ffSStefano Zampini ierr = MatDenseRestoreArray(GKins,&data);CHKERRQ(ierr); 1001a13144ffSStefano Zampini /* Gins kernel quadratures */ 1002a13144ffSStefano Zampini for (j=0;j<ncc;j++) { 1003a13144ffSStefano Zampini ierr = VecSetValueLocal(quads[j],cols[nch+j],1.,INSERT_VALUES);CHKERRQ(ierr); 1004a13144ffSStefano Zampini } 1005a13144ffSStefano Zampini /* H1 average */ 1006a13144ffSStefano Zampini ierr = VecSetValuesLocal(quads[nquads-1],nch,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1007a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&cols);CHKERRQ(ierr); 1008a13144ffSStefano Zampini } 1009a13144ffSStefano Zampini ierr = ISDestroy(&extrows[i]);CHKERRQ(ierr); 1010a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 1011a13144ffSStefano Zampini ierr = MatDestroy(&Gins);CHKERRQ(ierr); 1012a13144ffSStefano Zampini ierr = MatDestroy(&GKins);CHKERRQ(ierr); 1013a13144ffSStefano Zampini } 1014a13144ffSStefano Zampini 1015a13144ffSStefano Zampini /* Start assembling */ 1016a13144ffSStefano Zampini ierr = MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1017a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1018a13144ffSStefano Zampini ierr = VecAssemblyBegin(quads[i]);CHKERRQ(ierr); 1019a13144ffSStefano Zampini } 1020a13144ffSStefano Zampini 1021a13144ffSStefano Zampini /* Free */ 1022c2151214SStefano Zampini if (fl2g) { 1023c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1024c2151214SStefano Zampini for (i=0;i<nee;i++) { 1025c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1026c2151214SStefano Zampini } 1027c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1028c2151214SStefano Zampini } 1029c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1030c2151214SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1031c2151214SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 1032a13144ffSStefano Zampini ierr = PetscFree(extrow);CHKERRQ(ierr); 1033a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&vl2g);CHKERRQ(ierr); 1034c2151214SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&el2g);CHKERRQ(ierr); 1035c2151214SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&fl2g);CHKERRQ(ierr); 1036a13144ffSStefano Zampini ierr = PetscFree2(work,rwork);CHKERRQ(ierr); 1037a13144ffSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1038a13144ffSStefano Zampini ierr = PetscFree(extrows);CHKERRQ(ierr); 1039a13144ffSStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 1040a13144ffSStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 1041a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 1042a13144ffSStefano Zampini ierr = MatDestroy(&conn);CHKERRQ(ierr); 1043a13144ffSStefano Zampini 1044a13144ffSStefano Zampini /* Complete assembling */ 1045a13144ffSStefano Zampini ierr = MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1046a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1047a13144ffSStefano Zampini ierr = VecAssemblyEnd(quads[i]);CHKERRQ(ierr); 1048a13144ffSStefano Zampini } 1049a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1050a13144ffSStefano Zampini ierr = VecDestroy(&quads[i]);CHKERRQ(ierr); 1051a13144ffSStefano Zampini } 1052a13144ffSStefano Zampini ierr = PetscFree(quads);CHKERRQ(ierr); 1053a13144ffSStefano Zampini 1054a13144ffSStefano Zampini /* set change of basis */ 1055a13144ffSStefano Zampini ierr = PCBDDCSetChangeOfBasisMat(pc,T,PETSC_FALSE);CHKERRQ(ierr); 1056a13144ffSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1057a13144ffSStefano Zampini 1058a13144ffSStefano Zampini /* set quadratures */ 1059a13144ffSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,nnsp);CHKERRQ(ierr); 1060a13144ffSStefano Zampini ierr = MatNullSpaceDestroy(&nnsp);CHKERRQ(ierr); 1061a13144ffSStefano Zampini 1062a13144ffSStefano Zampini PetscFunctionReturn(0); 1063a13144ffSStefano Zampini } 1064a13144ffSStefano Zampini 1065d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 1066d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 1067d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 1068d8203eabSStefano Zampini #undef __FUNCT__ 1069d8203eabSStefano Zampini #define __FUNCT__ "PCBDDCNullSpaceCreate" 1070d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 1071d8203eabSStefano Zampini { 1072d8203eabSStefano Zampini PetscErrorCode ierr; 1073d8203eabSStefano Zampini PetscInt i; 1074d8203eabSStefano Zampini 1075d8203eabSStefano Zampini PetscFunctionBegin; 1076d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 1077d8203eabSStefano Zampini PetscInt first,last; 1078d8203eabSStefano Zampini 1079d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 108086fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 1081d8203eabSStefano Zampini if (i>=first && i < last) { 1082d8203eabSStefano Zampini PetscScalar *data; 1083d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1084d8203eabSStefano Zampini if (!has_const) { 1085d8203eabSStefano Zampini data[i-first] = 1.; 1086d8203eabSStefano Zampini } else { 108786fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 108886fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 1089d8203eabSStefano Zampini } 1090d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1091d8203eabSStefano Zampini } 1092d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1093d8203eabSStefano Zampini } 1094d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 1095d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 1096d8203eabSStefano Zampini PetscInt first,last; 1097d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 1098d8203eabSStefano Zampini if (i>=first && i < last) { 1099d8203eabSStefano Zampini PetscScalar *data; 1100d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1101d8203eabSStefano Zampini if (!has_const) { 1102d8203eabSStefano Zampini data[i-first] = 0.; 1103d8203eabSStefano Zampini } else { 110486fa73c5SStefano Zampini data[2*i-first] = 0.; 110586fa73c5SStefano Zampini data[2*i-first+1] = 0.; 1106d8203eabSStefano Zampini } 1107d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1108d8203eabSStefano Zampini } 1109d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1110d8203eabSStefano Zampini } 1111d8203eabSStefano Zampini PetscFunctionReturn(0); 1112d8203eabSStefano Zampini } 1113d8203eabSStefano Zampini 1114669cc0f4SStefano Zampini #undef __FUNCT__ 1115669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 11168ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 1117669cc0f4SStefano Zampini { 1118a198735bSStefano Zampini Mat loc_divudotp; 1119fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 11208ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 1121669cc0f4SStefano Zampini IS *faces,*edges; 1122669cc0f4SStefano Zampini PetscScalar *vals; 1123669cc0f4SStefano Zampini const PetscScalar *array; 1124669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 11251ae86dd6SStefano Zampini PetscMPIInt rank; 1126a198735bSStefano Zampini PetscErrorCode ierr; 1127669cc0f4SStefano Zampini 1128669cc0f4SStefano Zampini PetscFunctionBegin; 1129669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1130669cc0f4SStefano Zampini if (graph->twodim) { 1131669cc0f4SStefano Zampini lmaxneighs = 2; 1132669cc0f4SStefano Zampini } else { 1133669cc0f4SStefano Zampini lmaxneighs = 1; 1134669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1135669cc0f4SStefano Zampini const PetscInt *idxs; 1136669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1137669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 1138669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1139669cc0f4SStefano Zampini } 1140669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 1141669cc0f4SStefano Zampini } 1142669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1143669cc0f4SStefano Zampini maxsize = 0; 1144669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1145669cc0f4SStefano Zampini PetscInt nn; 1146669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1147669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1148669cc0f4SStefano Zampini } 1149669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1150669cc0f4SStefano Zampini PetscInt nn; 1151669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1152669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1153669cc0f4SStefano Zampini } 1154669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 1155669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 1156669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 11578ae0ca82SStefano Zampini if (!transpose) { 11588ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 11598ae0ca82SStefano Zampini } else { 11608ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 11618ae0ca82SStefano Zampini } 1162669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 11631ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 1164d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 1165669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 11668ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 1167669cc0f4SStefano Zampini } 1168d8203eabSStefano Zampini 1169669cc0f4SStefano Zampini /* compute local quad vec */ 1170a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 11718ae0ca82SStefano Zampini if (!transpose) { 1172a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 11738ae0ca82SStefano Zampini } else { 11748ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 11758ae0ca82SStefano Zampini } 1176669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 11778ae0ca82SStefano Zampini if (!transpose) { 1178a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 11798ae0ca82SStefano Zampini } else { 11808ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 11818ae0ca82SStefano Zampini } 1182fa23a32eSStefano Zampini if (vl2l) { 1183fa23a32eSStefano Zampini ierr = VecGetSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1184fa23a32eSStefano Zampini } else { 1185fa23a32eSStefano Zampini vins = v; 1186fa23a32eSStefano Zampini } 1187fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 1188669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 11899a962809SStefano Zampini 11901ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 11911ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 1192669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1193669cc0f4SStefano Zampini const PetscInt *idxs; 1194669cc0f4SStefano Zampini PetscInt idx,nn,j; 1195669cc0f4SStefano Zampini 1196669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 1197669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1198669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 11991ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1200669cc0f4SStefano Zampini idx = -(idx+1); 1201669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1202669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 1203669cc0f4SStefano Zampini } 1204669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1205669cc0f4SStefano Zampini const PetscInt *idxs; 1206669cc0f4SStefano Zampini PetscInt idx,nn,j; 1207669cc0f4SStefano Zampini 1208669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1209669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1210669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 12111ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1212669cc0f4SStefano Zampini idx = -(idx+1); 1213669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1214669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1215669cc0f4SStefano Zampini } 1216c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1217fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 1218fa23a32eSStefano Zampini if (vl2l) { 1219fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1220fa23a32eSStefano Zampini } 1221669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1222669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1223669cc0f4SStefano Zampini 1224669cc0f4SStefano Zampini /* assemble near null space */ 1225669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1226669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 1227669cc0f4SStefano Zampini } 1228669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1229669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 1230669cc0f4SStefano Zampini } 1231669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 1232669cc0f4SStefano Zampini PetscFunctionReturn(0); 1233669cc0f4SStefano Zampini } 1234669cc0f4SStefano Zampini 1235669cc0f4SStefano Zampini 1236a3df083aSStefano Zampini #undef __FUNCT__ 12371f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 12381f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 12391f4df5f7SStefano Zampini { 12401f4df5f7SStefano Zampini PetscErrorCode ierr; 12411f4df5f7SStefano Zampini Vec local,global; 12421f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12431f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 12441f4df5f7SStefano Zampini 12451f4df5f7SStefano Zampini PetscFunctionBegin; 12461f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 12471f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 12481f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 12491f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 12501f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 12511f4df5f7SStefano Zampini PetscInt i; 12521f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 12531f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 12541f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 12551f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 12561f4df5f7SStefano Zampini } 12571f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 12581f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 12591f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 12601f4df5f7SStefano Zampini } 12611f4df5f7SStefano Zampini } else { 1262986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 12631f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1264986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 1265986cdee1SStefano Zampini if (i > 1) { 1266986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 12671f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 12681f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 12691f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 12701f4df5f7SStefano Zampini } 12711f4df5f7SStefano Zampini } 12721f4df5f7SStefano Zampini } 1273986cdee1SStefano Zampini } 12741f4df5f7SStefano Zampini 12751f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 12761f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 12771f4df5f7SStefano Zampini } 12781f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 12791f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 12801f4df5f7SStefano Zampini } 12811f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 12821f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 12831f4df5f7SStefano Zampini } 12841f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 12851f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 12861f4df5f7SStefano Zampini PetscFunctionReturn(0); 12871f4df5f7SStefano Zampini } 12881f4df5f7SStefano Zampini 12891f4df5f7SStefano Zampini #undef __FUNCT__ 12903e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 12913e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 12923e589ea0SStefano Zampini { 12933e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 12943e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 12953e589ea0SStefano Zampini PetscErrorCode ierr; 12963e589ea0SStefano Zampini 12973e589ea0SStefano Zampini PetscFunctionBegin; 12983e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 12993e589ea0SStefano Zampini PetscFunctionReturn(0); 13003e589ea0SStefano Zampini } 13013e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 13023e589ea0SStefano Zampini Vec swap; 13033e589ea0SStefano Zampini 13043e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 13053e589ea0SStefano Zampini swap = pcbddc->work_change; 13063e589ea0SStefano Zampini pcbddc->work_change = r; 13073e589ea0SStefano Zampini r = swap; 13083e589ea0SStefano Zampini } 13093e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 13103e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 13113e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 13123e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 13133e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 13143e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 13153e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1316f913dca9SStefano Zampini pcbddc->work_change = r; 13173e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 13183e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 13193e589ea0SStefano Zampini } 13203e589ea0SStefano Zampini PetscFunctionReturn(0); 13213e589ea0SStefano Zampini } 13223e589ea0SStefano Zampini 13233e589ea0SStefano Zampini #undef __FUNCT__ 1324a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 1325a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1326a3df083aSStefano Zampini { 1327a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1328a3df083aSStefano Zampini PetscErrorCode ierr; 1329a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1330a3df083aSStefano Zampini 1331a3df083aSStefano Zampini PetscFunctionBegin; 1332a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1333a3df083aSStefano Zampini if (transpose) { 1334a3df083aSStefano Zampini apply_right = ctx->apply_left; 1335a3df083aSStefano Zampini apply_left = ctx->apply_right; 1336a3df083aSStefano Zampini } else { 1337a3df083aSStefano Zampini apply_right = ctx->apply_right; 1338a3df083aSStefano Zampini apply_left = ctx->apply_left; 1339a3df083aSStefano Zampini } 1340a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1341a3df083aSStefano Zampini if (apply_right) { 1342a3df083aSStefano Zampini const PetscScalar *ax; 1343a3df083aSStefano Zampini PetscInt nl,i; 1344a3df083aSStefano Zampini 1345a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1346a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1347a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 1348a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1349a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1350a3df083aSStefano Zampini PetscScalar sum,val; 1351a3df083aSStefano Zampini const PetscInt *idxs; 1352a3df083aSStefano Zampini PetscInt nz,j; 1353a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1354a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1355a3df083aSStefano Zampini sum = 0.; 1356a3df083aSStefano Zampini if (ctx->apply_p0) { 1357a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1358a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1359a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1360a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1361a3df083aSStefano Zampini } 1362a3df083aSStefano Zampini } else { 1363a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1364a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1365a3df083aSStefano Zampini } 1366a3df083aSStefano Zampini } 1367a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1368a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1369a3df083aSStefano Zampini } 1370a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1371a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1372a3df083aSStefano Zampini } 1373a3df083aSStefano Zampini if (transpose) { 1374a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1375a3df083aSStefano Zampini } else { 1376a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1377a3df083aSStefano Zampini } 1378a3df083aSStefano Zampini if (reset_x) { 1379a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1380a3df083aSStefano Zampini } 1381a3df083aSStefano Zampini if (apply_left) { 1382a3df083aSStefano Zampini PetscScalar *ay; 1383a3df083aSStefano Zampini PetscInt i; 1384a3df083aSStefano Zampini 1385a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1386a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1387a3df083aSStefano Zampini PetscScalar sum,val; 1388a3df083aSStefano Zampini const PetscInt *idxs; 1389a3df083aSStefano Zampini PetscInt nz,j; 1390a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1391a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1392a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 1393a3df083aSStefano Zampini if (ctx->apply_p0) { 1394a3df083aSStefano Zampini sum = 0.; 1395a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1396a3df083aSStefano Zampini sum += ay[idxs[j]]; 1397a3df083aSStefano Zampini ay[idxs[j]] += val; 1398a3df083aSStefano Zampini } 1399a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 1400a3df083aSStefano Zampini } else { 1401a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1402a3df083aSStefano Zampini ay[idxs[j]] += val; 1403a3df083aSStefano Zampini } 1404a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 1405a3df083aSStefano Zampini } 1406a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1407a3df083aSStefano Zampini } 1408a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 1409a3df083aSStefano Zampini } 1410a3df083aSStefano Zampini PetscFunctionReturn(0); 1411a3df083aSStefano Zampini } 1412a3df083aSStefano Zampini 1413a3df083aSStefano Zampini #undef __FUNCT__ 1414a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 1415a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 1416a3df083aSStefano Zampini { 1417a3df083aSStefano Zampini PetscErrorCode ierr; 1418a3df083aSStefano Zampini 1419a3df083aSStefano Zampini PetscFunctionBegin; 1420a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 1421a3df083aSStefano Zampini PetscFunctionReturn(0); 1422a3df083aSStefano Zampini } 1423a3df083aSStefano Zampini 1424a3df083aSStefano Zampini #undef __FUNCT__ 1425a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 1426a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 1427a3df083aSStefano Zampini { 1428a3df083aSStefano Zampini PetscErrorCode ierr; 1429a3df083aSStefano Zampini 1430a3df083aSStefano Zampini PetscFunctionBegin; 1431a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 1432a3df083aSStefano Zampini PetscFunctionReturn(0); 1433a3df083aSStefano Zampini } 1434a3df083aSStefano Zampini 1435a3df083aSStefano Zampini #undef __FUNCT__ 1436a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 1437a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 1438a3df083aSStefano Zampini { 1439a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1440a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1441a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1442a3df083aSStefano Zampini PetscErrorCode ierr; 1443a3df083aSStefano Zampini 1444a3df083aSStefano Zampini PetscFunctionBegin; 1445a3df083aSStefano Zampini if (!restore) { 14461dd7afcfSStefano Zampini Mat A_IB,A_BI; 1447a3df083aSStefano Zampini PetscScalar *work; 1448b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 1449a3df083aSStefano Zampini 14509a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 14519a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 1452a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 1453a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 1454a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1455a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 1456a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 1457a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 1458a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 1459a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 1460a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 1461a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 1462a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 1463a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 1464059032f7SStefano Zampini if (reuse) { 1465a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 14661dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 1467059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 1468059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 1469059032f7SStefano Zampini PetscInt i; 1470059032f7SStefano Zampini 1471059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 1472059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1473059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1474059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1475059032f7SStefano Zampini } 1476059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 14771dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 1478059032f7SStefano Zampini } 1479a3df083aSStefano Zampini ctx->A = pcis->A_IB; 1480a3df083aSStefano Zampini ctx->work = work; 1481a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 1482a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1483a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1484a3df083aSStefano Zampini pcis->A_IB = A_IB; 1485a3df083aSStefano Zampini 1486a3df083aSStefano Zampini /* A_BI as A_IB^T */ 1487a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 1488a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 1489a3df083aSStefano Zampini pcis->A_BI = A_BI; 1490a3df083aSStefano Zampini } else { 14911dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 14921dd7afcfSStefano Zampini PetscFunctionReturn(0); 14931dd7afcfSStefano Zampini } 1494a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 1495a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 1496a3df083aSStefano Zampini pcis->A_IB = ctx->A; 14971dd7afcfSStefano Zampini ctx->A = NULL; 14981dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 14991dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 15001dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 15011dd7afcfSStefano Zampini if (ctx->free) { 1502059032f7SStefano Zampini PetscInt i; 15031dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1504059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1505059032f7SStefano Zampini } 1506059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1507059032f7SStefano Zampini } 1508a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 1509a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 1510a3df083aSStefano Zampini } 1511a3df083aSStefano Zampini PetscFunctionReturn(0); 1512a3df083aSStefano Zampini } 1513a3df083aSStefano Zampini 1514a3df083aSStefano Zampini /* used just in bddc debug mode */ 1515a3df083aSStefano Zampini #undef __FUNCT__ 1516a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 1517a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 1518a3df083aSStefano Zampini { 1519a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1520a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1521a3df083aSStefano Zampini Mat An; 1522a3df083aSStefano Zampini PetscErrorCode ierr; 1523a3df083aSStefano Zampini 1524a3df083aSStefano Zampini PetscFunctionBegin; 1525a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 1526a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 1527a3df083aSStefano Zampini if (is1) { 1528a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 1529a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 1530a3df083aSStefano Zampini } else { 1531a3df083aSStefano Zampini *B = An; 1532a3df083aSStefano Zampini } 1533a3df083aSStefano Zampini PetscFunctionReturn(0); 1534a3df083aSStefano Zampini } 1535a3df083aSStefano Zampini 15361cf9b237SStefano Zampini /* TODO: add reuse flag */ 15371cf9b237SStefano Zampini #undef __FUNCT__ 15381cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 15391cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 15401cf9b237SStefano Zampini { 15411cf9b237SStefano Zampini Mat Bt; 15421cf9b237SStefano Zampini PetscScalar *a,*bdata; 15431cf9b237SStefano Zampini const PetscInt *ii,*ij; 15441cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 15451cf9b237SStefano Zampini PetscBool flg_row; 15461cf9b237SStefano Zampini PetscErrorCode ierr; 15471cf9b237SStefano Zampini 15481cf9b237SStefano Zampini PetscFunctionBegin; 15491cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 15501cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 15511cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 15521cf9b237SStefano Zampini nnz = n; 15531cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 15541cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 15551cf9b237SStefano Zampini } 15561cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 15571cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 15581cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 15591cf9b237SStefano Zampini nnz = 0; 15601cf9b237SStefano Zampini bii[0] = 0; 15611cf9b237SStefano Zampini for (i=0;i<n;i++) { 15621cf9b237SStefano Zampini PetscInt j; 15631cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 15641cf9b237SStefano Zampini PetscScalar entry = a[j]; 15651cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 15661cf9b237SStefano Zampini bij[nnz] = ij[j]; 15671cf9b237SStefano Zampini bdata[nnz] = entry; 15681cf9b237SStefano Zampini nnz++; 15691cf9b237SStefano Zampini } 15701cf9b237SStefano Zampini } 15711cf9b237SStefano Zampini bii[i+1] = nnz; 15721cf9b237SStefano Zampini } 15731cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 15741cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 15751cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 15761cf9b237SStefano Zampini { 15771cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 15781cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 15791cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 15801cf9b237SStefano Zampini } 15811cf9b237SStefano Zampini *B = Bt; 15821cf9b237SStefano Zampini PetscFunctionReturn(0); 15831cf9b237SStefano Zampini } 15841cf9b237SStefano Zampini 1585674ae819SStefano Zampini #undef __FUNCT__ 15864f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 15874f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 15884f1b2e48SStefano Zampini { 15894f1b2e48SStefano Zampini Mat B; 15904f1b2e48SStefano Zampini IS is_dummy,*cc_n; 15914f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 15924f1b2e48SStefano Zampini PCBDDCGraph graph; 15934f1b2e48SStefano Zampini PetscInt i,n; 15944f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 15954f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 15964f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 15974f1b2e48SStefano Zampini PetscErrorCode ierr; 15984f1b2e48SStefano Zampini 15994f1b2e48SStefano Zampini PetscFunctionBegin; 160063c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 160163c961adSStefano Zampini *ncc = 0; 160263c961adSStefano Zampini *cc = NULL; 160363c961adSStefano Zampini PetscFunctionReturn(0); 160463c961adSStefano Zampini } 16054f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 16064f1b2e48SStefano Zampini if (!isseqaij && filter) { 16071cf9b237SStefano Zampini PetscBool isseqdense; 16081cf9b237SStefano Zampini 16091cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 16101cf9b237SStefano Zampini if (!isseqdense) { 16114f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 16121cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 16131cf9b237SStefano Zampini PetscScalar *array; 16141cf9b237SStefano Zampini PetscReal chop=1.e-6; 16151cf9b237SStefano Zampini 16161cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 16171cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 16181cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 16191cf9b237SStefano Zampini for (i=0;i<n;i++) { 16201cf9b237SStefano Zampini PetscInt j; 16211cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 16221cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 16231cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 16241cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 16251cf9b237SStefano Zampini } 16261cf9b237SStefano Zampini } 16271cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 16289d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 16291cf9b237SStefano Zampini } 16304f1b2e48SStefano Zampini } else { 16314f1b2e48SStefano Zampini B = A; 16324f1b2e48SStefano Zampini } 16334f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 16344f1b2e48SStefano Zampini 16354f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 16364f1b2e48SStefano Zampini if (filter) { 16374f1b2e48SStefano Zampini PetscScalar *data; 16384f1b2e48SStefano Zampini PetscInt j,cum; 16394f1b2e48SStefano Zampini 16404f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 16414f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 16424f1b2e48SStefano Zampini cum = 0; 16434f1b2e48SStefano Zampini for (i=0;i<n;i++) { 16444f1b2e48SStefano Zampini PetscInt t; 16454f1b2e48SStefano Zampini 16464f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 16474f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 16484f1b2e48SStefano Zampini continue; 16494f1b2e48SStefano Zampini } 16504f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 16514f1b2e48SStefano Zampini } 16524f1b2e48SStefano Zampini t = xadj_filtered[i]; 16534f1b2e48SStefano Zampini xadj_filtered[i] = cum; 16544f1b2e48SStefano Zampini cum += t; 16554f1b2e48SStefano Zampini } 16564f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 16574f1b2e48SStefano Zampini } else { 16584f1b2e48SStefano Zampini xadj_filtered = NULL; 16594f1b2e48SStefano Zampini adjncy_filtered = NULL; 16604f1b2e48SStefano Zampini } 16614f1b2e48SStefano Zampini 16624f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 16634f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 16644f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 16654f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 16664f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 1667be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 16684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 16694f1b2e48SStefano Zampini if (xadj_filtered) { 16704f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 16714f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 16724f1b2e48SStefano Zampini } else { 16734f1b2e48SStefano Zampini graph->xadj = xadj; 16744f1b2e48SStefano Zampini graph->adjncy = adjncy; 16754f1b2e48SStefano Zampini } 16764f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 16774f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 16784f1b2e48SStefano Zampini /* partial clean up */ 16794f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 16804f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 16811cf9b237SStefano Zampini if (A != B) { 16824f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 16834f1b2e48SStefano Zampini } 16844f1b2e48SStefano Zampini 16854f1b2e48SStefano Zampini /* get back data */ 16861cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 16871cf9b237SStefano Zampini if (cc) { 16884f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 16894f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 16904f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 16914f1b2e48SStefano Zampini } 16924f1b2e48SStefano Zampini *cc = cc_n; 16931cf9b237SStefano Zampini } 16944f1b2e48SStefano Zampini /* clean up graph */ 16954f1b2e48SStefano Zampini graph->xadj = 0; 16964f1b2e48SStefano Zampini graph->adjncy = 0; 16974f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 16984f1b2e48SStefano Zampini PetscFunctionReturn(0); 16994f1b2e48SStefano Zampini } 17004f1b2e48SStefano Zampini 17014f1b2e48SStefano Zampini #undef __FUNCT__ 17025408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 17035408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 17045408967cSStefano Zampini { 17055408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 17065408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1707dee84bffSStefano Zampini IS dirIS = NULL; 17084f1b2e48SStefano Zampini PetscInt i; 17095408967cSStefano Zampini PetscErrorCode ierr; 17105408967cSStefano Zampini 17115408967cSStefano Zampini PetscFunctionBegin; 1712dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 17135408967cSStefano Zampini if (zerodiag) { 17145408967cSStefano Zampini Mat A; 17155408967cSStefano Zampini Vec vec3_N; 17165408967cSStefano Zampini PetscScalar *vals; 17175408967cSStefano Zampini const PetscInt *idxs; 1718d12d3064SStefano Zampini PetscInt nz,*count; 17195408967cSStefano Zampini 17205408967cSStefano Zampini /* p0 */ 17215408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 17225408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 17235408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 17245408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 17254f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 17265408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 17275408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 17285408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 17295408967cSStefano Zampini /* v_I */ 17305408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 17315408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 17325408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 17335408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 17345408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 17355408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 17365408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 17375408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 17385408967cSStefano Zampini if (dirIS) { 17395408967cSStefano Zampini PetscInt n; 17405408967cSStefano Zampini 17415408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 17425408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 17435408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 17445408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 17455408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 17465408967cSStefano Zampini } 17475408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 17485408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 17495408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 17505408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 1751669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 17525408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 17535408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 17549a962809SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)",PetscAbsScalar(vals[0])); 17555408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 17565408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 1757d12d3064SStefano Zampini 1758d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 1759d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 1760d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 1761d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 1762d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 1763d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 17649a962809SStefano Zampini for (i=0;i<nz;i++) if (count[idxs[i]]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 1765d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 1766d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 17675408967cSStefano Zampini } 1768dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 17695408967cSStefano Zampini 17705408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 17715408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 17724f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 17735408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 17744f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 17755408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 1776f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1777f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 1778f2a566d8SStefano Zampini if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %d instead of %g\n",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i);CHKERRQ(ierr); 1779f2a566d8SStefano Zampini } 17805408967cSStefano Zampini PetscFunctionReturn(0); 17815408967cSStefano Zampini } 17825408967cSStefano Zampini 17835408967cSStefano Zampini #undef __FUNCT__ 1784339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 1785339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 1786339f8db1SStefano Zampini { 1787339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 17884f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 1789b0f5fe93SStefano Zampini PetscInt nz,n; 17901f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 17914f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 1792339f8db1SStefano Zampini PetscErrorCode ierr; 1793339f8db1SStefano Zampini 1794339f8db1SStefano Zampini PetscFunctionBegin; 17959f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 17969f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 1797a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 1798a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 1799a3df083aSStefano Zampini } 1800a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1801a3df083aSStefano Zampini pcbddc->benign_n = 0; 18021ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 18034f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 18044f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 18054f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 18061ae86dd6SStefano Zampini since the local Schur complements are already SPD 18074f1b2e48SStefano Zampini */ 18084f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 18094f1b2e48SStefano Zampini have_null = PETSC_TRUE; 181040fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 18114f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 18124f1b2e48SStefano Zampini 18134f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 18144f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 18154f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 18164f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 1817ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 181840fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 181940fa8d13SStefano Zampini if (!sorted) { 182040fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 182140fa8d13SStefano Zampini } 182240fa8d13SStefano Zampini } else { 182340fa8d13SStefano Zampini pressures = NULL; 182440fa8d13SStefano Zampini } 182597d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 182697d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 182727b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 182897d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 1829339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 1830339f8db1SStefano Zampini if (!sorted) { 1831339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 1832339f8db1SStefano Zampini } 1833339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 18344f1b2e48SStefano Zampini if (!nz) { 18354f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 18364f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 183740fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 183840fa8d13SStefano Zampini } 18394f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 18404f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 18414f1b2e48SStefano Zampini zerodiag_subs = NULL; 18424f1b2e48SStefano Zampini pcbddc->benign_n = 0; 18431f4df5f7SStefano Zampini n_interior_dofs = 0; 18441f4df5f7SStefano Zampini interior_dofs = NULL; 18451f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 18461f4df5f7SStefano Zampini PetscInt n,i,j; 18471f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 18481f4df5f7SStefano Zampini PetscInt *iwork; 18491f4df5f7SStefano Zampini 18501f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 18511f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 18521f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 18531f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 185490648384SStefano Zampini for (i=1;i<n_neigh;i++) 18551f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 18561f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 18571f4df5f7SStefano Zampini for (i=0;i<n;i++) 18581f4df5f7SStefano Zampini if (!iwork[i]) 18591f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 18601f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 18611f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 18621f4df5f7SStefano Zampini } 18634f1b2e48SStefano Zampini if (has_null_pressures) { 18644f1b2e48SStefano Zampini IS *subs; 18651f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 18661f4df5f7SStefano Zampini const PetscInt *idxs; 18671f4df5f7SStefano Zampini PetscScalar *array; 18681f4df5f7SStefano Zampini Vec *work; 18691f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 18704f1b2e48SStefano Zampini 18714f1b2e48SStefano Zampini subs = pcbddc->local_subs; 18724f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 18731f4df5f7SStefano Zampini /* these vectors are needed to check if the constant on pressures is in the kernel of the local operator B (i.e. B(v_I,p0) should be zero) */ 18741f4df5f7SStefano Zampini if (pcbddc->current_level) { 18751f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 18761f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 18771f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 18781f4df5f7SStefano Zampini /* work[0] = 1_p */ 18791f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 18801f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 18811f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 18821f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 18831f4df5f7SStefano Zampini /* work[0] = 1_v */ 18841f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 18851f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 18861f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 18871f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 18881f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 18891f4df5f7SStefano Zampini } 18904f1b2e48SStefano Zampini if (nsubs > 1) { 18914f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 18924f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 18934f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 18944f1b2e48SStefano Zampini IS t_zerodiag_subs; 18954f1b2e48SStefano Zampini PetscInt nl; 18964f1b2e48SStefano Zampini 18974f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 18984f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 18994f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 19004f1b2e48SStefano Zampini if (nl) { 19014f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 19024f1b2e48SStefano Zampini 19031f4df5f7SStefano Zampini if (pcbddc->current_level) { 19041f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 19051f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 19061f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 19071f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 19081f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 19091f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 19101f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 19111f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 19121f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 19131f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 19141f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 19151f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 19161f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 19171f4df5f7SStefano Zampini valid = PETSC_FALSE; 19181f4df5f7SStefano Zampini break; 19191f4df5f7SStefano Zampini } 19201f4df5f7SStefano Zampini } 19211f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 19221f4df5f7SStefano Zampini } 19231f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 19241f4df5f7SStefano Zampini IS t_bc; 19251f4df5f7SStefano Zampini PetscInt nzb; 19261f4df5f7SStefano Zampini 19271f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 19281f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 19291f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 19301f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 19311f4df5f7SStefano Zampini } 19321f4df5f7SStefano Zampini if (valid && pressures) { 19334f1b2e48SStefano Zampini IS t_pressure_subs; 19344f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 19354f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 19364f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 19374f1b2e48SStefano Zampini } 19384f1b2e48SStefano Zampini if (valid) { 19394f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 19404f1b2e48SStefano Zampini pcbddc->benign_n++; 19414f1b2e48SStefano Zampini } else { 19424f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 19434f1b2e48SStefano Zampini } 19444f1b2e48SStefano Zampini } 19454f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 19464f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 19474f1b2e48SStefano Zampini } 19484f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 19494f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 19501f4df5f7SStefano Zampini 19511f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 19521f4df5f7SStefano Zampini PetscInt nzb; 19531f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 19541f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 19551f4df5f7SStefano Zampini } 19561f4df5f7SStefano Zampini if (valid && pressures) { 19574f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 19584f1b2e48SStefano Zampini } 19591f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 19601f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 19611f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 19621f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 19631f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 19641f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 19651f4df5f7SStefano Zampini valid = PETSC_FALSE; 19661f4df5f7SStefano Zampini break; 19671f4df5f7SStefano Zampini } 19681f4df5f7SStefano Zampini } 19691f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 19701f4df5f7SStefano Zampini } 19714f1b2e48SStefano Zampini if (valid) { 19724f1b2e48SStefano Zampini pcbddc->benign_n = 1; 1973ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 19744f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 19754f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 19764f1b2e48SStefano Zampini } 19774f1b2e48SStefano Zampini } 19781f4df5f7SStefano Zampini if (pcbddc->current_level) { 19791f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 19804f1b2e48SStefano Zampini } 19811f4df5f7SStefano Zampini } 19821f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 19834f1b2e48SStefano Zampini 19844f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 1985b9b0e38cSStefano Zampini PetscInt n; 1986b9b0e38cSStefano Zampini 19874f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 19884f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 1989b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 1990b9b0e38cSStefano Zampini if (n) { 19914f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 19924f1b2e48SStefano Zampini have_null = PETSC_FALSE; 19934f1b2e48SStefano Zampini } 1994b9b0e38cSStefano Zampini } 19954f1b2e48SStefano Zampini 19964f1b2e48SStefano Zampini /* final check for null pressures */ 19974f1b2e48SStefano Zampini if (zerodiag && pressures) { 19984f1b2e48SStefano Zampini PetscInt nz,np; 19994f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 20004f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 20014f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 20024f1b2e48SStefano Zampini } 20034f1b2e48SStefano Zampini 20044f1b2e48SStefano Zampini if (recompute_zerodiag) { 20054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 20064f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 20074f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 20084f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 20094f1b2e48SStefano Zampini } else { 20104f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 20114f1b2e48SStefano Zampini 20124f1b2e48SStefano Zampini nzn = 0; 20134f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20144f1b2e48SStefano Zampini PetscInt ns; 20154f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 20164f1b2e48SStefano Zampini nzn += ns; 20174f1b2e48SStefano Zampini } 20184f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 20194f1b2e48SStefano Zampini nzn = 0; 20204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 20214f1b2e48SStefano Zampini PetscInt ns,*idxs; 20224f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 20234f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 20244f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 20254f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 20264f1b2e48SStefano Zampini nzn += ns; 20274f1b2e48SStefano Zampini } 20284f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 20294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 20304f1b2e48SStefano Zampini } 20314f1b2e48SStefano Zampini have_null = PETSC_FALSE; 20324f1b2e48SStefano Zampini } 20334f1b2e48SStefano Zampini 2034669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2035a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2036a198735bSStefano Zampini Mat A,loc_divudotp; 2037a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2038a198735bSStefano Zampini IS row,col,isused = NULL; 2039a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2040a198735bSStefano Zampini 20411f4df5f7SStefano Zampini if (pressures) { 20421f4df5f7SStefano Zampini isused = pressures; 20431f4df5f7SStefano Zampini } else { 20441f4df5f7SStefano Zampini isused = zerodiag; 20451f4df5f7SStefano Zampini } 2046a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2047669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 20481ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 20491ae86dd6SStefano Zampini if (!isused && n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Don't know how to extract div u dot p! Please provide the pressure field"); 2050a198735bSStefano Zampini n_isused = 0; 2051a198735bSStefano Zampini if (isused) { 2052a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2053a198735bSStefano Zampini } 2054a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2055a198735bSStefano Zampini st = st-n_isused; 20561ae86dd6SStefano Zampini if (n) { 2057a198735bSStefano Zampini const PetscInt *gidxs; 2058a198735bSStefano Zampini 2059a198735bSStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2060a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2061a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2062a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2063a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2064a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 20651ae86dd6SStefano Zampini } else { 2066a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2067a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2068a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2069a198735bSStefano Zampini } 2070a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2071a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2072a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2073a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2074a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2075a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2076a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2077a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2078a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2079a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2080a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2081a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2082a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2083a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 20841ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 20851ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 20861ae86dd6SStefano Zampini } 2087b3afcdbeSStefano Zampini 2088b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 20894f1b2e48SStefano Zampini if (has_null_pressures) { 20904f1b2e48SStefano Zampini IS zerodiagc; 20914f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 20924f1b2e48SStefano Zampini PetscInt i,s,*nnz; 20934f1b2e48SStefano Zampini 20944f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2095339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 2096339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 2097339f8db1SStefano Zampini /* local change of basis for pressures */ 2098339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 209997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2100339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2101339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2102339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 21034f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 21044f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 21054f1b2e48SStefano Zampini PetscInt nzs,j; 21064f1b2e48SStefano Zampini 21074f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 21084f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 21094f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 21104f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 21114f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 21124f1b2e48SStefano Zampini } 2113339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2114339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2115339f8db1SStefano Zampini /* set identity on velocities */ 2116339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 2117339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 2118339f8db1SStefano Zampini } 21194f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 21204f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 21219f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 21224f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 2123339f8db1SStefano Zampini /* set change on pressures */ 21244f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 21254f1b2e48SStefano Zampini PetscScalar *array; 21264f1b2e48SStefano Zampini PetscInt nzs; 21274f1b2e48SStefano Zampini 21284f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 21294f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 21304f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2131339f8db1SStefano Zampini PetscScalar vals[2]; 2132339f8db1SStefano Zampini PetscInt cols[2]; 2133339f8db1SStefano Zampini 2134339f8db1SStefano Zampini cols[0] = idxs[i]; 21354f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2136339f8db1SStefano Zampini vals[0] = 1.; 2137b0f5fe93SStefano Zampini vals[1] = 1.; 21384f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2139339f8db1SStefano Zampini } 21404f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 21414f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 21424f1b2e48SStefano Zampini array[nzs-1] = 1.; 21434f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 21444f1b2e48SStefano Zampini /* store local idxs for p0 */ 21454f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 21464f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2147339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 21484f1b2e48SStefano Zampini } 2149339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2150339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2151a3df083aSStefano Zampini /* project if needed */ 2152a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 21531dd7afcfSStefano Zampini Mat M; 21541dd7afcfSStefano Zampini 21551dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2156339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 21571dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 21581dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2159a3df083aSStefano Zampini } 21604f1b2e48SStefano Zampini /* store global idxs for p0 */ 21614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2162339f8db1SStefano Zampini } 2163ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 21644f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 2165b0f5fe93SStefano Zampini 2166b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 2167b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 216827b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 216927b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2170339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2171339f8db1SStefano Zampini PetscFunctionReturn(0); 2172339f8db1SStefano Zampini } 2173339f8db1SStefano Zampini 2174339f8db1SStefano Zampini #undef __FUNCT__ 2175015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 2176015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 2177efc2fbd9SStefano Zampini { 2178efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2179de9d7bd0SStefano Zampini PetscScalar *array; 2180efc2fbd9SStefano Zampini PetscErrorCode ierr; 2181efc2fbd9SStefano Zampini 2182efc2fbd9SStefano Zampini PetscFunctionBegin; 2183efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 2184efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 21854f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2186efc2fbd9SStefano Zampini } 2187de9d7bd0SStefano Zampini if (get) { 2188efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 21894f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 21904f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 2191efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 2192de9d7bd0SStefano Zampini } else { 2193de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 2194de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2195de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2196de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 2197efc2fbd9SStefano Zampini } 2198efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2199efc2fbd9SStefano Zampini } 2200efc2fbd9SStefano Zampini 2201efc2fbd9SStefano Zampini #undef __FUNCT__ 2202c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 2203c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 2204c263805aSStefano Zampini { 2205c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2206c263805aSStefano Zampini PetscErrorCode ierr; 2207c263805aSStefano Zampini 2208c263805aSStefano Zampini PetscFunctionBegin; 2209c263805aSStefano Zampini /* TODO: add error checking 2210c263805aSStefano Zampini - avoid nested pop (or push) calls. 2211c263805aSStefano Zampini - cannot push before pop. 22121c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 2213c263805aSStefano Zampini */ 22144f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2215efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2216efc2fbd9SStefano Zampini } 2217c263805aSStefano Zampini if (pop) { 2218a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 22194f1b2e48SStefano Zampini IS is_p0; 22204f1b2e48SStefano Zampini MatReuse reuse; 2221c263805aSStefano Zampini 2222c263805aSStefano Zampini /* extract B_0 */ 22234f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 22244f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 22254f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 22264f1b2e48SStefano Zampini } 22274f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 22284f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 2229c263805aSStefano Zampini /* remove rows and cols from local problem */ 2230c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 223197d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 22324f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 22334f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 2234a3df083aSStefano Zampini } else { 2235a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2236a3df083aSStefano Zampini PetscScalar *vals; 2237a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 2238a3df083aSStefano Zampini 2239a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 2240a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 2241a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 22420b5adadeSStefano Zampini PetscInt *nnz; 2243a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 2244a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 2245a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2246331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 2247331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2248331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 2249331e053bSStefano Zampini nnz[i] = n - nnz[i]; 2250331e053bSStefano Zampini } 2251331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 2252331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2253331e053bSStefano Zampini } 2254a3df083aSStefano Zampini 2255a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2256a3df083aSStefano Zampini PetscScalar *array; 2257a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 2258a3df083aSStefano Zampini 2259a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 2260a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2261a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2262a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 2263a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2264a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 2265a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 2266a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 2267a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 2268a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 2269a3df083aSStefano Zampini cum = 0; 2270a3df083aSStefano Zampini for (j=0;j<n;j++) { 227122db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 2272a3df083aSStefano Zampini vals[cum] = array[j]; 2273a3df083aSStefano Zampini idxs_ins[cum] = j; 2274a3df083aSStefano Zampini cum++; 2275a3df083aSStefano Zampini } 2276a3df083aSStefano Zampini } 2277a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 2278a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 2279a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2280a3df083aSStefano Zampini } 2281a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2282a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2283a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 2284a3df083aSStefano Zampini } 2285c263805aSStefano Zampini } else { /* push */ 2286a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 22874f1b2e48SStefano Zampini PetscInt i; 22884f1b2e48SStefano Zampini 22894f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 22904f1b2e48SStefano Zampini PetscScalar *B0_vals; 22914f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 22924f1b2e48SStefano Zampini 22934f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 22944f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 22957b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 22964f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 22974f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 22984f1b2e48SStefano Zampini } 2299c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2300c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2301a3df083aSStefano Zampini } else { 2302a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 2303a3df083aSStefano Zampini } 2304c263805aSStefano Zampini } 2305c263805aSStefano Zampini PetscFunctionReturn(0); 2306c263805aSStefano Zampini } 2307c263805aSStefano Zampini 2308c263805aSStefano Zampini #undef __FUNCT__ 2309b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 231008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 2311b1b3d7a2SStefano Zampini { 2312b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 231308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 231408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 231508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 231608122e43SStefano Zampini PetscScalar *work,lwork; 231708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 231808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 231908122e43SStefano Zampini PetscReal *eigs,thresh; 23201b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 2321f6f667cfSStefano Zampini PetscBool allocated_S_St; 232208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 232308122e43SStefano Zampini PetscReal *rwork; 232408122e43SStefano Zampini #endif 2325b1b3d7a2SStefano Zampini PetscErrorCode ierr; 2326b1b3d7a2SStefano Zampini 2327b1b3d7a2SStefano Zampini PetscFunctionBegin; 2328b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 2329af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 2330af25d912SStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) 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); 233106a4e24aSStefano Zampini 2332fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2333fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2334fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2335fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 23361575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2337fd14bc51SStefano Zampini } 2338fd14bc51SStefano Zampini 2339e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 2340e496cd5dSStefano 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); 2341e496cd5dSStefano Zampini } 2342e496cd5dSStefano Zampini 234308122e43SStefano Zampini /* max size of subsets */ 234408122e43SStefano Zampini mss = 0; 234508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 234608122e43SStefano Zampini PetscInt subset_size; 2347862806e4SStefano Zampini 234808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 234908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 235008122e43SStefano Zampini } 235108122e43SStefano Zampini 235208122e43SStefano Zampini /* min/max and threshold */ 235308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 2354f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 235508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 2356f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 2357f6f667cfSStefano Zampini if (nmin) { 2358f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 2359f6f667cfSStefano Zampini } 236008122e43SStefano Zampini 236108122e43SStefano Zampini /* allocate lapack workspace */ 236208122e43SStefano Zampini cum = cum2 = 0; 236308122e43SStefano Zampini maxneigs = 0; 236408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 236508122e43SStefano Zampini PetscInt n,subset_size; 2366f6f667cfSStefano Zampini 236708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 236808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 23699162d606SStefano Zampini cum += subset_size; 23709162d606SStefano Zampini cum2 += subset_size*n; 237108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 237208122e43SStefano Zampini } 237308122e43SStefano Zampini if (mss) { 23749ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 237508122e43SStefano Zampini PetscBLASInt B_itype = 1; 237608122e43SStefano Zampini PetscBLASInt B_N = mss; 23774c6709b3SStefano Zampini PetscReal zero = 0.0; 23784c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 237908122e43SStefano Zampini 238008122e43SStefano Zampini B_lwork = -1; 238108122e43SStefano Zampini S = NULL; 238208122e43SStefano Zampini St = NULL; 2383a58a30b4SStefano Zampini eigs = NULL; 2384a58a30b4SStefano Zampini eigv = NULL; 2385a58a30b4SStefano Zampini B_iwork = NULL; 2386a58a30b4SStefano Zampini B_ifail = NULL; 2387d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2388d1710679SStefano Zampini rwork = NULL; 2389d1710679SStefano Zampini #endif 23908bec7fa6SStefano Zampini thresh = 1.0; 239108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 239208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 239308122e43SStefano 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)); 239408122e43SStefano Zampini #else 239508122e43SStefano 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)); 239608122e43SStefano Zampini #endif 239708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 239808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 239908122e43SStefano Zampini } else { 240008122e43SStefano Zampini /* TODO */ 240108122e43SStefano Zampini } 240208122e43SStefano Zampini } else { 240308122e43SStefano Zampini lwork = 0; 240408122e43SStefano Zampini } 240508122e43SStefano Zampini 240608122e43SStefano Zampini nv = 0; 2407d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 2408d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 240908122e43SStefano Zampini } 24104c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 2411f6f667cfSStefano Zampini if (allocated_S_St) { 2412f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 2413f6f667cfSStefano Zampini } 2414f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 241508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 241608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 241708122e43SStefano Zampini #endif 24189162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 24199162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 24209162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 242108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 24229162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 242308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 242408122e43SStefano Zampini 242508122e43SStefano Zampini maxneigs = 0; 242672b8c272SStefano Zampini cum = cumarray = 0; 24279162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 24289162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 2429d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 243008122e43SStefano Zampini const PetscInt *idxs; 243108122e43SStefano Zampini 2432d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 243308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 243408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 243508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 243608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 24379162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 24389162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 243908122e43SStefano Zampini } 2440d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 244108122e43SStefano Zampini } 244208122e43SStefano Zampini 244308122e43SStefano Zampini if (mss) { /* multilevel */ 244408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 244508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 244608122e43SStefano Zampini } 244708122e43SStefano Zampini 2448ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 244908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 245008122e43SStefano Zampini const PetscInt *idxs; 24519d54b7f4SStefano Zampini PetscReal upper,lower; 2452862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 245308122e43SStefano Zampini PetscBLASInt B_N; 2454aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 245508122e43SStefano Zampini 24569d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 24579d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 24589d54b7f4SStefano Zampini lower = thresh; 24599d54b7f4SStefano Zampini } else { 24609d54b7f4SStefano Zampini upper = 1./thresh; 24619d54b7f4SStefano Zampini lower = 0.; 24629d54b7f4SStefano Zampini } 2463862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 2464ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 2465f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 2466f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 24679ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 2468aff50787SStefano Zampini PetscInt j,k; 2469aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 2470aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 2471aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 247208122e43SStefano Zampini } 247308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 2474aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 2475aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 2476aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 2477aff50787SStefano Zampini } 247808122e43SStefano Zampini } 247908122e43SStefano Zampini } else { 248008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 248108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 248208122e43SStefano Zampini } 24838bec7fa6SStefano Zampini } else { 2484f6f667cfSStefano Zampini S = Sarray + cumarray; 2485f6f667cfSStefano Zampini St = Starray + cumarray; 24868bec7fa6SStefano Zampini } 2487aff50787SStefano Zampini /* see if we can save some work */ 2488b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 2489aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 2490aff50787SStefano Zampini } 2491aff50787SStefano Zampini 2492b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 2493aff50787SStefano Zampini B_neigs = 0; 2494aff50787SStefano Zampini } else { 24959ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 249608122e43SStefano Zampini PetscBLASInt B_itype = 1; 2497f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 24984c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 24999552c7c7SStefano Zampini PetscInt nmin_s; 2500b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 250108122e43SStefano Zampini 2502fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 25038bec7fa6SStefano 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]]); 2504fd14bc51SStefano Zampini } 2505d16cbb6bSStefano Zampini 2506b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 2507b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 2508b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 2509b7ab4a40SStefano Zampini } 2510b7ab4a40SStefano Zampini 251108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2512b7ab4a40SStefano Zampini if (compute_range) { 2513d16cbb6bSStefano Zampini 2514d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 251508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 25169d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 251708122e43SStefano Zampini #else 25189d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 251908122e43SStefano Zampini #endif 2520b7ab4a40SStefano Zampini } else if (!same_data) { 2521d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 2522d16cbb6bSStefano Zampini B_IL = 1; 2523d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 25249d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 2525d16cbb6bSStefano Zampini #else 25269d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 2527d16cbb6bSStefano Zampini #endif 2528b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 2529b7ab4a40SStefano Zampini PetscInt k; 2530b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 2531b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 2532b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 2533b7ab4a40SStefano Zampini nmin = nmax; 2534b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 2535b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 2536b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 2537b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 2538b7ab4a40SStefano Zampini } 2539d16cbb6bSStefano Zampini } 254008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 254108122e43SStefano Zampini if (B_ierr) { 25426c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 25436c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 25446c4ed002SBarry Smith else 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); 254508122e43SStefano Zampini } 254608122e43SStefano Zampini 254708122e43SStefano Zampini if (B_neigs > nmax) { 2548fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2549fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 2550fd14bc51SStefano Zampini } 25519d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 255208122e43SStefano Zampini B_neigs = nmax; 255308122e43SStefano Zampini } 255408122e43SStefano Zampini 25559552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 25569552c7c7SStefano Zampini if (B_neigs < nmin_s) { 255708122e43SStefano Zampini PetscBLASInt B_neigs2; 255808122e43SStefano Zampini 25599d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2560f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 25619d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 25629d54b7f4SStefano Zampini } else { 25639d54b7f4SStefano Zampini B_IL = B_neigs + 1; 25649d54b7f4SStefano Zampini B_IU = nmin_s; 25659d54b7f4SStefano Zampini } 2566fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2567fd14bc51SStefano 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); 2568fd14bc51SStefano Zampini } 25699ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 25701ae86dd6SStefano Zampini PetscInt j,k; 257108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 25721ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 25731ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 25741ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 257508122e43SStefano Zampini } 257608122e43SStefano Zampini } 257708122e43SStefano Zampini } else { 257808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 257908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 258008122e43SStefano Zampini } 258108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 258208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 25839d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&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)); 258408122e43SStefano Zampini #else 25859d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&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)); 258608122e43SStefano Zampini #endif 258708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 258808122e43SStefano Zampini B_neigs += B_neigs2; 258908122e43SStefano Zampini } 259008122e43SStefano Zampini if (B_ierr) { 25916c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 25926c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 25936c4ed002SBarry Smith else 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); 259408122e43SStefano Zampini } 2595fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2596ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 259708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 259808122e43SStefano Zampini if (eigs[j] == 0.0) { 2599ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 260008122e43SStefano Zampini } else { 26019d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2602ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 26039d54b7f4SStefano Zampini } else { 26049d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 26059d54b7f4SStefano Zampini } 2606fd14bc51SStefano Zampini } 260708122e43SStefano Zampini } 260808122e43SStefano Zampini } 260908122e43SStefano Zampini } else { 261008122e43SStefano Zampini /* TODO */ 261108122e43SStefano Zampini } 2612aff50787SStefano Zampini } 26136c3e6151SStefano Zampini /* change the basis back to the original one */ 26146c3e6151SStefano Zampini if (sub_schurs->change) { 261572b8c272SStefano Zampini Mat change,phi,phit; 26166c3e6151SStefano Zampini 26176c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 26186c3e6151SStefano Zampini PetscInt ii; 26196c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 26206c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 26216c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 2622684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 2623684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 2624684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 2625684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 2626684229deSStefano Zampini #else 26276c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 2628684229deSStefano Zampini #endif 26296c3e6151SStefano Zampini } 26306c3e6151SStefano Zampini } 26316c3e6151SStefano Zampini } 263272b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 26336c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 263472b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 26356c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 26366c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 26376c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 26386c3e6151SStefano Zampini } 26398bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 26408bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 26419162d606SStefano Zampini if (B_neigs) { 26429162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 2643fd14bc51SStefano Zampini 2644fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 26459552c7c7SStefano Zampini PetscInt ii; 26469552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 2647ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 26489552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2649ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 2650ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2651ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2652ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 2653ac47001eSStefano Zampini #else 2654ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 2655ac47001eSStefano Zampini #endif 26569552c7c7SStefano Zampini } 26579552c7c7SStefano Zampini } 2658fd14bc51SStefano Zampini } 26599162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 26609162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 26619162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 26629162d606SStefano Zampini cum++; 266308122e43SStefano Zampini } 266408122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 266508122e43SStefano Zampini /* shift for next computation */ 266608122e43SStefano Zampini cumarray += subset_size*subset_size; 266708122e43SStefano Zampini } 2668fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2669fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2670fd14bc51SStefano Zampini } 267108122e43SStefano Zampini 267208122e43SStefano Zampini if (mss) { 267308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 267408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 2675f6f667cfSStefano Zampini /* destroy matrices (junk) */ 2676f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 2677f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 267808122e43SStefano Zampini } 2679f6f667cfSStefano Zampini if (allocated_S_St) { 2680f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 2681f6f667cfSStefano Zampini } 2682f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 268308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 268408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 268508122e43SStefano Zampini #endif 268608122e43SStefano Zampini if (pcbddc->dbg_flag) { 26871b968477SStefano Zampini PetscInt maxneigs_r; 2688b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 26899b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 269008122e43SStefano Zampini } 269108122e43SStefano Zampini PetscFunctionReturn(0); 269208122e43SStefano Zampini } 2693b1b3d7a2SStefano Zampini 2694674ae819SStefano Zampini #undef __FUNCT__ 2695c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 2696c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 2697c8587f34SStefano Zampini { 26988629588bSStefano Zampini PetscScalar *coarse_submat_vals; 2699c8587f34SStefano Zampini PetscErrorCode ierr; 2700c8587f34SStefano Zampini 2701c8587f34SStefano Zampini PetscFunctionBegin; 2702f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 27035e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 2704c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 2705c8587f34SStefano Zampini 2706684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 27070fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 2708684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2709c8587f34SStefano Zampini 27108629588bSStefano Zampini /* 27118629588bSStefano Zampini Setup local correction and local part of coarse basis. 27128629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 27138629588bSStefano Zampini */ 271447f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 27158629588bSStefano Zampini 27168629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 27178629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 27188629588bSStefano Zampini 27198629588bSStefano Zampini /* free */ 27208629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 2721c8587f34SStefano Zampini PetscFunctionReturn(0); 2722c8587f34SStefano Zampini } 2723c8587f34SStefano Zampini 2724c8587f34SStefano Zampini #undef __FUNCT__ 2725674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 2726674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 2727674ae819SStefano Zampini { 2728674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2729674ae819SStefano Zampini PetscErrorCode ierr; 2730674ae819SStefano Zampini 2731674ae819SStefano Zampini PetscFunctionBegin; 2732674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 273330368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 2734674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 2735785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2736674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 2737f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2738f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2739785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 274063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 274163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 2742674ae819SStefano Zampini PetscFunctionReturn(0); 2743674ae819SStefano Zampini } 2744674ae819SStefano Zampini 2745674ae819SStefano Zampini #undef __FUNCT__ 2746674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 2747674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 2748674ae819SStefano Zampini { 2749674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 27504f1b2e48SStefano Zampini PetscInt i; 2751674ae819SStefano Zampini PetscErrorCode ierr; 2752674ae819SStefano Zampini 2753674ae819SStefano Zampini PetscFunctionBegin; 2754a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 2755b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2756674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 275716909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 27581dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 2759674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2760669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 2761fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 2762a13144ffSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2763674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 27644f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 27654f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 27664f1b2e48SStefano Zampini } 27674f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 2768b334f244SStefano Zampini if (pcbddc->sub_schurs) { 2769b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 2770b334f244SStefano Zampini } 2771c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 27728af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 2773674ae819SStefano Zampini PetscFunctionReturn(0); 2774674ae819SStefano Zampini } 2775674ae819SStefano Zampini 2776674ae819SStefano Zampini #undef __FUNCT__ 2777674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 2778674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 2779674ae819SStefano Zampini { 2780674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2781674ae819SStefano Zampini PetscErrorCode ierr; 2782674ae819SStefano Zampini 2783674ae819SStefano Zampini PetscFunctionBegin; 2784674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 278558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 2786ca92afb2SStefano Zampini PetscScalar *array; 278706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 278806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 278958da7f69SStefano Zampini } 2790674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2791674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 279215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 279315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2794674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 2795674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 2796674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 279706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 2798674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 2799674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 28008ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2801674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2802674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2803674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 2804f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 2805f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 2806f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 2807f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2808727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 28090e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2810f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 281170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 281281d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 28130369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 28141dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 28154f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 28168b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 2817ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 2818ca92afb2SStefano Zampini PetscInt i; 2819ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2820ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 2821ca92afb2SStefano Zampini } 2822ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 2823ca92afb2SStefano Zampini } 28244f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 2825674ae819SStefano Zampini PetscFunctionReturn(0); 2826674ae819SStefano Zampini } 2827674ae819SStefano Zampini 2828674ae819SStefano Zampini #undef __FUNCT__ 2829f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 2830f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 28316bfb1811SStefano Zampini { 28326bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 28336bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 28346bfb1811SStefano Zampini VecType impVecType; 28354f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 28366bfb1811SStefano Zampini PetscErrorCode ierr; 28376bfb1811SStefano Zampini 28386bfb1811SStefano Zampini PetscFunctionBegin; 28396c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 2840e7b262bdSStefano Zampini /* get sizes */ 28414f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 2842b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 28436bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 2844e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 2845e7b262bdSStefano Zampini /* R nodes */ 2846e7b262bdSStefano Zampini old_size = -1; 2847e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 2848e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 2849e7b262bdSStefano Zampini } 2850e7b262bdSStefano Zampini if (n_R != old_size) { 2851e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 2852e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 28536bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 28546bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 28556bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 28566bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 2857e7b262bdSStefano Zampini } 2858e7b262bdSStefano Zampini /* local primal dofs */ 2859e7b262bdSStefano Zampini old_size = -1; 2860e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 2861e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 2862e7b262bdSStefano Zampini } 2863e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 2864e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 286583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 2866e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 28676bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 2868e7b262bdSStefano Zampini } 2869e7b262bdSStefano Zampini /* local explicit constraints */ 2870e7b262bdSStefano Zampini old_size = -1; 2871e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 2872e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 2873e7b262bdSStefano Zampini } 2874e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 2875e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 287683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 287783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 287883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 287983b7ccabSStefano Zampini } 28806bfb1811SStefano Zampini PetscFunctionReturn(0); 28816bfb1811SStefano Zampini } 28826bfb1811SStefano Zampini 28836bfb1811SStefano Zampini #undef __FUNCT__ 288447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 288547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 288688ebb749SStefano Zampini { 288725084f0cSStefano Zampini PetscErrorCode ierr; 288825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 288988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 289088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2891d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 289225084f0cSStefano Zampini /* submatrices of local problem */ 289380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 289406656605SStefano Zampini /* submatrices of local coarse problem */ 289506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 289625084f0cSStefano Zampini /* working matrices */ 289706656605SStefano Zampini Mat C_CR; 289825084f0cSStefano Zampini /* additional working stuff */ 289906656605SStefano Zampini PC pc_R; 29004f1b2e48SStefano Zampini Mat F; 29015cbda25cSStefano Zampini Vec dummy_vec; 2902a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 290325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 290406656605SStefano Zampini PetscScalar *work; 290506656605SStefano Zampini PetscInt *idx_V_B; 2906ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 290706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 2908ffd830a3SStefano Zampini 290925084f0cSStefano Zampini /* some shortcuts to scalars */ 291006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 291188ebb749SStefano Zampini 291288ebb749SStefano Zampini PetscFunctionBegin; 29139a962809SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 2914ffd830a3SStefano Zampini 2915ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 2916b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 29174f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 2918b371cd4fSStefano Zampini n_B = pcis->n_B; 2919b371cd4fSStefano Zampini n_D = pcis->n - n_B; 292088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 292188ebb749SStefano Zampini 292288ebb749SStefano Zampini /* vertices in boundary numbering */ 2923785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 29240e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 29256c4ed002SBarry Smith if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D\n",n_vertices,i); 292688ebb749SStefano Zampini 292706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 2928019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 292906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 293006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 293106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 293206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 293306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 293406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 293506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 293606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 293706656605SStefano Zampini 293806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 293906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 294006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 294106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 294206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 2943ffd830a3SStefano Zampini lda_rhs = n_R; 2944a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 294506656605SStefano Zampini if (isLU || isILU || isCHOL) { 294606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 2947b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 2948df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2949d62866d3SStefano Zampini MatFactorType type; 2950d62866d3SStefano Zampini 2951df4d28bfSStefano Zampini F = reuse_solver->F; 29526816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 2953d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 2954ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 295522db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 295606656605SStefano Zampini } else { 295706656605SStefano Zampini F = NULL; 295806656605SStefano Zampini } 295906656605SStefano Zampini 2960ffd830a3SStefano Zampini /* allocate workspace */ 2961ffd830a3SStefano Zampini n = 0; 2962ffd830a3SStefano Zampini if (n_constraints) { 2963ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 2964ffd830a3SStefano Zampini } 2965ffd830a3SStefano Zampini if (n_vertices) { 2966ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 2967ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 2968ffd830a3SStefano Zampini } 2969ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 2970ffd830a3SStefano Zampini 29715cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 29725cbda25cSStefano Zampini dummy_vec = NULL; 29735cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 29745cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 29755cbda25cSStefano Zampini } 29765cbda25cSStefano Zampini 297788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 297888ebb749SStefano Zampini if (n_constraints) { 297972b8c272SStefano Zampini Mat M1,M2,M3,C_B; 298006656605SStefano Zampini IS is_aux; 298180677318SStefano Zampini PetscScalar *array,*array2; 298206656605SStefano Zampini 2983f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 298480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 298588ebb749SStefano Zampini 298625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 298725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 29888ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 298972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 299088ebb749SStefano Zampini 299180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 299280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 2993ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 299488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 299506656605SStefano Zampini const PetscScalar *row_cmat_values; 299606656605SStefano Zampini const PetscInt *row_cmat_indices; 299706656605SStefano Zampini PetscInt size_of_constraint,j; 299888ebb749SStefano Zampini 299906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 300006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 3001ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 300206656605SStefano Zampini } 300306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 300406656605SStefano Zampini } 3005ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 300606656605SStefano Zampini if (F) { 300706656605SStefano Zampini Mat B; 300806656605SStefano Zampini 3009ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 3010a3df083aSStefano Zampini if (need_benign_correction) { 3011df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3012a3df083aSStefano Zampini 301372b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 301472b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 3015a3df083aSStefano Zampini } 301680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 3017a3df083aSStefano Zampini if (need_benign_correction) { 3018a3df083aSStefano Zampini PetscScalar *marr; 3019df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3020a3df083aSStefano Zampini 3021a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 30225cbda25cSStefano Zampini if (lda_rhs != n_R) { 30235cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 30245cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 30255cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 30265cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 30275cbda25cSStefano Zampini } 30285cbda25cSStefano Zampini } else { 3029a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 3030a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 30315cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3032a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3033a3df083aSStefano Zampini } 30345cbda25cSStefano Zampini } 3035a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 3036a3df083aSStefano Zampini } 303706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 303806656605SStefano Zampini } else { 303980677318SStefano Zampini PetscScalar *marr; 304080677318SStefano Zampini 304180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 304206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 3043ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3044ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 304506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 304606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 304706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 304806656605SStefano Zampini } 304980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 305006656605SStefano Zampini } 305180677318SStefano Zampini if (!pcbddc->switch_static) { 305280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 305380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 305480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 305580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 3056ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 305780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 305880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 306180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 306280677318SStefano Zampini } 306380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 306480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 306572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 306680677318SStefano Zampini } else { 3067ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3068ffd830a3SStefano Zampini IS dummy; 3069ffd830a3SStefano Zampini 3070ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 307172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 3072ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 3073ffd830a3SStefano Zampini } else { 307480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 307580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 3076ffd830a3SStefano Zampini } 307725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 307880677318SStefano Zampini } 307980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 308080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 308180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 308206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 308306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 308480677318SStefano Zampini if (isCHOL) { 308580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 308680677318SStefano Zampini } else { 308725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 308880677318SStefano Zampini } 308980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 309006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 309125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 309225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 309325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 309480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 309572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 309672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 309706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 309806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 3099f4ddd8eeSStefano Zampini } 3100fc227af8SStefano Zampini 3101fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 310288ebb749SStefano Zampini if (n_vertices) { 310306656605SStefano Zampini IS is_aux; 31043a50541eSStefano Zampini 3105b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 31066816873aSStefano Zampini IS tis; 31076816873aSStefano Zampini 31086816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 31096816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 31106816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 31116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 31126816873aSStefano Zampini } else { 31133a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 31146816873aSStefano Zampini } 31159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 31169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 311704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 311825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 311988ebb749SStefano Zampini } 312088ebb749SStefano Zampini 312188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 3122f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 312306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 312406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 312506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 312606656605SStefano Zampini } 3127f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 312806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 312906656605SStefano Zampini PetscScalar *marray; 313006656605SStefano Zampini 313106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 313206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 3133f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3134f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3135f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3136f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3137f4ddd8eeSStefano Zampini } 3138f4ddd8eeSStefano Zampini } 313906656605SStefano Zampini 3140f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 314106656605SStefano Zampini PetscScalar *marray; 314288ebb749SStefano Zampini 314306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 31448eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 314506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 314688ebb749SStefano Zampini } 31473301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 314806656605SStefano Zampini n *= 2; 314988ebb749SStefano Zampini } 315006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 315106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 315206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 31538eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 315406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 315506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 315688ebb749SStefano Zampini } 31573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 315806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 31598eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 316006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 316106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 316288ebb749SStefano Zampini } 316388ebb749SStefano Zampini } else { 3164c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 3165c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 31661b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3167c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 3168c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 3169c0553b1fSStefano Zampini } 317088ebb749SStefano Zampini } 317106656605SStefano Zampini } 3172019a44ceSStefano Zampini 317306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 31744f1b2e48SStefano Zampini p0_lidx_I = NULL; 31754f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 3176d12edf2fSStefano Zampini const PetscInt *idxs; 3177d12edf2fSStefano Zampini 3178d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 31794f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 31804f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 31814f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 31824f1b2e48SStefano Zampini } 3183d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 3184d12edf2fSStefano Zampini } 3185d16cbb6bSStefano Zampini 318606656605SStefano Zampini /* vertices */ 318706656605SStefano Zampini if (n_vertices) { 318816f15bc4SStefano Zampini 3189af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 319004708bb6SStefano Zampini 319116f15bc4SStefano Zampini if (n_R) { 319214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 319306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 319416f15bc4SStefano Zampini PetscScalar *x,*y; 319504708bb6SStefano Zampini PetscBool isseqaij; 319606656605SStefano Zampini 319721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 319814393ed6SStefano Zampini if (need_benign_correction) { 319914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 320014393ed6SStefano Zampini IS is_p0; 320114393ed6SStefano Zampini PetscInt *idxs_p0,n; 320214393ed6SStefano Zampini 320314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 320414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 320514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 3206af25d912SStefano Zampini if (n != pcbddc->benign_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 320714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 320814393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 320914393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 321014393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 321114393ed6SStefano Zampini } 321214393ed6SStefano Zampini 3213ffd830a3SStefano Zampini if (lda_rhs == n_R) { 3214af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3215ffd830a3SStefano Zampini } else { 3216ca92afb2SStefano Zampini PetscScalar *av,*array; 3217ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 3218ca92afb2SStefano Zampini PetscInt n; 3219ca92afb2SStefano Zampini PetscBool flg_row; 3220ffd830a3SStefano Zampini 3221ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 3222ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 32239d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3224ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3225ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 3226ca92afb2SStefano Zampini for (i=0;i<n;i++) { 3227ca92afb2SStefano Zampini PetscInt j; 3228ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 3229ffd830a3SStefano Zampini } 3230ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3231ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3232ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 3233ffd830a3SStefano Zampini } 3234ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3235a3df083aSStefano Zampini if (need_benign_correction) { 3236df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3237a3df083aSStefano Zampini PetscScalar *marr; 3238a3df083aSStefano Zampini 3239a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 324014393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 324114393ed6SStefano Zampini 324214393ed6SStefano Zampini | 0 0 0 | (V) 324314393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 324414393ed6SStefano Zampini | 0 0 -1 | (p0) 324514393ed6SStefano Zampini 324614393ed6SStefano Zampini */ 3247df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 324814393ed6SStefano Zampini const PetscScalar *vals; 324914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 325014393ed6SStefano Zampini PetscInt n,j,nz; 325114393ed6SStefano Zampini 3252df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3253df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 325414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 325514393ed6SStefano Zampini for (j=0;j<n;j++) { 325614393ed6SStefano Zampini PetscScalar val = vals[j]; 325714393ed6SStefano Zampini PetscInt k,col = idxs[j]; 325814393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 325914393ed6SStefano Zampini } 326014393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3261df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 326214393ed6SStefano Zampini } 326372b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 326472b8c272SStefano Zampini } 326572b8c272SStefano Zampini if (F) { 326614393ed6SStefano Zampini /* need to correct the rhs */ 326772b8c272SStefano Zampini if (need_benign_correction) { 326872b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 326972b8c272SStefano Zampini PetscScalar *marr; 327072b8c272SStefano Zampini 327172b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 32725cbda25cSStefano Zampini if (lda_rhs != n_R) { 32735cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 32745cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 32755cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 32765cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 32775cbda25cSStefano Zampini } 32785cbda25cSStefano Zampini } else { 3279a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3280a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 32815cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3282a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3283a3df083aSStefano Zampini } 32845cbda25cSStefano Zampini } 3285a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 3286a3df083aSStefano Zampini } 328706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 328814393ed6SStefano Zampini /* need to correct the solution */ 3289a3df083aSStefano Zampini if (need_benign_correction) { 3290df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3291a3df083aSStefano Zampini PetscScalar *marr; 3292a3df083aSStefano Zampini 3293a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 32945cbda25cSStefano Zampini if (lda_rhs != n_R) { 32955cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 32965cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 32975cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 32985cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 32995cbda25cSStefano Zampini } 33005cbda25cSStefano Zampini } else { 3301a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3302a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 33035cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3304a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3305a3df083aSStefano Zampini } 33065cbda25cSStefano Zampini } 3307a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 3308a3df083aSStefano Zampini } 330906656605SStefano Zampini } else { 331006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 331106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 3312ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 3313ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 331406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 331506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 331606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 331706656605SStefano Zampini } 331806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 331906656605SStefano Zampini } 332080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3321ffd830a3SStefano Zampini /* S_VV and S_CV */ 332206656605SStefano Zampini if (n_constraints) { 332306656605SStefano Zampini Mat B; 332480677318SStefano Zampini 3325ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 332680677318SStefano Zampini for (i=0;i<n_vertices;i++) { 3327ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3328ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 332980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 333080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 333180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 333280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 333380677318SStefano Zampini } 3334ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 333580677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 333680677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3337ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 333880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 333906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 3340ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 3341ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 334206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 334306656605SStefano Zampini } 334404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 334504708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 3346511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 334704708bb6SStefano Zampini } 3348ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3349ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 3350ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3351ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 3352ffd830a3SStefano Zampini } 335306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 335414393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 335514393ed6SStefano Zampini if (need_benign_correction) { 3356df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 335714393ed6SStefano Zampini PetscScalar *marr,*sums; 335814393ed6SStefano Zampini 335914393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 3360f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 3361df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 336214393ed6SStefano Zampini const PetscScalar *vals; 336314393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 336414393ed6SStefano Zampini PetscInt n,j,nz; 336514393ed6SStefano Zampini 3366df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3367df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 336814393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 336914393ed6SStefano Zampini PetscInt k; 337014393ed6SStefano Zampini sums[j] = 0.; 337114393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 337214393ed6SStefano Zampini } 337314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 337414393ed6SStefano Zampini for (j=0;j<n;j++) { 337514393ed6SStefano Zampini PetscScalar val = vals[j]; 337614393ed6SStefano Zampini PetscInt k; 337714393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 337814393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 337914393ed6SStefano Zampini } 338014393ed6SStefano Zampini } 338114393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3382df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 338314393ed6SStefano Zampini } 338414393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 3385f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 338614393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 338714393ed6SStefano Zampini } 338880677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 338906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 339006656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 339106656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 339206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 339306656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 339406656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 339506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3396d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 3397019a44ceSStefano Zampini } else { 3398d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3399d16cbb6bSStefano Zampini } 340021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 3401d16cbb6bSStefano Zampini 340206656605SStefano Zampini /* coarse basis functions */ 340306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 340416f15bc4SStefano Zampini PetscScalar *y; 340516f15bc4SStefano Zampini 3406ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 340706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 340806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 340906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 341106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 341206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 341306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 341406656605SStefano Zampini 341506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 34164f1b2e48SStefano Zampini PetscInt j; 34174f1b2e48SStefano Zampini 341806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 341906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 342006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 34234f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 342406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 342506656605SStefano Zampini } 342606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 342706656605SStefano Zampini } 342804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 342904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 343006656605SStefano Zampini } 34315cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 343206656605SStefano Zampini 343306656605SStefano Zampini if (n_constraints) { 343406656605SStefano Zampini Mat B; 343506656605SStefano Zampini 3436ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 343706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 343880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 343906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 344006656605SStefano Zampini if (n_vertices) { 344180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 344280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 344380677318SStefano Zampini } else { 344480677318SStefano Zampini Mat S_VCt; 344580677318SStefano Zampini 3446ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3447ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 344872b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 3449ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 3450ffd830a3SStefano Zampini } 345180677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 345280677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 345380677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 345480677318SStefano Zampini } 345506656605SStefano Zampini } 345606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 345706656605SStefano Zampini /* coarse basis functions */ 345806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 345906656605SStefano Zampini PetscScalar *y; 346006656605SStefano Zampini 3461ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 346206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 346306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 346406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 346506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 346606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 346706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 346806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 34694f1b2e48SStefano Zampini PetscInt j; 34704f1b2e48SStefano Zampini 347106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 347206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 347306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 347406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 347506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 34764f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 347706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 347806656605SStefano Zampini } 347906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 348006656605SStefano Zampini } 348106656605SStefano Zampini } 348280677318SStefano Zampini if (n_constraints) { 348380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 348480677318SStefano Zampini } 34854f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 348672b8c272SStefano Zampini 348772b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 348872b8c272SStefano Zampini if (pcbddc->benign_n) { 348972b8c272SStefano Zampini Mat B0_B,B0_BPHI; 349072b8c272SStefano Zampini IS is_dummy; 349172b8c272SStefano Zampini PetscScalar *data; 349272b8c272SStefano Zampini PetscInt j; 349372b8c272SStefano Zampini 349472b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 349572b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 349672b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 349772b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 349886c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 349972b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 350072b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 350172b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 350272b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 350372b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 350472b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 350572b8c272SStefano Zampini } 350672b8c272SStefano Zampini } 350772b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 350872b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 350972b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 351072b8c272SStefano Zampini } 3511019a44ceSStefano Zampini 351206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 35133301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 3514ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 3515ffd830a3SStefano Zampini PetscScalar *marray; 351606656605SStefano Zampini 351706656605SStefano Zampini if (n_constraints) { 3518ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 351906656605SStefano Zampini 3520af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 352106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 3522ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 352316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 352406656605SStefano Zampini if (n_vertices) { 3525ffd830a3SStefano Zampini Mat S_VCT; 352606656605SStefano Zampini 352706656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 3528ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 352916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 353006656605SStefano Zampini } 3531ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 35325b782168SStefano Zampini } else { 35335b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 353406656605SStefano Zampini } 353516f15bc4SStefano Zampini if (n_vertices && n_R) { 3536ffd830a3SStefano Zampini PetscScalar *av,*marray; 3537ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 3538ffd830a3SStefano Zampini PetscInt n; 3539ffd830a3SStefano Zampini PetscBool flg_row; 354006656605SStefano Zampini 3541ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 3542af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 3543ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3544ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 3545ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 3546ffd830a3SStefano Zampini for (i=0;i<n;i++) { 3547ffd830a3SStefano Zampini PetscInt j; 3548ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 3549ffd830a3SStefano Zampini } 3550ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 3551ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3552ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 355306656605SStefano Zampini } 355406656605SStefano Zampini 3555ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 3556ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 3557ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 3558ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 3559ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 356006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 356106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 356206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 356306656605SStefano Zampini } 3564ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 35655b782168SStefano Zampini if (B_C) { 3566ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 3567ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 3568ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 3569ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 3570ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 3571ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3572ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 357306656605SStefano Zampini } 3574ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 35755b782168SStefano Zampini } 357606656605SStefano Zampini /* coarse basis functions */ 357706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 357806656605SStefano Zampini PetscScalar *y; 357906656605SStefano Zampini 3580ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 358106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 358206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 358306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 358406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 358506656605SStefano Zampini if (i<n_vertices) { 358606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 358706656605SStefano Zampini } 358806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 358906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 359006656605SStefano Zampini 359106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 359206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 359306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 359406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 359506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 359606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 359706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 359806656605SStefano Zampini } 359906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 360006656605SStefano Zampini } 3601ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 3602ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 360306656605SStefano Zampini } 3604d62866d3SStefano Zampini /* free memory */ 360588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 360606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 360706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 360806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 360906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 3610d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3611d62866d3SStefano Zampini if (n_vertices) { 3612d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 3613d62866d3SStefano Zampini } 3614d62866d3SStefano Zampini if (n_constraints) { 3615d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 3616d62866d3SStefano Zampini } 361788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 361888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 361988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 3620d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 362188ebb749SStefano Zampini Mat coarse_sub_mat; 362225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 362388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 362488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 362588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 36268bec7fa6SStefano Zampini Mat C_B,CPHI; 36278bec7fa6SStefano Zampini IS is_dummy; 36288bec7fa6SStefano Zampini Vec mones; 362988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 363088ebb749SStefano Zampini PetscReal real_value; 363188ebb749SStefano Zampini 3632a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 3633a3df083aSStefano Zampini Mat A; 3634a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 3635a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 3636a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 3637a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 3638a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 3639a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 3640a3df083aSStefano Zampini } else { 364188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 364288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 364388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 364488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 3645a3df083aSStefano Zampini } 364688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 364788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 3648ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 364988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 365088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 365188ebb749SStefano Zampini } 365288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 365388ebb749SStefano Zampini 365425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 36553301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 365625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3657ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 365888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 365988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 366088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 366188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 366288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 366388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 366488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 366588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 366688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 366788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 366888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 366988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 367088ebb749SStefano Zampini } else { 367188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 367288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 367388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 367488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 367588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 367688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 367788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 367888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 367988ebb749SStefano Zampini } 368088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 368188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 368288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 3683511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 36844f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3685fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 3686d12edf2fSStefano Zampini PetscScalar *data,*data2; 36874f1b2e48SStefano Zampini PetscInt j; 3688d12edf2fSStefano Zampini 36894f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 3690fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 3691d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 369286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 3693d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 3694d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 36954f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 36964f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 3697d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 36984f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 36994f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 37004f1b2e48SStefano Zampini } 3701d12edf2fSStefano Zampini } 3702d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 3703d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 3704d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 3705d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3706d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 3707d12edf2fSStefano Zampini } 3708d12edf2fSStefano Zampini #if 0 3709d12edf2fSStefano Zampini { 3710d12edf2fSStefano Zampini PetscViewer viewer; 3711d12edf2fSStefano Zampini char filename[256]; 3712ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 3713d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3714d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3715ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 3716ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 3717ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 3718d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 371972b8c272SStefano Zampini if (save_change) { 372072b8c272SStefano Zampini Mat phi_B; 372172b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 372272b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 372372b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 372472b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 372572b8c272SStefano Zampini } else { 3726ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 3727ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 372872b8c272SStefano Zampini } 3729ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 3730ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 3731ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 3732ffd830a3SStefano Zampini } 3733ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 3734ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 3735ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 3736ffd830a3SStefano Zampini } 373772b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 3738ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 3739ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 3740ffd830a3SStefano Zampini } 3741d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3742d12edf2fSStefano Zampini } 3743d12edf2fSStefano Zampini #endif 374481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 37458bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 37461575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 374706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 37488bec7fa6SStefano Zampini 37498bec7fa6SStefano Zampini /* check constraints */ 3750a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 3751a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 37524f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 37538bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 3754a00504b5SStefano Zampini } else { 3755a00504b5SStefano Zampini PetscScalar *data; 3756a00504b5SStefano Zampini Mat tmat; 3757a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 3758a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 3759a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 3760a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 3761a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3762a00504b5SStefano Zampini } 37638bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 37648bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 37658bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 37668bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 3767bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 3768ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 3769bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 3770bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 3771bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 3772bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 3773bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 377488ebb749SStefano Zampini } 37758bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 37768bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 37778bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 37788bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 377925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 378088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 378188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 378288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 378388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 378488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 378588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 378688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 378788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 378888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 378988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 3790ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 379188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 379288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 379388ebb749SStefano Zampini } 379488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 379588ebb749SStefano Zampini } 37968629588bSStefano Zampini /* get back data */ 37978629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 379888ebb749SStefano Zampini PetscFunctionReturn(0); 379988ebb749SStefano Zampini } 380088ebb749SStefano Zampini 380188ebb749SStefano Zampini #undef __FUNCT__ 3802d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 3803d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 3804aa0d41d4SStefano Zampini { 3805d65f70fdSStefano Zampini Mat *work_mat; 3806d65f70fdSStefano Zampini IS isrow_s,iscol_s; 3807d65f70fdSStefano Zampini PetscBool rsorted,csorted; 3808c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 3809aa0d41d4SStefano Zampini PetscErrorCode ierr; 3810aa0d41d4SStefano Zampini 3811aa0d41d4SStefano Zampini PetscFunctionBegin; 3812d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 3813d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 3814d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 3815d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3816aa0d41d4SStefano Zampini 3817d65f70fdSStefano Zampini if (!rsorted) { 3818906d46d4SStefano Zampini const PetscInt *idxs; 3819906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 3820aa0d41d4SStefano Zampini 3821d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 3822d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 3823d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 3824d65f70fdSStefano Zampini idxs_perm_r[i] = i; 3825aa0d41d4SStefano Zampini } 3826d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 3827d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 3828d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 3829d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 3830aa0d41d4SStefano Zampini } 3831d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 3832d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 3833d65f70fdSStefano Zampini } else { 3834d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 3835d65f70fdSStefano Zampini isrow_s = isrow; 3836aa0d41d4SStefano Zampini } 3837906d46d4SStefano Zampini 3838d65f70fdSStefano Zampini if (!csorted) { 3839d65f70fdSStefano Zampini if (isrow == iscol) { 3840d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 3841d65f70fdSStefano Zampini iscol_s = isrow_s; 3842d65f70fdSStefano Zampini } else { 3843d65f70fdSStefano Zampini const PetscInt *idxs; 3844d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 3845906d46d4SStefano Zampini 3846d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 3847d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 3848d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 3849d65f70fdSStefano Zampini idxs_perm_c[i] = i; 3850d65f70fdSStefano Zampini } 3851d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 3852d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 3853d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 3854d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 3855d65f70fdSStefano Zampini } 3856d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 3857d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 3858d65f70fdSStefano Zampini } 3859d65f70fdSStefano Zampini } else { 3860d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 3861d65f70fdSStefano Zampini iscol_s = iscol; 3862d65f70fdSStefano Zampini } 3863d65f70fdSStefano Zampini 3864d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 3865d65f70fdSStefano Zampini 3866d65f70fdSStefano Zampini if (!rsorted || !csorted) { 3867906d46d4SStefano Zampini Mat new_mat; 3868d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 3869906d46d4SStefano Zampini 3870d65f70fdSStefano Zampini if (!rsorted) { 3871d65f70fdSStefano Zampini PetscInt *idxs_r,i; 3872d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 3873d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 3874d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 3875906d46d4SStefano Zampini } 3876d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 3877d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 3878d65f70fdSStefano Zampini } else { 3879d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 3880906d46d4SStefano Zampini } 3881d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 3882d65f70fdSStefano Zampini 3883d65f70fdSStefano Zampini if (!csorted) { 3884d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 3885d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 3886d65f70fdSStefano Zampini is_perm_c = is_perm_r; 3887d65f70fdSStefano Zampini } else { 3888d65f70fdSStefano Zampini PetscInt *idxs_c,i; 3889f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 3890d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 3891d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 3892d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 3893d65f70fdSStefano Zampini } 3894d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 3895d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 3896d65f70fdSStefano Zampini } 3897d65f70fdSStefano Zampini } else { 3898d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 3899d65f70fdSStefano Zampini } 3900d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 3901d65f70fdSStefano Zampini 3902d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 3903d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 3904d65f70fdSStefano Zampini work_mat[0] = new_mat; 3905d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 3906d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 3907d65f70fdSStefano Zampini } 3908d65f70fdSStefano Zampini 3909d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 3910d65f70fdSStefano Zampini *B = work_mat[0]; 3911d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 3912d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 3913d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 3914d65f70fdSStefano Zampini PetscFunctionReturn(0); 3915d65f70fdSStefano Zampini } 3916d65f70fdSStefano Zampini 3917d65f70fdSStefano Zampini #undef __FUNCT__ 39185e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 39195e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 3920aa0d41d4SStefano Zampini { 3921aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 39225e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3923d65f70fdSStefano Zampini Mat new_mat; 39245e8657edSStefano Zampini IS is_local,is_global; 3925d65f70fdSStefano Zampini PetscInt local_size; 3926d65f70fdSStefano Zampini PetscBool isseqaij; 3927aa0d41d4SStefano Zampini PetscErrorCode ierr; 3928aa0d41d4SStefano Zampini 3929aa0d41d4SStefano Zampini PetscFunctionBegin; 3930aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 39315e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 39325e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 3933b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 3934aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 3935d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 3936aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 3937906d46d4SStefano Zampini 3938906d46d4SStefano Zampini /* check */ 3939906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3940906d46d4SStefano Zampini Vec x,x_change; 3941906d46d4SStefano Zampini PetscReal error; 3942906d46d4SStefano Zampini 39435e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 3944906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 39455e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 3946e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3947e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3948d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 3949e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3950e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3951906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3952906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3953906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3954906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 3955906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3956906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3957906d46d4SStefano Zampini } 3958906d46d4SStefano Zampini 395922d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 39609b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 396122d5777bSStefano Zampini if (isseqaij) { 3962a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3963a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 3964aa0d41d4SStefano Zampini } else { 3965a00504b5SStefano Zampini Mat work_mat; 39661cf9b237SStefano Zampini 3967a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3968aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 3969a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 39701d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 3971aa0d41d4SStefano Zampini } 39723301b35fSStefano Zampini if (matis->A->symmetric_set) { 39733301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 3974e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 39753301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 3976e496cd5dSStefano Zampini #endif 39773301b35fSStefano Zampini } 3978d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 3979aa0d41d4SStefano Zampini PetscFunctionReturn(0); 3980aa0d41d4SStefano Zampini } 3981aa0d41d4SStefano Zampini 3982aa0d41d4SStefano Zampini #undef __FUNCT__ 3983a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 39848ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 3985a64d13efSStefano Zampini { 3986a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3987a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3988d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 398953892102SStefano Zampini PetscInt *idx_R_local=NULL; 39903a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 39913a50541eSStefano Zampini PetscInt vbs,bs; 39926816873aSStefano Zampini PetscBT bitmask=NULL; 3993a64d13efSStefano Zampini PetscErrorCode ierr; 3994a64d13efSStefano Zampini 3995a64d13efSStefano Zampini PetscFunctionBegin; 3996b23d619eSStefano Zampini /* 3997b23d619eSStefano Zampini No need to setup local scatters if 3998b23d619eSStefano Zampini - primal space is unchanged 3999b23d619eSStefano Zampini AND 4000b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 4001b23d619eSStefano Zampini AND 4002b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 4003b23d619eSStefano Zampini */ 4004b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 4005f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 4006f4ddd8eeSStefano Zampini } 4007f4ddd8eeSStefano Zampini /* destroy old objects */ 4008f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 4009f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 4010f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 4011a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 4012b371cd4fSStefano Zampini n_B = pcis->n_B; 4013b371cd4fSStefano Zampini n_D = pcis->n - n_B; 4014b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 40153a50541eSStefano Zampini 4016a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 40176816873aSStefano Zampini 401853892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 4019b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4020854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 4021a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 4022a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 40230e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 4024a64d13efSStefano Zampini } 4025a64d13efSStefano Zampini 4026a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 40274641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 40286816873aSStefano Zampini idx_R_local[n_R++] = i; 4029a64d13efSStefano Zampini } 4030a64d13efSStefano Zampini } 4031df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 4032df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 40336816873aSStefano Zampini 4034df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4035df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 40366816873aSStefano Zampini } 40373a50541eSStefano Zampini 40383a50541eSStefano Zampini /* Block code */ 40393a50541eSStefano Zampini vbs = 1; 40403a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 40413a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 40423a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 40433a50541eSStefano Zampini PetscInt *vary; 4044b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4045785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 40463a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 4047d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 4048d3df7717SStefano Zampini /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */ 40490e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 4050d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 40513a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 40523a50541eSStefano Zampini is_blocked = PETSC_FALSE; 40533a50541eSStefano Zampini break; 40543a50541eSStefano Zampini } 40553a50541eSStefano Zampini } 4056d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 4057d3df7717SStefano Zampini } else { 4058d3df7717SStefano Zampini /* Verify directly the R set */ 4059d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 4060d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 4061d3df7717SStefano Zampini for (j=1; j<bs; j++) { 4062d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 4063d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 4064d3df7717SStefano Zampini break; 4065d3df7717SStefano Zampini } 4066d3df7717SStefano Zampini } 4067d3df7717SStefano Zampini } 4068d3df7717SStefano Zampini } 40693a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 40703a50541eSStefano Zampini vbs = bs; 40713a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 40723a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 40733a50541eSStefano Zampini } 40743a50541eSStefano Zampini } 40753a50541eSStefano Zampini } 40763a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 4077b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4078df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 407953892102SStefano Zampini 4080df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4081df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 408253892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 4083df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 408453892102SStefano Zampini } else { 40853a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 408653892102SStefano Zampini } 4087a64d13efSStefano Zampini 4088a64d13efSStefano Zampini /* print some info if requested */ 4089a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 4090a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4091a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 40921575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4093a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 4094a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 40954f1b2e48SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"r_size = %d, v_size = %d, constraints = %d, local_primal_size = %d\n",n_R,n_vertices,pcbddc->local_primal_size-n_vertices-pcbddc->benign_n,pcbddc->local_primal_size);CHKERRQ(ierr); 4096a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4097a64d13efSStefano Zampini } 4098a64d13efSStefano Zampini 4099a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 4100b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 41016816873aSStefano Zampini IS is_aux1,is_aux2; 41026816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 41036816873aSStefano Zampini 41043a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4105854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 4106854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 4107a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 41084641a718SStefano Zampini for (i=0; i<n_D; i++) { 41094641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 41104641a718SStefano Zampini } 4111a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4112a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 41134641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 41144641a718SStefano Zampini aux_array1[j++] = i; 4115a64d13efSStefano Zampini } 4116a64d13efSStefano Zampini } 4117a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4118a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4119a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 41204641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 41214641a718SStefano Zampini aux_array2[j++] = i; 4122a64d13efSStefano Zampini } 4123a64d13efSStefano Zampini } 4124a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4125a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 4126a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 4127a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4128a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 4129a64d13efSStefano Zampini 41308eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4131785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 4132a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 41334641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 41344641a718SStefano Zampini aux_array1[j++] = i; 4135a64d13efSStefano Zampini } 4136a64d13efSStefano Zampini } 4137a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4138a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 4139a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4140a64d13efSStefano Zampini } 41414641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 41423a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4143d62866d3SStefano Zampini } else { 4144df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 41456816873aSStefano Zampini IS tis; 41466816873aSStefano Zampini PetscInt schur_size; 41476816873aSStefano Zampini 4148df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 41496816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 4150df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 41516816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 41526816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 41536816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 41546816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 41556816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 4156d62866d3SStefano Zampini } 4157d62866d3SStefano Zampini } 4158a64d13efSStefano Zampini PetscFunctionReturn(0); 4159a64d13efSStefano Zampini } 4160a64d13efSStefano Zampini 4161304d26faSStefano Zampini 4162304d26faSStefano Zampini #undef __FUNCT__ 4163304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 4164684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 4165304d26faSStefano Zampini { 4166304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4167304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4168304d26faSStefano Zampini PC pc_temp; 4169304d26faSStefano Zampini Mat A_RR; 4170f4ddd8eeSStefano Zampini MatReuse reuse; 4171304d26faSStefano Zampini PetscScalar m_one = -1.0; 4172304d26faSStefano Zampini PetscReal value; 417304708bb6SStefano Zampini PetscInt n_D,n_R; 4174c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 4175304d26faSStefano Zampini PetscErrorCode ierr; 4176e604994aSStefano Zampini /* prefixes stuff */ 4177312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 4178e604994aSStefano Zampini size_t len; 4179304d26faSStefano Zampini 4180304d26faSStefano Zampini PetscFunctionBegin; 4181304d26faSStefano Zampini 4182e604994aSStefano Zampini /* compute prefixes */ 4183e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 4184e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 4185e604994aSStefano Zampini if (!pcbddc->current_level) { 4186e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4187e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4188e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4189e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4190e604994aSStefano Zampini } else { 4191e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4192312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4193e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4194e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 4195312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4196312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 419734d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 419834d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4199e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4200e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4201e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 4202e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 4203e604994aSStefano Zampini } 4204e604994aSStefano Zampini 4205304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 4206684f6988SStefano Zampini if (dirichlet) { 4207d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4208450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 42099a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 4210450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 4211a3df083aSStefano Zampini Mat A_IIn; 4212a3df083aSStefano Zampini 4213a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 4214a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 4215a3df083aSStefano Zampini pcis->A_II = A_IIn; 4216a3df083aSStefano Zampini } 4217450f8f5eSStefano Zampini } 42183301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 42193301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 4220964fefecSStefano Zampini } 4221ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 4222964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 4223304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 4224304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 4225304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 4226304d26faSStefano Zampini /* default */ 4227304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 4228e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 42299577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 4230304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 42319577ea80SStefano Zampini if (issbaij) { 42329577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 42339577ea80SStefano Zampini } else { 4234304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 42359577ea80SStefano Zampini } 4236304d26faSStefano Zampini /* Allow user's customization */ 4237304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 4238304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4239304d26faSStefano Zampini } 4240d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 4241b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4242df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4243d62866d3SStefano Zampini 4244df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 4245d5574798SStefano Zampini } 4246304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4247304d26faSStefano Zampini if (!n_D) { 4248304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 4249304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4250304d26faSStefano Zampini } 4251304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 4252304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 4253304d26faSStefano Zampini /* set ksp_D into pcis data */ 4254304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 4255304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 4256304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 4257684f6988SStefano Zampini } 4258304d26faSStefano Zampini 4259304d26faSStefano Zampini /* NEUMANN PROBLEM */ 4260684f6988SStefano Zampini A_RR = 0; 4261684f6988SStefano Zampini if (neumann) { 4262d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 426304708bb6SStefano Zampini PetscInt ibs,mbs; 426404708bb6SStefano Zampini PetscBool issbaij; 426504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4266f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 42678ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 4268f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 4269f4ddd8eeSStefano Zampini PetscInt nn_R; 427081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 4271f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4272f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 4273f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 4274f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 4275f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4276f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4277f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 4278727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 4279f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4280f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4281f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 4282f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 4283f4ddd8eeSStefano Zampini } 4284f4ddd8eeSStefano Zampini } 4285f4ddd8eeSStefano Zampini /* last check */ 4286d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 4287f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4288f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4289f4ddd8eeSStefano Zampini } 4290f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 4291f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4292f4ddd8eeSStefano Zampini } 4293a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 4294af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 4295af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 429604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 429704708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 429804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 429904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 430004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4301af732b37SStefano Zampini } else { 4302511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 43036816873aSStefano Zampini } 430404708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 430504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 430604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 430704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 430804708bb6SStefano Zampini } else { 4309511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 431004708bb6SStefano Zampini } 431104708bb6SStefano Zampini } 4312a00504b5SStefano Zampini /* extract A_RR */ 4313b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4314a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4315a00504b5SStefano Zampini 4316a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 431716e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4318a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 431916e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 432016e386b8SStefano Zampini } else { 4321a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 4322a00504b5SStefano Zampini } 4323a00504b5SStefano Zampini } else { 4324a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4325a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 4326a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4327a00504b5SStefano Zampini } 4328a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 4329f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 433016e386b8SStefano Zampini } 43313301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 43323301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 43336816873aSStefano Zampini } 4334f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 4335304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 4336304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 4337304d26faSStefano Zampini /* default */ 4338304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 4339e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 4340304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 43419577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 43429577ea80SStefano Zampini if (issbaij) { 43439577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 43449577ea80SStefano Zampini } else { 4345304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 43469577ea80SStefano Zampini } 4347304d26faSStefano Zampini /* Allow user's customization */ 4348304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 4349304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4350304d26faSStefano Zampini } 4351304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4352304d26faSStefano Zampini if (!n_R) { 4353304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 4354304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4355304d26faSStefano Zampini } 43565cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 4357df4d28bfSStefano Zampini /* Reuse solver if it is present */ 4358b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4359df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4360d62866d3SStefano Zampini 4361df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 4362d62866d3SStefano Zampini } 4363304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 4364304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 4365684f6988SStefano Zampini } 4366304d26faSStefano Zampini 4367684f6988SStefano Zampini if (pcbddc->dbg_flag) { 4368684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43691575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4370684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4371684f6988SStefano Zampini } 4372c7017625SStefano Zampini 4373c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 4374c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 4375c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 4376c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 4377c7017625SStefano Zampini } 4378c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 4379c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 4380c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 4381c7017625SStefano Zampini } 4382c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 4383c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 4384c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 4385c7017625SStefano Zampini } 4386c7017625SStefano Zampini 4387c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 4388c7017625SStefano Zampini if (pcbddc->dbg_flag) { 4389684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 43900fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 43910fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 43920fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 43930fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 43940fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 4395e604994aSStefano 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); 4396c7017625SStefano Zampini if (check_corr[0]) { 4397c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 4398c7017625SStefano Zampini } 4399304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4400304d26faSStefano Zampini } 4401684f6988SStefano Zampini if (neumann) { /* Neumann */ 44020fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 44030fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 44040fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 44050fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 44060fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 4407e604994aSStefano 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); 4408c7017625SStefano Zampini if (check_corr[1]) { 4409c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 4410c7017625SStefano Zampini } 4411304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4412304d26faSStefano Zampini } 4413684f6988SStefano Zampini } 44145cbda25cSStefano Zampini /* free Neumann problem's matrix */ 44155cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4416304d26faSStefano Zampini PetscFunctionReturn(0); 4417304d26faSStefano Zampini } 4418304d26faSStefano Zampini 4419304d26faSStefano Zampini #undef __FUNCT__ 4420ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 442180677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 4422674ae819SStefano Zampini { 4423674ae819SStefano Zampini PetscErrorCode ierr; 4424674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4425be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4426b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 4427674ae819SStefano Zampini 4428674ae819SStefano Zampini PetscFunctionBegin; 4429b334f244SStefano Zampini if (!reuse_solver) { 443080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 443120c7b377SStefano Zampini } 443280677318SStefano Zampini if (!pcbddc->switch_static) { 443380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 443480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 443580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 443620c7b377SStefano Zampini } 4437b334f244SStefano Zampini if (!reuse_solver) { 443880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 443980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 444020c7b377SStefano Zampini } else { 4441df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4442be83ff47SStefano Zampini 4443df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4444df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 444520c7b377SStefano Zampini } 4446be83ff47SStefano Zampini } else { 444780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 444880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 444980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 445080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 445180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 445280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 445380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 445480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 445580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4456674ae819SStefano Zampini } 4457674ae819SStefano Zampini } 4458b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 445980677318SStefano Zampini if (applytranspose) { 446080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 446180677318SStefano Zampini } else { 446280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 446380677318SStefano Zampini } 4464be83ff47SStefano Zampini } else { 4465df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4466be83ff47SStefano Zampini 4467be83ff47SStefano Zampini if (applytranspose) { 4468df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4469be83ff47SStefano Zampini } else { 4470df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4471be83ff47SStefano Zampini } 4472be83ff47SStefano Zampini } 447380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 447480677318SStefano Zampini if (!pcbddc->switch_static) { 4475b334f244SStefano Zampini if (!reuse_solver) { 447680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 447780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4478be83ff47SStefano Zampini } else { 4479df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4480be83ff47SStefano Zampini 4481df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4482df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4483be83ff47SStefano Zampini } 448480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 448580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 448680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 448780677318SStefano Zampini } 448880677318SStefano Zampini } else { 448980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 449080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 449180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 449280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 449380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 449480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 449580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 449680677318SStefano Zampini } 449780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 449880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 449980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 450080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4501674ae819SStefano Zampini } 4502674ae819SStefano Zampini PetscFunctionReturn(0); 4503674ae819SStefano Zampini } 4504674ae819SStefano Zampini 4505dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 4506674ae819SStefano Zampini #undef __FUNCT__ 4507674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 4508dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 4509674ae819SStefano Zampini { 4510674ae819SStefano Zampini PetscErrorCode ierr; 4511674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4512674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 4513674ae819SStefano Zampini const PetscScalar zero = 0.0; 4514674ae819SStefano Zampini 4515674ae819SStefano Zampini PetscFunctionBegin; 4516dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 45174fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 4518dc359a40SStefano Zampini if (applytranspose) { 4519674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 45208eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 4521dc359a40SStefano Zampini } else { 4522674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 4523674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 452415aaf578SStefano Zampini } 45254fee134fSStefano Zampini } else { 45264fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 45274fee134fSStefano Zampini } 4528efc2fbd9SStefano Zampini 4529efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 45304f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4531efc2fbd9SStefano Zampini PetscScalar *array; 45324f1b2e48SStefano Zampini PetscInt j; 4533efc2fbd9SStefano Zampini 4534efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 45354f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 4536efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 4537efc2fbd9SStefano Zampini } 4538efc2fbd9SStefano Zampini 453912edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 454012edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 454112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 454212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 454312edc857SStefano Zampini 45449f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 454512edc857SStefano Zampini if (pcbddc->coarse_ksp) { 454651694757SStefano Zampini Mat coarse_mat; 4547964fefecSStefano Zampini Vec rhs,sol; 454851694757SStefano Zampini MatNullSpace nullsp; 454927b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 4550964fefecSStefano Zampini 455127b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 455227b6a85dSStefano Zampini PC coarse_pc; 455327b6a85dSStefano Zampini 455427b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 455527b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 455627b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 455727b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 455827b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 455927b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 45603bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 456127b6a85dSStefano Zampini } 456227b6a85dSStefano Zampini } 4563964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 4564964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 456551694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 456651694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 456751694757SStefano Zampini if (nullsp) { 456851694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 456951694757SStefano Zampini } 457012edc857SStefano Zampini if (applytranspose) { 45719a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 4572964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 45732701bc32SStefano Zampini } else { 45741f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 45752701bc32SStefano Zampini PC coarse_pc; 45762701bc32SStefano Zampini 45772701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 45782701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 45793e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 45802701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 458112edc857SStefano Zampini } else { 4582964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 458312edc857SStefano Zampini } 45842701bc32SStefano Zampini } 45851d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 458627b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 458727b6a85dSStefano Zampini PC coarse_pc; 458827b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 458927b6a85dSStefano Zampini 459027b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 459127b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 459227b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 45933bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 459427b6a85dSStefano Zampini } 459551694757SStefano Zampini if (nullsp) { 459651694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 459751694757SStefano Zampini } 459812edc857SStefano Zampini } 4599674ae819SStefano Zampini 4600674ae819SStefano Zampini /* Local solution on R nodes */ 46014fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 460280677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 46039f00e9b4SStefano Zampini } 46049f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 46059f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 460612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4607674ae819SStefano Zampini 46084fee134fSStefano Zampini /* Sum contributions from the two levels */ 46094fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 4610dc359a40SStefano Zampini if (applytranspose) { 4611dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 4612dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 4613dc359a40SStefano Zampini } else { 4614674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 46158eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 4616dc359a40SStefano Zampini } 4617efc2fbd9SStefano Zampini /* store p0 */ 46184f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4619efc2fbd9SStefano Zampini PetscScalar *array; 46204f1b2e48SStefano Zampini PetscInt j; 4621efc2fbd9SStefano Zampini 4622efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 46234f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 4624efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 4625efc2fbd9SStefano Zampini } 46264fee134fSStefano Zampini } else { /* expand the coarse solution */ 46274fee134fSStefano Zampini if (applytranspose) { 46284fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 46294fee134fSStefano Zampini } else { 46304fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 46314fee134fSStefano Zampini } 46324fee134fSStefano Zampini } 4633674ae819SStefano Zampini PetscFunctionReturn(0); 4634674ae819SStefano Zampini } 4635674ae819SStefano Zampini 4636674ae819SStefano Zampini #undef __FUNCT__ 4637674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 463812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 4639674ae819SStefano Zampini { 4640674ae819SStefano Zampini PetscErrorCode ierr; 4641674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 464258da7f69SStefano Zampini PetscScalar *array; 464312edc857SStefano Zampini Vec from,to; 4644674ae819SStefano Zampini 4645674ae819SStefano Zampini PetscFunctionBegin; 464612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 464712edc857SStefano Zampini from = pcbddc->coarse_vec; 464812edc857SStefano Zampini to = pcbddc->vec1_P; 464912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 465012edc857SStefano Zampini Vec tvec; 465158da7f69SStefano Zampini 465258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 465358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 465412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 465558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 465658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 465758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 465812edc857SStefano Zampini } 465912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 466012edc857SStefano Zampini from = pcbddc->vec1_P; 466112edc857SStefano Zampini to = pcbddc->coarse_vec; 466212edc857SStefano Zampini } 466312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 4664674ae819SStefano Zampini PetscFunctionReturn(0); 4665674ae819SStefano Zampini } 4666674ae819SStefano Zampini 4667674ae819SStefano Zampini #undef __FUNCT__ 4668674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 466912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 4670674ae819SStefano Zampini { 4671674ae819SStefano Zampini PetscErrorCode ierr; 4672674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 467358da7f69SStefano Zampini PetscScalar *array; 467412edc857SStefano Zampini Vec from,to; 4675674ae819SStefano Zampini 4676674ae819SStefano Zampini PetscFunctionBegin; 467712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 467812edc857SStefano Zampini from = pcbddc->coarse_vec; 467912edc857SStefano Zampini to = pcbddc->vec1_P; 468012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 468112edc857SStefano Zampini from = pcbddc->vec1_P; 468212edc857SStefano Zampini to = pcbddc->coarse_vec; 468312edc857SStefano Zampini } 468412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 468512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 468612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 468712edc857SStefano Zampini Vec tvec; 468858da7f69SStefano Zampini 468912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 469058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 469158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 469258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 469358da7f69SStefano Zampini } 469458da7f69SStefano Zampini } else { 469558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 469658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 469712edc857SStefano Zampini } 469812edc857SStefano Zampini } 4699674ae819SStefano Zampini PetscFunctionReturn(0); 4700674ae819SStefano Zampini } 4701674ae819SStefano Zampini 4702984c4197SStefano Zampini /* uncomment for testing purposes */ 4703984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 4704674ae819SStefano Zampini #undef __FUNCT__ 4705674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 4706674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 4707674ae819SStefano Zampini { 4708674ae819SStefano Zampini PetscErrorCode ierr; 4709674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4710674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4711674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4712984c4197SStefano Zampini /* one and zero */ 4713984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 4714984c4197SStefano Zampini /* space to store constraints and their local indices */ 47159162d606SStefano Zampini PetscScalar *constraints_data; 47169162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 47179162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 47189162d606SStefano Zampini PetscInt *constraints_n; 4719984c4197SStefano Zampini /* iterators */ 4720b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 4721984c4197SStefano Zampini /* BLAS integers */ 4722e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 4723e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 4724c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 4725727cdba6SStefano Zampini /* reuse */ 47260e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 47270e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 4728984c4197SStefano Zampini /* change of basis */ 4729b3d85658SStefano Zampini PetscBool qr_needed; 47309162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 4731984c4197SStefano Zampini /* auxiliary stuff */ 473264efe560SStefano Zampini PetscInt *nnz,*is_indices; 47338a0068c3SStefano Zampini PetscInt ncc; 4734984c4197SStefano Zampini /* some quantities */ 473545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 4736a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 4737984c4197SStefano Zampini 4738674ae819SStefano Zampini PetscFunctionBegin; 47398e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 47408e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 47418e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 474216909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 4743088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 4744088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 47450e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 47460e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 47470e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 47480e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 47490e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 4750088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4751cf5a6209SStefano Zampini 4752cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 47539162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 4754cf5a6209SStefano Zampini MatNullSpace nearnullsp; 4755cf5a6209SStefano Zampini const Vec *nearnullvecs; 4756cf5a6209SStefano Zampini Vec *localnearnullsp; 4757cf5a6209SStefano Zampini PetscScalar *array; 4758cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 4759cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 4760674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 4761b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 4762674ae819SStefano Zampini PetscScalar *work; 4763674ae819SStefano Zampini PetscReal *singular_vals; 4764674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4765674ae819SStefano Zampini PetscReal *rwork; 4766674ae819SStefano Zampini #endif 4767674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4768674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 4769674ae819SStefano Zampini #else 4770964fefecSStefano Zampini PetscBLASInt dummy_int=1; 4771964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 4772674ae819SStefano Zampini #endif 4773674ae819SStefano Zampini 4774674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 4775d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 4776e4d548c7SStefano Zampini /* print some info */ 4777*5c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 4778e4d548c7SStefano Zampini PetscInt nv; 4779e4d548c7SStefano Zampini 4780c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4781e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 4782e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4783e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4784e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 4785e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 4786e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 4787e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4788e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4789e4d548c7SStefano Zampini } 4790e4d548c7SStefano Zampini 4791d06fc5fdSStefano Zampini /* free unneeded index sets */ 4792d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 4793d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 4794674ae819SStefano Zampini } 4795d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 4796d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4797d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4798d06fc5fdSStefano Zampini } 4799d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4800d06fc5fdSStefano Zampini n_ISForEdges = 0; 4801d06fc5fdSStefano Zampini } 4802d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 4803d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4804d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4805d06fc5fdSStefano Zampini } 4806d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4807d06fc5fdSStefano Zampini n_ISForFaces = 0; 4808d06fc5fdSStefano Zampini } 480970022509SStefano Zampini 4810674ae819SStefano Zampini /* check if near null space is attached to global mat */ 4811674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 4812674ae819SStefano Zampini if (nearnullsp) { 4813674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 4814f4ddd8eeSStefano Zampini /* remove any stored info */ 4815f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 4816f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 4817f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 4818f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 4819f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 4820473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 4821f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 4822f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 4823f4ddd8eeSStefano Zampini } 4824984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 4825984c4197SStefano Zampini nnsp_size = 0; 4826674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 4827674ae819SStefano Zampini } 4828984c4197SStefano Zampini /* get max number of constraints on a single cc */ 4829984c4197SStefano Zampini max_constraints = nnsp_size; 4830984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 4831984c4197SStefano Zampini 4832674ae819SStefano Zampini /* 4833674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 48349162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 48359162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 48369162d606SStefano Zampini There can be multiple constraints per connected component 4837674ae819SStefano Zampini */ 4838674ae819SStefano Zampini n_vertices = 0; 4839674ae819SStefano Zampini if (ISForVertices) { 4840674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 4841674ae819SStefano Zampini } 48429162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 48439162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 48449162d606SStefano Zampini 48459162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 48469162d606SStefano Zampini total_counts *= max_constraints; 4847674ae819SStefano Zampini total_counts += n_vertices; 48484641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 48499162d606SStefano Zampini 4850674ae819SStefano Zampini total_counts = 0; 4851674ae819SStefano Zampini max_size_of_constraint = 0; 4852674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 48539162d606SStefano Zampini IS used_is; 4854674ae819SStefano Zampini if (i<n_ISForEdges) { 48559162d606SStefano Zampini used_is = ISForEdges[i]; 4856674ae819SStefano Zampini } else { 48579162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 4858674ae819SStefano Zampini } 48599162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 4860674ae819SStefano Zampini total_counts += j; 4861674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 4862674ae819SStefano Zampini } 48639162d606SStefano Zampini ierr = PetscMalloc3(total_counts*max_constraints+n_vertices,&constraints_data,total_counts+n_vertices,&constraints_idxs,total_counts+n_vertices,&constraints_idxs_B);CHKERRQ(ierr); 48649162d606SStefano Zampini 4865984c4197SStefano Zampini /* get local part of global near null space vectors */ 4866785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 4867984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4868984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 4869e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4870e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4871984c4197SStefano Zampini } 4872674ae819SStefano Zampini 4873242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 4874242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 4875a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 4876242a89d7SStefano Zampini 4877984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 4878a773dcb8SStefano Zampini if (!skip_lapack) { 4879674ae819SStefano Zampini PetscScalar temp_work; 4880911cabfeSStefano Zampini 4881674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4882984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 4883785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 4884785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 4885785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 4886674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4887785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 4888674ae819SStefano Zampini #endif 4889674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 4890c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 4891c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 4892674ae819SStefano Zampini lwork = -1; 4893674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4894674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 4895c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 4896674ae819SStefano Zampini #else 4897c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 4898674ae819SStefano Zampini #endif 4899674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4900984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 4901674ae819SStefano Zampini #else /* on missing GESVD */ 4902674ae819SStefano Zampini /* SVD */ 4903674ae819SStefano Zampini PetscInt max_n,min_n; 4904674ae819SStefano Zampini max_n = max_size_of_constraint; 4905984c4197SStefano Zampini min_n = max_constraints; 4906984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 4907674ae819SStefano Zampini min_n = max_size_of_constraint; 4908984c4197SStefano Zampini max_n = max_constraints; 4909674ae819SStefano Zampini } 4910785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 4911674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4912785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 4913674ae819SStefano Zampini #endif 4914674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 4915674ae819SStefano Zampini lwork = -1; 4916e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 4917e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 4918b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 4919674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4920674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 49219162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr)); 4922674ae819SStefano Zampini #else 49239162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr)); 4924674ae819SStefano Zampini #endif 4925674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4926984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 4927984c4197SStefano Zampini #endif /* on missing GESVD */ 4928674ae819SStefano Zampini /* Allocate optimal workspace */ 4929674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 4930854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 4931674ae819SStefano Zampini } 4932674ae819SStefano Zampini /* Now we can loop on constraining sets */ 4933674ae819SStefano Zampini total_counts = 0; 49349162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 49359162d606SStefano Zampini constraints_data_ptr[0] = 0; 4936674ae819SStefano Zampini /* vertices */ 49379162d606SStefano Zampini if (n_vertices) { 4938674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 49399162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 4940674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 49419162d606SStefano Zampini constraints_n[total_counts] = 1; 49429162d606SStefano Zampini constraints_data[total_counts] = 1.0; 49439162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 49449162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 4945674ae819SStefano Zampini total_counts++; 4946674ae819SStefano Zampini } 4947674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4948674ae819SStefano Zampini n_vertices = total_counts; 4949674ae819SStefano Zampini } 4950984c4197SStefano Zampini 4951674ae819SStefano Zampini /* edges and faces */ 49529162d606SStefano Zampini total_counts_cc = total_counts; 4953911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 49549162d606SStefano Zampini IS used_is; 49559162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 49569162d606SStefano Zampini 4957911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 49589162d606SStefano Zampini used_is = ISForEdges[ncc]; 4959984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 4960674ae819SStefano Zampini } else { 49619162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 4962984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 4963674ae819SStefano Zampini } 4964674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 49659162d606SStefano Zampini 49669162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 49679162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4968984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 4969984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 4970674ae819SStefano Zampini if (nnsp_has_cnst) { 49715b08dc53SStefano Zampini PetscScalar quad_value; 49729162d606SStefano Zampini 49739162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 49749162d606SStefano Zampini idxs_copied = PETSC_TRUE; 49759162d606SStefano Zampini 4976a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 4977674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 4978a773dcb8SStefano Zampini } else { 4979a773dcb8SStefano Zampini quad_value = 1.0; 4980a773dcb8SStefano Zampini } 4981674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 49829162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 4983674ae819SStefano Zampini } 49849162d606SStefano Zampini temp_constraints++; 4985674ae819SStefano Zampini total_counts++; 4986674ae819SStefano Zampini } 4987674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 4988984c4197SStefano Zampini PetscReal real_value; 49899162d606SStefano Zampini PetscScalar *ptr_to_data; 49909162d606SStefano Zampini 4991984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 49929162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 4993674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 49949162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 4995674ae819SStefano Zampini } 4996984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 4997984c4197SStefano Zampini /* check if array is null on the connected component */ 4998e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 49999162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 50005b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 5001674ae819SStefano Zampini temp_constraints++; 5002674ae819SStefano Zampini total_counts++; 50039162d606SStefano Zampini if (!idxs_copied) { 50049162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 50059162d606SStefano Zampini idxs_copied = PETSC_TRUE; 5006674ae819SStefano Zampini } 5007674ae819SStefano Zampini } 50089162d606SStefano Zampini } 50099162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 501045a1bb75SStefano Zampini valid_constraints = temp_constraints; 5011eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 5012a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 50139162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 50149162d606SStefano Zampini 50159162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5016a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 50179162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 5018a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 50199162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 5020a773dcb8SStefano Zampini } else { /* perform SVD */ 5021984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 50229162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5023674ae819SStefano Zampini 5024674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5025984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 5026984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 5027984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 5028984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 5029984c4197SStefano Zampini from that computed using LAPACKgesvd 5030984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 5031984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 5032984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 5033674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 5034e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5035984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5036674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 5037674ae819SStefano Zampini for (k=0;k<j+1;k++) { 50389162d606SStefano Zampini PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k] = BLASdot_(&Blas_N,ptr_to_data+k*size_of_constraint,&Blas_one,ptr_to_data+j*size_of_constraint,&Blas_one)); 5039674ae819SStefano Zampini } 5040674ae819SStefano Zampini } 5041e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 5042e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5043e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 5044674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5045c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 5046674ae819SStefano Zampini #else 5047c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 5048674ae819SStefano Zampini #endif 5049674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5050984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 5051984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 5052674ae819SStefano Zampini j = 0; 5053984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 5054674ae819SStefano Zampini total_counts = total_counts-j; 505545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 5056e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 5057c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5058c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5059c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 5060c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5061c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 5062c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5063674ae819SStefano Zampini if (j<temp_constraints) { 5064984c4197SStefano Zampini PetscInt ii; 5065984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 5066674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 50679162d606SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,ptr_to_data,&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC)); 5068674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5069984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 5070674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 50719162d606SStefano Zampini ptr_to_data[k*size_of_constraint+ii] = singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii]; 5072674ae819SStefano Zampini } 5073674ae819SStefano Zampini } 5074674ae819SStefano Zampini } 5075674ae819SStefano Zampini #else /* on missing GESVD */ 5076e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5077e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5078b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5079674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5080674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 50819162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr)); 5082674ae819SStefano Zampini #else 50839162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr)); 5084674ae819SStefano Zampini #endif 5085984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 5086674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5087984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 5088e310c8b4SStefano Zampini k = temp_constraints; 5089e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 5090674ae819SStefano Zampini j = 0; 5091e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 509245a1bb75SStefano Zampini valid_constraints = k-j; 5093911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 5094984c4197SStefano Zampini #endif /* on missing GESVD */ 5095674ae819SStefano Zampini } 5096a773dcb8SStefano Zampini } 50979162d606SStefano Zampini /* update pointers information */ 50989162d606SStefano Zampini if (valid_constraints) { 50999162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 51009162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 51019162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 51029162d606SStefano Zampini /* set change_of_basis flag */ 510345a1bb75SStefano Zampini if (boolforchange) { 5104b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 51059162d606SStefano Zampini } 5106b3d85658SStefano Zampini total_counts_cc++; 510745a1bb75SStefano Zampini } 510845a1bb75SStefano Zampini } 5109984c4197SStefano Zampini /* free workspace */ 51108f1c130eSStefano Zampini if (!skip_lapack) { 5111984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 5112984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5113984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 5114984c4197SStefano Zampini #endif 5115984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 5116984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5117984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 5118984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 5119984c4197SStefano Zampini #endif 5120984c4197SStefano Zampini } 5121984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5122984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 5123984c4197SStefano Zampini } 5124984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 5125cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 5126cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5127cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5128cf5a6209SStefano Zampini } 5129cf5a6209SStefano Zampini if (n_ISForFaces) { 5130cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5131cf5a6209SStefano Zampini } 5132cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5133cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5134cf5a6209SStefano Zampini } 5135cf5a6209SStefano Zampini if (n_ISForEdges) { 5136cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5137cf5a6209SStefano Zampini } 5138cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 513908122e43SStefano Zampini } else { 514008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5141984c4197SStefano Zampini 514208122e43SStefano Zampini total_counts = 0; 514308122e43SStefano Zampini n_vertices = 0; 5144d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 5145d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 514608122e43SStefano Zampini } 514708122e43SStefano Zampini max_constraints = 0; 51489162d606SStefano Zampini total_counts_cc = 0; 514908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 515008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 51519162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 515208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 515308122e43SStefano Zampini } 51549162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 51559162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 51569162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 51579162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 515874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 51599162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 51609162d606SStefano Zampini total_counts_cc = 0; 51619162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 51629162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 51639162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 516408122e43SStefano Zampini } 516508122e43SStefano Zampini } 51669162d606SStefano Zampini #if 0 51679162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 51689162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 51699162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 51709162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 51719162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 51729162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 51739162d606SStefano Zampini } 51749162d606SStefano Zampini printf("\n"); 51759162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 51769162d606SStefano Zampini } 51771b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 51788bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 51791b968477SStefano Zampini } 51801b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 51818bec7fa6SStefano 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]); 51821b968477SStefano Zampini } 518308122e43SStefano Zampini #endif 518408122e43SStefano Zampini 51858bec7fa6SStefano Zampini max_size_of_constraint = 0; 51869162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) max_size_of_constraint = PetscMax(max_size_of_constraint,constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]); 51879162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 518808122e43SStefano Zampini /* Change of basis */ 5189b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 519008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 519108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 519208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 5193b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 519408122e43SStefano Zampini } 519508122e43SStefano Zampini } 519608122e43SStefano Zampini } 519708122e43SStefano Zampini } 5198984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 51994f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 520008122e43SStefano Zampini 52019162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 52029162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 52036c4ed002SBarry Smith if (i != constraints_idxs_ptr[total_counts_cc]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %D != %D\n",constraints_idxs_ptr[total_counts_cc],i); 5204674ae819SStefano Zampini 5205674ae819SStefano Zampini /* Create constraint matrix */ 5206674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 520716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 5208984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 5209984c4197SStefano Zampini 5210984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 5211a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 5212a717540cSStefano Zampini qr_needed = PETSC_FALSE; 521374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 5214984c4197SStefano Zampini total_primal_vertices=0; 5215b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 52169162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 52179162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 521872b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 52199162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 5220b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 522164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 52229162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 52239162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 5224a717540cSStefano Zampini } 5225b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 522691af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 5227a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 5228a717540cSStefano Zampini qr_needed = PETSC_TRUE; 5229a717540cSStefano Zampini } 5230fa434743SStefano Zampini } else { 5231b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 5232fa434743SStefano Zampini } 5233a717540cSStefano Zampini } 5234b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 5235b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 5236674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 523770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5238b3d85658SStefano Zampini 52394f1b2e48SStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 52400e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 52410e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 5242984c4197SStefano Zampini 5243984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 524474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 5245785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 5246984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 524774d5cdf7SStefano Zampini 5248984c4197SStefano Zampini j = total_primal_vertices; 524974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 5250b3d85658SStefano Zampini cum = total_primal_vertices; 52519162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 52524641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 5253b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 5254b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 5255b3d85658SStefano Zampini cum++; 52569162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 525774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 525874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 525974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 526074d5cdf7SStefano Zampini } 52619162d606SStefano Zampini j += constraints_n[i]; 5262674ae819SStefano Zampini } 5263674ae819SStefano Zampini } 5264674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 5265674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 5266088faed8SStefano Zampini 5267674ae819SStefano Zampini /* set values in constraint matrix */ 5268984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 52690e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5270674ae819SStefano Zampini } 5271984c4197SStefano Zampini total_counts = total_primal_vertices; 52729162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 52734641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 52749162d606SStefano Zampini PetscInt *cols; 52759162d606SStefano Zampini 52769162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 52779162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 52789162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 52799162d606SStefano Zampini PetscInt row = total_counts+k; 52809162d606SStefano Zampini PetscScalar *vals; 52819162d606SStefano Zampini 52829162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 52839162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 52849162d606SStefano Zampini } 52859162d606SStefano Zampini total_counts += constraints_n[i]; 5286674ae819SStefano Zampini } 5287674ae819SStefano Zampini } 5288674ae819SStefano Zampini /* assembling */ 5289674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5290674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5291088faed8SStefano Zampini 5292984c4197SStefano Zampini /* 52936a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5294984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 5295f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5296984c4197SStefano Zampini */ 5297674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 5298674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 5299026de310SStefano Zampini /* dual and primal dofs on a single cc */ 5300984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 5301984c4197SStefano Zampini /* working stuff for GEQRF */ 530281d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 5303984c4197SStefano Zampini PetscBLASInt lqr_work; 5304984c4197SStefano Zampini /* working stuff for UNGQR */ 5305984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 5306984c4197SStefano Zampini PetscBLASInt lgqr_work; 5307984c4197SStefano Zampini /* working stuff for TRTRS */ 5308984c4197SStefano Zampini PetscScalar *trs_rhs; 53093f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 5310984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 5311984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 5312984c4197SStefano Zampini PetscScalar *start_vals; 5313984c4197SStefano Zampini /* working stuff for values insertion */ 53144641a718SStefano Zampini PetscBT is_primal; 531564efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 5316906d46d4SStefano Zampini /* matrix sizes */ 5317906d46d4SStefano Zampini PetscInt global_size,local_size; 5318906d46d4SStefano Zampini /* temporary change of basis */ 5319906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 5320cf5a6209SStefano Zampini /* extra space for debugging */ 5321cf5a6209SStefano Zampini PetscScalar *dbg_work; 5322984c4197SStefano Zampini 5323906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 5324906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 532516f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 5326bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 5327906d46d4SStefano Zampini /* nonzeros for local mat */ 5328bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 53291dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5330bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 53311dd7afcfSStefano Zampini } else { 53321dd7afcfSStefano Zampini const PetscInt *ii; 53331dd7afcfSStefano Zampini PetscInt n; 53341dd7afcfSStefano Zampini PetscBool flg_row; 53351dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 53361dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 53371dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 53381dd7afcfSStefano Zampini } 53399162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 5340a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 53419162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 5342a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 53439162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 5344a717540cSStefano Zampini } else { 53459162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 53469162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 5347a717540cSStefano Zampini } 5348a717540cSStefano Zampini } 5349a717540cSStefano Zampini } 5350906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 5351bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 53521dd7afcfSStefano Zampini /* Set interior change in the matrix */ 53531dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5354bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 5355906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 5356a717540cSStefano Zampini } 53571dd7afcfSStefano Zampini } else { 53581dd7afcfSStefano Zampini const PetscInt *ii,*jj; 53591dd7afcfSStefano Zampini PetscScalar *aa; 53601dd7afcfSStefano Zampini PetscInt n; 53611dd7afcfSStefano Zampini PetscBool flg_row; 53621dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 53631dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 53641dd7afcfSStefano Zampini for (i=0;i<n;i++) { 53651dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 53661dd7afcfSStefano Zampini } 53671dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 53681dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 53691dd7afcfSStefano Zampini } 5370a717540cSStefano Zampini 5371a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5372a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5373a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5374a717540cSStefano Zampini } 5375a717540cSStefano Zampini 5376a717540cSStefano Zampini 5377a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 5378a717540cSStefano Zampini /* 5379a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 5380a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 5381a717540cSStefano Zampini 5382a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 5383a717540cSStefano Zampini 5384a6b551f4SStefano 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) 5385a6b551f4SStefano Zampini 5386a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 5387a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 5388a717540cSStefano Zampini | ... | 5389a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 5390a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 5391a717540cSStefano Zampini 5392a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 5393a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 5394a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 5395a6b551f4SStefano Zampini 5396a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 5397a717540cSStefano Zampini */ 5398a717540cSStefano Zampini if (qr_needed) { 5399984c4197SStefano Zampini /* space to store Q */ 5400854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 54014e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 54024e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 5403984c4197SStefano Zampini /* first we issue queries for optimal work */ 54043f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 54053f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 54063f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5407984c4197SStefano Zampini lqr_work = -1; 54083f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 5409984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 5410984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 5411785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 5412984c4197SStefano Zampini lgqr_work = -1; 54133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 54143f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 54153f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 54163f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 54173f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 54183f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 5419984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 5420984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 5421785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 5422984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 5423785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 5424a717540cSStefano Zampini /* allocating workspace for check */ 5425a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5426cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 5427a717540cSStefano Zampini } 5428a717540cSStefano Zampini } 5429984c4197SStefano Zampini /* array to store whether a node is primal or not */ 54304641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 5431473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 54320e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 54336c4ed002SBarry Smith if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %D != %D\n",total_primal_vertices,i); 543439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 543539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 543639e2fb2aSStefano Zampini } 543739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 5438984c4197SStefano Zampini 5439a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 54409162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 54419162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 54424641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 5443984c4197SStefano Zampini /* get constraint info */ 54449162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 5445984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 5446984c4197SStefano Zampini 5447984c4197SStefano Zampini if (pcbddc->dbg_flag) { 54489162d606SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %d: %d need a change of basis (size %d)\n",total_counts,primal_dofs,size_of_constraint);CHKERRQ(ierr); 5449674ae819SStefano Zampini } 5450984c4197SStefano Zampini 5451fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 5452a717540cSStefano Zampini 5453a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 5454a717540cSStefano Zampini if (pcbddc->dbg_flag) { 54559162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5456a717540cSStefano Zampini } 5457984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 54589162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5459984c4197SStefano Zampini 5460984c4197SStefano Zampini /* compute QR decomposition of constraints */ 54613f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 54623f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 54633f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5464674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 54653f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 5466984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 5467674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5468984c4197SStefano Zampini 5469984c4197SStefano Zampini /* explictly compute R^-T */ 5470984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 5471984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 54723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 54733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 54743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 54753f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 5476984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 54773f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 5478984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 5479984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5480984c4197SStefano Zampini 5481a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 54823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 54833f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 54843f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 54853f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5486984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 54873f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 5488984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 5489984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5490984c4197SStefano Zampini 5491984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 5492984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 5493984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 54943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 54953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 54963f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 54973f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 54983f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 54993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5500984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 55019162d606SStefano Zampini PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,constraints_data+constraints_data_ptr[total_counts],&Blas_LDC)); 5502984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 55039162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5504984c4197SStefano Zampini 5505984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 55069162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 5507984c4197SStefano Zampini /* insert cols for primal dofs */ 5508984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 5509984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 55109162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 5511906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 5512984c4197SStefano Zampini } 5513984c4197SStefano Zampini /* insert cols for dual dofs */ 5514984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 55159162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 5516984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 55179162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 5518906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 5519984c4197SStefano Zampini j++; 5520674ae819SStefano Zampini } 5521674ae819SStefano Zampini } 5522984c4197SStefano Zampini 5523984c4197SStefano Zampini /* check change of basis */ 5524984c4197SStefano Zampini if (pcbddc->dbg_flag) { 5525984c4197SStefano Zampini PetscInt ii,jj; 5526984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 5527c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 5528c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5529c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 5530c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5531c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 5532c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 5533984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5534cf5a6209SStefano 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)); 5535984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5536984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 5537984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 5538cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 5539cf5a6209SStefano 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; 5540674ae819SStefano Zampini } 5541674ae819SStefano Zampini } 5542984c4197SStefano Zampini if (!valid_qr) { 554322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 5544984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 5545984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 5546cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 5547cf5a6209SStefano 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])); 5548674ae819SStefano Zampini } 5549cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 5550cf5a6209SStefano 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])); 5551984c4197SStefano Zampini } 5552984c4197SStefano Zampini } 5553984c4197SStefano Zampini } 5554674ae819SStefano Zampini } else { 555522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 5556674ae819SStefano Zampini } 5557674ae819SStefano Zampini } 5558a717540cSStefano Zampini } else { /* simple transformation block */ 5559a717540cSStefano Zampini PetscInt row,col; 5560a6b551f4SStefano Zampini PetscScalar val,norm; 5561a6b551f4SStefano Zampini 5562a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 55639162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,constraints_data+constraints_data_ptr[total_counts],&Blas_one,constraints_data+constraints_data_ptr[total_counts],&Blas_one)); 5564a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 55659162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 55669162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 5567bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 55689162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 5569906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 55709162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 5571a717540cSStefano Zampini } else { 5572a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 55739162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 5574a717540cSStefano Zampini if (row != col) { 55759162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 5576a717540cSStefano Zampini } else { 55779162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 5578a717540cSStefano Zampini } 5579906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 5580a717540cSStefano Zampini } 5581a717540cSStefano Zampini } 5582a717540cSStefano Zampini } 558398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 558422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 5585a717540cSStefano Zampini } 5586674ae819SStefano Zampini } 5587984c4197SStefano Zampini } else { 5588984c4197SStefano Zampini if (pcbddc->dbg_flag) { 55899162d606SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %d does not need a change of basis (size %d)\n",total_counts,size_of_constraint);CHKERRQ(ierr); 5590674ae819SStefano Zampini } 5591674ae819SStefano Zampini } 5592674ae819SStefano Zampini } 5593a717540cSStefano Zampini 5594a717540cSStefano Zampini /* free workspace */ 5595a717540cSStefano Zampini if (qr_needed) { 5596984c4197SStefano Zampini if (pcbddc->dbg_flag) { 5597cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 5598984c4197SStefano Zampini } 5599984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 5600984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 5601984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 5602984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 5603984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 5604674ae819SStefano Zampini } 5605a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 5606906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5607906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5608906d46d4SStefano Zampini 5609906d46d4SStefano Zampini /* assembling of global change of variable */ 561088c03ad3SStefano Zampini if (!pcbddc->fake_change) { 5611bbb9e6c6SStefano Zampini Mat tmat; 561216f15bc4SStefano Zampini PetscInt bs; 561316f15bc4SStefano Zampini 5614906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 5615906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 5616bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 5617bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 5618bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 5619bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 562016f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 562116f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 5622906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 5623bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 5624bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 5625bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 5626bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5627bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5628e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5629e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5630bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 5631bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 563288c03ad3SStefano Zampini 5633906d46d4SStefano Zampini /* check */ 5634906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 5635906d46d4SStefano Zampini PetscReal error; 5636906d46d4SStefano Zampini Vec x,x_change; 5637906d46d4SStefano Zampini 5638906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 5639906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 5640906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 5641906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 5642e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5643e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5644bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 5645e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5646e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5647906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 5648906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 5649906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 5650906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5651bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 5652906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 5653906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 5654906d46d4SStefano Zampini } 5655b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 5656b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 5657b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5658bf3a8328SStefano Zampini 56599a962809SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 5660b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 5661ac632422SStefano Zampini Mat S_new,tmat; 5662bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 5663bbb9e6c6SStefano Zampini 5664bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 56656816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 5666bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 5667bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 5668bf3a8328SStefano Zampini IS is_V; 5669b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 5670b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 5671b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 5672b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 5673b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 5674bf3a8328SStefano Zampini } 5675bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 5676ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 5677b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 5678ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 5679bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 5680bf3a8328SStefano Zampini const PetscScalar *array; 5681bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 5682bf3a8328SStefano Zampini PetscInt i,n_V; 5683bf3a8328SStefano Zampini 5684b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 5685b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 5686b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 5687b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 5688b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 5689b087196eSStefano Zampini for (i=0;i<n_V;i++) { 5690b087196eSStefano Zampini PetscScalar val; 5691b087196eSStefano Zampini PetscInt idx; 5692b087196eSStefano Zampini 5693b087196eSStefano Zampini idx = idxs_V[i]; 5694b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 5695b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 5696b087196eSStefano Zampini } 5697b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5698b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5699bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 5700bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 5701bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 5702bf3a8328SStefano Zampini } 5703ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 5704ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 5705ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 5706ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 5707b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 5708ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 5709bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 5710b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 5711bf3a8328SStefano Zampini } 5712ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 5713ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 5714ac632422SStefano Zampini } 5715b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 571688c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 5717b96c3477SStefano Zampini } 5718c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 5719b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 5720c9db6a07SStefano Zampini PetscInt i; 5721c9db6a07SStefano Zampini 5722c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5723c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 5724c9db6a07SStefano Zampini } 5725c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 5726c9db6a07SStefano Zampini } 5727b96c3477SStefano Zampini } 572816909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 572916909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 573016909a7fSStefano Zampini } else { 5731906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 573216909a7fSStefano Zampini } 57331dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 573427b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 573572b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 573672b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 573772b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 573872b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 573972b8c272SStefano Zampini } 57401dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 574127b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 5742b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 5743b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 5744906d46d4SStefano Zampini } else { 57451dd7afcfSStefano Zampini Mat benign_global = NULL; 574627b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 57471dd7afcfSStefano Zampini Mat tmat; 57481dd7afcfSStefano Zampini 57491dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 57501dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 57511dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 57521dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 57531dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 57541dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 57551dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 57561dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 57571dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 57581dd7afcfSStefano Zampini if (pcbddc->benign_change) { 57591dd7afcfSStefano Zampini Mat M; 57601dd7afcfSStefano Zampini 57611dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 57621dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 57631dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 57641dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 5765906d46d4SStefano Zampini } else { 57661dd7afcfSStefano Zampini Mat eye; 57671dd7afcfSStefano Zampini PetscScalar *array; 57681dd7afcfSStefano Zampini 57691dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 57701dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 57711dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 57721dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 5773906d46d4SStefano Zampini } 57741dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 57751dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57761dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 57771dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 57781dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 57791dd7afcfSStefano Zampini } 57801dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 57811dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 57821dd7afcfSStefano Zampini } 57831dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 57841dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 57851dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 578627b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 57871dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 57881dd7afcfSStefano Zampini } 57891dd7afcfSStefano Zampini } 579016909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 579116909a7fSStefano Zampini IS is_global; 579216909a7fSStefano Zampini const PetscInt *gidxs; 579316909a7fSStefano Zampini 579416909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 579516909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 579616909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 579716909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 579816909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 579916909a7fSStefano Zampini } 58001dd7afcfSStefano Zampini } 58011dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 58021dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 5803b9b85e73SStefano Zampini } 5804a717540cSStefano Zampini 580572b8c272SStefano Zampini if (!pcbddc->fake_change) { 58064f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 58074f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58084f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 58094f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 5810019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 5811019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 5812019a44ceSStefano Zampini pcbddc->local_primal_size++; 5813019a44ceSStefano Zampini } 5814019a44ceSStefano Zampini 5815019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 5816727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 5817727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 58189f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 5819c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 58200e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 58219f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 5822727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 5823727cdba6SStefano Zampini } 58240e6343abSStefano Zampini } 5825727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 5826b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 582772b8c272SStefano Zampini } 582872b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 5829727cdba6SStefano Zampini 5830a717540cSStefano Zampini /* flush dbg viewer */ 5831b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 5832b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5833b8ffe317SStefano Zampini } 5834a717540cSStefano Zampini 5835e310c8b4SStefano Zampini /* free workspace */ 5836a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 58374641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 583808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 58399162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 58409162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 584108122e43SStefano Zampini } else { 58429162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 58439162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 58449162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 584508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 584608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 58479162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 58489162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 584908122e43SStefano Zampini } 5850674ae819SStefano Zampini PetscFunctionReturn(0); 5851674ae819SStefano Zampini } 5852674ae819SStefano Zampini 5853674ae819SStefano Zampini #undef __FUNCT__ 5854674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 5855674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 5856674ae819SStefano Zampini { 585771582508SStefano Zampini ISLocalToGlobalMapping map; 5858674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5859674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 586014f95afaSStefano Zampini PetscInt ierr,i,N; 5861674ae819SStefano Zampini 5862674ae819SStefano Zampini PetscFunctionBegin; 58638af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 58648e61c736SStefano Zampini /* Reset previously computed graph */ 58658e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 5866674ae819SStefano Zampini /* Init local Graph struct */ 58677fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 586871582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 5869be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 5870674ae819SStefano Zampini 5871575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 58729a962809SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %d, expected %d\n",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs); 58739577ea80SStefano Zampini 5874674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 5875d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 58764d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 58774d379d7bSStefano Zampini PetscInt nvtxs; 5878e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 5879674ae819SStefano Zampini 58802fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 58812fffb893SStefano Zampini if (flg_row) { 58824d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 5883b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 58842fffb893SStefano Zampini } 58852fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5886674ae819SStefano Zampini } 58879b28b941SStefano Zampini if (pcbddc->dbg_flag) { 58889b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5889674ae819SStefano Zampini } 5890674ae819SStefano Zampini 5891674ae819SStefano Zampini /* Setup of Graph */ 58924b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 589314f95afaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,pcbddc->vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 5894674ae819SStefano Zampini 58954f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 58964f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 58974f1b2e48SStefano Zampini PetscInt *local_subs; 58984f1b2e48SStefano Zampini 58994f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 59004f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 59014f1b2e48SStefano Zampini const PetscInt *idxs; 59024f1b2e48SStefano Zampini PetscInt nl,j; 59034f1b2e48SStefano Zampini 59044f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 59054f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 590671582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 59074f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 59084f1b2e48SStefano Zampini } 59094f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 59104f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 59114f1b2e48SStefano Zampini } 59128af8fcf9SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 59138af8fcf9SStefano Zampini } 59144f1b2e48SStefano Zampini 5915cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 5916674ae819SStefano Zampini /* Graph's connected components analysis */ 5917674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 591871582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 59198af8fcf9SStefano Zampini } 5920674ae819SStefano Zampini PetscFunctionReturn(0); 5921674ae819SStefano Zampini } 5922674ae819SStefano Zampini 59239a7d3425SStefano Zampini #undef __FUNCT__ 59249a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 59259a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 59269a7d3425SStefano Zampini { 59279a7d3425SStefano Zampini PetscInt i,j; 59289a7d3425SStefano Zampini PetscScalar *alphas; 59299a7d3425SStefano Zampini PetscErrorCode ierr; 59309a7d3425SStefano Zampini 59319a7d3425SStefano Zampini PetscFunctionBegin; 5932785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 59339a7d3425SStefano Zampini for (i=0;i<n;i++) { 59349a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 5935669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 5936669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 5937669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 59389a7d3425SStefano Zampini } 59399a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 59409a7d3425SStefano Zampini PetscFunctionReturn(0); 59419a7d3425SStefano Zampini } 59429a7d3425SStefano Zampini 5943e7931f94SStefano Zampini #undef __FUNCT__ 594470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 594557de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 5946e7931f94SStefano Zampini { 594757de7509SStefano Zampini Mat A; 5948e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 5949e7931f94SStefano Zampini PetscMPIInt size,rank,color; 595052e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 595152e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 595227b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 595357de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 595427b6a85dSStefano Zampini PetscInt xadj_count, *count; 595527b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 595627b6a85dSStefano Zampini PetscSubcomm psubcomm; 595727b6a85dSStefano Zampini MPI_Comm subcomm; 595852e5ac9dSStefano Zampini PetscErrorCode ierr; 5959a57a6d2fSStefano Zampini 5960e7931f94SStefano Zampini PetscFunctionBegin; 596157de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 596257de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 596357de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 596457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 596557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 596657de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 596757de7509SStefano Zampini 596857de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 596957de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 597057de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 597157de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 597257de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 597357de7509SStefano Zampini im_active = !!(n); 597457de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 597557de7509SStefano Zampini void_procs = size - active_procs; 597657de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 597757de7509SStefano Zampini if (void_procs) { 597857de7509SStefano Zampini PetscInt ncand; 597957de7509SStefano Zampini 598057de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 598157de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 598257de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 598357de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 598457de7509SStefano Zampini if (!procs_candidates[i]) { 598557de7509SStefano Zampini procs_candidates[ncand++] = i; 598657de7509SStefano Zampini } 598757de7509SStefano Zampini } 598857de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 598957de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 599057de7509SStefano Zampini } 599157de7509SStefano Zampini 599214f0bfb9SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix 599314f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 599414f0bfb9SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1) { 599514f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 599614f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 599714f0bfb9SStefano Zampini else dest = rank; 599857de7509SStefano Zampini if (im_active) { 599957de7509SStefano Zampini issize = 1; 600057de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 600114f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 600257de7509SStefano Zampini } else { 600314f0bfb9SStefano Zampini isidx = dest; 600457de7509SStefano Zampini } 600557de7509SStefano Zampini } else { 600657de7509SStefano Zampini issize = 0; 600757de7509SStefano Zampini isidx = -1; 600857de7509SStefano Zampini } 600957de7509SStefano Zampini *n_subdomains = active_procs; 601057de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 6011daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 601257de7509SStefano Zampini PetscFunctionReturn(0); 601357de7509SStefano Zampini } 6014c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 6015c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 601627b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 6017e7931f94SStefano Zampini 6018e7931f94SStefano Zampini /* Get info on mapping */ 60193bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 60203bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6021e7931f94SStefano Zampini 6022e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 6023785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 6024e7931f94SStefano Zampini xadj[0] = 0; 6025e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 6026785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 6027785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 602827b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 602927b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 603027b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 603127b6a85dSStefano Zampini count[shared[i][j]] += 1; 6032e7931f94SStefano Zampini 603327b6a85dSStefano Zampini xadj_count = 0; 60342b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 603527b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 603627b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 6037d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 6038d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 6039d023bfaeSStefano Zampini xadj_count++; 604027b6a85dSStefano Zampini break; 604127b6a85dSStefano Zampini } 6042e7931f94SStefano Zampini } 6043e7931f94SStefano Zampini } 6044d023bfaeSStefano Zampini xadj[1] = xadj_count; 604527b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 60463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6047e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 6048e7931f94SStefano Zampini 60493837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 6050e7931f94SStefano Zampini 605127b6a85dSStefano Zampini /* Restrict work on active processes only */ 605227b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 605327b6a85dSStefano Zampini if (void_procs) { 605427b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 605527b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 605627b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 605727b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 605827b6a85dSStefano Zampini } else { 605927b6a85dSStefano Zampini psubcomm = NULL; 606027b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 606127b6a85dSStefano Zampini } 606227b6a85dSStefano Zampini 606327b6a85dSStefano Zampini v_wgt = NULL; 606427b6a85dSStefano Zampini if (!color) { 6065e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 6066e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 6067e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6068c8587f34SStefano Zampini } else { 606952e5ac9dSStefano Zampini Mat subdomain_adj; 607052e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 607152e5ac9dSStefano Zampini MatPartitioning partitioner; 607227b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 607352e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 607457de7509SStefano Zampini PetscMPIInt size; 6075b0c7d250SStefano Zampini PetscBool aggregate; 6076b0c7d250SStefano Zampini 607727b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 607827b6a85dSStefano Zampini if (void_procs) { 607927b6a85dSStefano Zampini PetscInt prank = rank; 6080785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 608127b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 6082e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 6083e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 6084c8587f34SStefano Zampini } 6085e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 608627b6a85dSStefano Zampini } else { 608727b6a85dSStefano Zampini oldranks = NULL; 608827b6a85dSStefano Zampini } 6089b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 609027b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 6091b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 6092b0c7d250SStefano Zampini PetscMPIInt nrank; 6093b0c7d250SStefano Zampini PetscScalar *vals; 6094b0c7d250SStefano Zampini 609527b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 6096b0c7d250SStefano Zampini lrows = 0; 6097b0c7d250SStefano Zampini if (nrank<redprocs) { 6098b0c7d250SStefano Zampini lrows = size/redprocs; 6099b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 6100b0c7d250SStefano Zampini } 610127b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 6102b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 6103b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6104b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6105b0c7d250SStefano Zampini row = nrank; 6106b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 6107b0c7d250SStefano Zampini cols = adjncy; 6108b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 6109b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 6110b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 6111b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6112b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 611352e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 611452e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 611552e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6116b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 611727b6a85dSStefano Zampini if (use_vwgt) { 611827b6a85dSStefano Zampini Vec v; 611927b6a85dSStefano Zampini const PetscScalar *array; 612027b6a85dSStefano Zampini PetscInt nl; 612127b6a85dSStefano Zampini 612227b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 612327b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 612427b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 612527b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 612627b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 612727b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 612827b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 612922db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 613027b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 613127b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 613227b6a85dSStefano Zampini } 6133b0c7d250SStefano Zampini } else { 613427b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 613527b6a85dSStefano Zampini if (use_vwgt) { 613627b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 613727b6a85dSStefano Zampini v_wgt[0] = local_size; 613827b6a85dSStefano Zampini } 6139b0c7d250SStefano Zampini } 614022b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 6141e7931f94SStefano Zampini 6142e7931f94SStefano Zampini /* Partition */ 614327b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 6144e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 614527b6a85dSStefano Zampini if (v_wgt) { 6146e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 6147c8587f34SStefano Zampini } 614857de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 614957de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 6150e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 6151e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 615222b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 6153e7931f94SStefano Zampini 615452e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 61556583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 615652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 615752e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 615857de7509SStefano Zampini if (!aggregate) { 615957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 616027b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 616127b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 616227b6a85dSStefano Zampini #endif 616357de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 616427b6a85dSStefano Zampini } else if (oldranks) { 6165b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 616627b6a85dSStefano Zampini } else { 616727b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 616857de7509SStefano Zampini } 616928143c3dSStefano Zampini } else { 6170b0c7d250SStefano Zampini PetscInt idxs[1]; 6171b0c7d250SStefano Zampini PetscMPIInt tag; 6172b0c7d250SStefano Zampini MPI_Request *reqs; 6173b0c7d250SStefano Zampini 6174b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 6175b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 6176b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 617727b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 617828143c3dSStefano Zampini } 617927b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 6180b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6181b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 618257de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 618327b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 618427b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 618527b6a85dSStefano Zampini #endif 618657de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 618727b6a85dSStefano Zampini } else if (oldranks) { 6188b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 618927b6a85dSStefano Zampini } else { 619027b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 6191e7931f94SStefano Zampini } 619257de7509SStefano Zampini } 619352e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6194e7931f94SStefano Zampini /* clean up */ 6195e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 619652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 6197e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 6198e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 6199e7931f94SStefano Zampini } 620027b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 620157de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 6202e7931f94SStefano Zampini 6203e7931f94SStefano Zampini /* assemble parallel IS for sends */ 6204e7931f94SStefano Zampini i = 1; 620527b6a85dSStefano Zampini if (!color) i=0; 620657de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 6207e7931f94SStefano Zampini PetscFunctionReturn(0); 6208e7931f94SStefano Zampini } 6209e7931f94SStefano Zampini 6210e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 6211e7931f94SStefano Zampini 6212e7931f94SStefano Zampini #undef __FUNCT__ 6213e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 62141ae86dd6SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, PetscBool reuse, Mat *mat_n, PetscInt nis, IS isarray[], PetscInt nvecs, Vec nnsp_vec[]) 6215e7931f94SStefano Zampini { 621670cf5478SStefano Zampini Mat local_mat; 6217e7931f94SStefano Zampini IS is_sends_internal; 62189d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 62191ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 62209d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 6221e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 6222e7931f94SStefano Zampini PetscInt* l2gmap_indices; 6223e7931f94SStefano Zampini const PetscInt* is_indices; 6224e7931f94SStefano Zampini MatType new_local_type; 6225e7931f94SStefano Zampini /* buffers */ 6226e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 622728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 62289d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 6229e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 62301ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 6231e7931f94SStefano Zampini /* MPI */ 623228143c3dSStefano Zampini MPI_Comm comm,comm_n; 623328143c3dSStefano Zampini PetscSubcomm subcomm; 6234e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 623528143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 623628143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 62371ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 62381ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 62391ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 6240e7931f94SStefano Zampini PetscErrorCode ierr; 6241e7931f94SStefano Zampini 6242e7931f94SStefano Zampini PetscFunctionBegin; 624357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 6244e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 624528143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 624657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 624757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 624857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 624957de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 625057de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 62511ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 62521ae86dd6SStefano Zampini if (nvecs) { 62531ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 62541ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 62551ae86dd6SStefano Zampini } 625657de7509SStefano Zampini /* further checks */ 6257e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6258e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 6259e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 6260e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 6261e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 626257de7509SStefano Zampini if (reuse && *mat_n) { 626370cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 626457de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 626570cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 626628143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 626770cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 626870cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 626970cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 627070cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 627170cf5478SStefano Zampini } 6272e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 6273e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 627457de7509SStefano Zampini 6275e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 6276e7931f94SStefano Zampini if (!is_sends) { 627728143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 627857de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 6279c8587f34SStefano Zampini } else { 6280e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 6281e7931f94SStefano Zampini is_sends_internal = is_sends; 6282c8587f34SStefano Zampini } 6283e7931f94SStefano Zampini 6284e7931f94SStefano Zampini /* get comm */ 6285a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 6286e7931f94SStefano Zampini 6287e7931f94SStefano Zampini /* compute number of sends */ 6288e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 6289e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 6290e7931f94SStefano Zampini 6291e7931f94SStefano Zampini /* compute number of receives */ 6292e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 6293785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 6294e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 6295e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6296e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 6297e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 6298e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 6299e7931f94SStefano Zampini 630028143c3dSStefano Zampini /* restrict comm if requested */ 630128143c3dSStefano Zampini subcomm = 0; 630228143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 630328143c3dSStefano Zampini if (restrict_comm) { 6304779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 6305779c1cceSStefano Zampini 630628143c3dSStefano Zampini color = 0; 630753a05cb3SStefano Zampini if (restrict_full) { 630853a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 630953a05cb3SStefano Zampini } else { 631053a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 631153a05cb3SStefano Zampini } 6312b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 631328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 631428143c3dSStefano Zampini /* check if reuse has been requested */ 631557de7509SStefano Zampini if (reuse) { 631628143c3dSStefano Zampini if (*mat_n) { 631728143c3dSStefano Zampini PetscMPIInt subcommsize2; 631828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 631928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 632028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 632128143c3dSStefano Zampini } else { 632228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 632328143c3dSStefano Zampini } 632428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 6325779c1cceSStefano Zampini PetscMPIInt rank; 6326779c1cceSStefano Zampini 6327779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 632828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 632928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 633028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 6331306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 633228143c3dSStefano Zampini } 633328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 633428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 633528143c3dSStefano Zampini } else { 633628143c3dSStefano Zampini comm_n = comm; 633728143c3dSStefano Zampini } 633828143c3dSStefano Zampini 6339e7931f94SStefano Zampini /* prepare send/receive buffers */ 6340785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 6341e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 6342785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 6343e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 634428143c3dSStefano Zampini if (nis) { 6345854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 634628143c3dSStefano Zampini } 6347e7931f94SStefano Zampini 634828143c3dSStefano Zampini /* Get data from local matrices */ 63496c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 6350e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 6351e7931f94SStefano Zampini /* 6352e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 6353e7931f94SStefano Zampini send_buffer_idxs should contain: 6354e7931f94SStefano Zampini - MatType_PRIVATE type 6355e7931f94SStefano Zampini - PetscInt size_of_l2gmap 6356e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 6357e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 6358e7931f94SStefano Zampini */ 63596c4ed002SBarry Smith else { 6360e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 63613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 6362854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 6363e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 6364e7931f94SStefano Zampini send_buffer_idxs[1] = i; 63653bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6366e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 63673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6368e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 6369e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6370e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 6371e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 6372c8587f34SStefano Zampini } 6373c8587f34SStefano Zampini } 6374e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 637528143c3dSStefano Zampini /* additional is (if any) */ 637628143c3dSStefano Zampini if (nis) { 637728143c3dSStefano Zampini PetscMPIInt psum; 637828143c3dSStefano Zampini PetscInt j; 637928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 638028143c3dSStefano Zampini PetscInt plen; 638128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 638228143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 638328143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 638428143c3dSStefano Zampini } 6385854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 638628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 638728143c3dSStefano Zampini PetscInt plen; 638828143c3dSStefano Zampini const PetscInt *is_array_idxs; 638928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 639028143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 639128143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 639228143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 639328143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 639428143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 639528143c3dSStefano Zampini } 639628143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 639728143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 639828143c3dSStefano Zampini } 639928143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 640028143c3dSStefano Zampini } 640128143c3dSStefano Zampini 6402e7931f94SStefano Zampini buf_size_idxs = 0; 6403e7931f94SStefano Zampini buf_size_vals = 0; 640428143c3dSStefano Zampini buf_size_idxs_is = 0; 64051ae86dd6SStefano Zampini buf_size_vecs = 0; 6406e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6407e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 6408e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 640928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 64101ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 6411e7931f94SStefano Zampini } 6412785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 6413785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 641495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 64151ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 6416e7931f94SStefano Zampini 6417e7931f94SStefano Zampini /* get new tags for clean communications */ 6418e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 6419e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 642028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 64211ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 6422e7931f94SStefano Zampini 6423e7931f94SStefano Zampini /* allocate for requests */ 6424785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 6425785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 642695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 64271ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 6428785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 6429785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 643095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 64311ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 6432e7931f94SStefano Zampini 6433e7931f94SStefano Zampini /* communications */ 6434e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6435e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 643628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 64371ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 6438e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6439e7931f94SStefano Zampini source_dest = onodes[i]; 6440e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 6441e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 6442e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6443e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 644428143c3dSStefano Zampini if (nis) { 644557de7509SStefano Zampini source_dest = onodes_is[i]; 644628143c3dSStefano 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); 644728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 644828143c3dSStefano Zampini } 64491ae86dd6SStefano Zampini if (nvecs) { 64501ae86dd6SStefano Zampini source_dest = onodes[i]; 64511ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 64521ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 64531ae86dd6SStefano Zampini } 6454e7931f94SStefano Zampini } 6455e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6456e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 6457e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 6458e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 645928143c3dSStefano Zampini if (nis) { 646028143c3dSStefano 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); 646128143c3dSStefano Zampini } 64621ae86dd6SStefano Zampini if (nvecs) { 64631ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 64641ae86dd6SStefano Zampini ierr = MPI_Isend(send_buffer_vecs,ilengths_idxs[source_dest]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&send_req_vecs[i]);CHKERRQ(ierr); 64651ae86dd6SStefano Zampini } 6466e7931f94SStefano Zampini } 6467e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6468e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 6469e7931f94SStefano Zampini 6470e7931f94SStefano Zampini /* assemble new l2g map */ 6471e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6472e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 64739d30be91SStefano Zampini new_local_rows = 0; 6474e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 64759d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6476e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6477e7931f94SStefano Zampini } 64789d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 6479e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 64809d30be91SStefano Zampini new_local_rows = 0; 6481e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 64829d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 64839d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6484e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6485e7931f94SStefano Zampini } 64869d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 64879d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 6488e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 6489e7931f94SStefano Zampini 6490e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 6491e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 6492e7931f94SStefano 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) */ 6493e7931f94SStefano Zampini if (n_recvs) { 649428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 6495e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6496e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6497e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 6498e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 6499e7931f94SStefano Zampini break; 6500e7931f94SStefano Zampini } 6501e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6502e7931f94SStefano Zampini } 6503e7931f94SStefano Zampini switch (new_local_type_private) { 650428143c3dSStefano Zampini case MATDENSE_PRIVATE: 650528143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 6506e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 6507e7931f94SStefano Zampini bs = 1; 650828143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 650928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 651028143c3dSStefano Zampini bs = 1; 651128143c3dSStefano Zampini } 6512e7931f94SStefano Zampini break; 6513e7931f94SStefano Zampini case MATAIJ_PRIVATE: 6514e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 6515e7931f94SStefano Zampini bs = 1; 6516e7931f94SStefano Zampini break; 6517e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 6518e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 6519e7931f94SStefano Zampini break; 6520e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 6521e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 6522e7931f94SStefano Zampini break; 6523e7931f94SStefano Zampini default: 65249d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 6525e7931f94SStefano Zampini break; 6526e7931f94SStefano Zampini } 652728143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 652828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 652928143c3dSStefano Zampini bs = 1; 6530e7931f94SStefano Zampini } 6531e7931f94SStefano Zampini 653270cf5478SStefano Zampini /* create MATIS object if needed */ 653357de7509SStefano Zampini if (!reuse) { 6534e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 6535e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 653670cf5478SStefano Zampini } else { 653770cf5478SStefano Zampini /* it also destroys the local matrices */ 653857de7509SStefano Zampini if (*mat_n) { 653970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 654057de7509SStefano Zampini } else { /* this is a fake object */ 654157de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 654257de7509SStefano Zampini } 654370cf5478SStefano Zampini } 654470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 6545e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 65469d30be91SStefano Zampini 65479d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 65489d30be91SStefano Zampini 65499d30be91SStefano Zampini /* Global to local map of received indices */ 65509d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 65519d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 65529d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 65539d30be91SStefano Zampini 65549d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 65559d30be91SStefano Zampini buf_size_idxs = 0; 65569d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 65579d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 65589d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 65599d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 65609d30be91SStefano Zampini } 65619d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 65629d30be91SStefano Zampini 65639d30be91SStefano Zampini /* set preallocation */ 65649d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 65659d30be91SStefano Zampini if (!newisdense) { 65669d30be91SStefano Zampini PetscInt *new_local_nnz=0; 65679d30be91SStefano Zampini 65689d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 65699d30be91SStefano Zampini if (n_recvs) { 65709d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 65719d30be91SStefano Zampini } 65729d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 65739d30be91SStefano Zampini PetscInt j; 65749d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 65759d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 65769d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 65779d30be91SStefano Zampini } 65789d30be91SStefano Zampini } else { 65799d30be91SStefano Zampini /* TODO */ 65809d30be91SStefano Zampini } 65819d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 65829d30be91SStefano Zampini } 65839d30be91SStefano Zampini if (new_local_nnz) { 65849d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 65859d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 65869d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 65879d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 65889d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 65899d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 65909d30be91SStefano Zampini } else { 65919d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 65929d30be91SStefano Zampini } 65939d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 65949d30be91SStefano Zampini } else { 65959d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 65969d30be91SStefano Zampini } 6597e7931f94SStefano Zampini 6598e7931f94SStefano Zampini /* set values */ 6599e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 66009d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 6601e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6602e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 6603e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 66049d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 6605e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 6606e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 6607e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 660828143c3dSStefano Zampini } else { 660928143c3dSStefano Zampini /* TODO */ 6610e7931f94SStefano Zampini } 6611e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6612e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 6613e7931f94SStefano Zampini } 6614e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6615e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 661670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 661770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 66189d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 6619e7931f94SStefano Zampini 6620dfd14d43SStefano Zampini #if 0 662128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 6622e7931f94SStefano Zampini Vec lvec,rvec; 6623e7931f94SStefano Zampini PetscReal infty_error; 6624e7931f94SStefano Zampini 66252a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 6626e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 6627e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 6628e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 662970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 6630e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 6631e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 6632e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 6633e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 6634e7931f94SStefano Zampini } 663528143c3dSStefano Zampini #endif 6636e7931f94SStefano Zampini 663728143c3dSStefano Zampini /* assemble new additional is (if any) */ 663828143c3dSStefano Zampini if (nis) { 663928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 664028143c3dSStefano Zampini 664128143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6642854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 664328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 664428143c3dSStefano Zampini psum = 0; 664528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 664628143c3dSStefano Zampini for (j=0;j<nis;j++) { 664728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 664828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 664928143c3dSStefano Zampini psum += plen; 665028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 665128143c3dSStefano Zampini } 665228143c3dSStefano Zampini } 6653854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 6654854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 665528143c3dSStefano Zampini for (i=1;i<nis;i++) { 665628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 665728143c3dSStefano Zampini } 665828143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 665928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 666028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 666128143c3dSStefano Zampini for (j=0;j<nis;j++) { 666228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 666328143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 666428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 666528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 666628143c3dSStefano Zampini } 666728143c3dSStefano Zampini } 666828143c3dSStefano Zampini for (i=0;i<nis;i++) { 666928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 667028143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 667128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 667228143c3dSStefano Zampini } 667328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 667428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 667528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 667628143c3dSStefano Zampini } 6677e7931f94SStefano Zampini /* free workspace */ 667828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 6679e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6680e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 6681e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6682e7931f94SStefano Zampini if (isdense) { 6683e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6684e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 6685e7931f94SStefano Zampini } else { 6686e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 6687e7931f94SStefano Zampini } 668828143c3dSStefano Zampini if (nis) { 668928143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 669028143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 669128143c3dSStefano Zampini } 66921ae86dd6SStefano Zampini 66931ae86dd6SStefano Zampini if (nvecs) { 66941ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 66951ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 66961ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 66971ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 66981ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 66991ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 67001ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 67011ae86dd6SStefano Zampini /* set values */ 67021ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 67031ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 67041ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 67051ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 67061ae86dd6SStefano Zampini PetscInt j; 67071ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 67081ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 67091ae86dd6SStefano Zampini } 67101ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 67111ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 67121ae86dd6SStefano Zampini } 67131ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 67141ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 67151ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 67161ae86dd6SStefano Zampini } 67171ae86dd6SStefano Zampini 67181ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 67191ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 6720e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 6721e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 67221ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 672328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 6724e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 6725e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 67261ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 672728143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 6728e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 6729e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 6730e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 6731e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 6732e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 673328143c3dSStefano Zampini if (nis) { 673428143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 673528143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 673628143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 673728143c3dSStefano Zampini } 673828143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 673928143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 674028143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 674128143c3dSStefano Zampini for (i=0;i<nis;i++) { 674228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 674328143c3dSStefano Zampini } 67441ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 67451ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 67461ae86dd6SStefano Zampini } 674753a05cb3SStefano Zampini *mat_n = NULL; 674828143c3dSStefano Zampini } 6749e7931f94SStefano Zampini PetscFunctionReturn(0); 6750e7931f94SStefano Zampini } 6751a57a6d2fSStefano Zampini 675212edc857SStefano Zampini /* temporary hack into ksp private data structure */ 6753af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 675412edc857SStefano Zampini 6755c8587f34SStefano Zampini #undef __FUNCT__ 6756c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 6757c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 6758c8587f34SStefano Zampini { 6759c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 6760c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 676120a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 67621ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 67639881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 676420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 67656e683305SStefano Zampini IS coarse_is,*isarray; 67666e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 676730368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 6768f9eb5b7dSStefano Zampini PC pc_temp; 6769c8587f34SStefano Zampini PCType coarse_pc_type; 6770c8587f34SStefano Zampini KSPType coarse_ksp_type; 6771f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 67724f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 67736e683305SStefano Zampini Mat t_coarse_mat_is; 677457de7509SStefano Zampini PetscInt ncoarse; 677568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 677622bc73bbSStefano Zampini PetscScalar *array; 677757de7509SStefano Zampini MatReuse coarse_mat_reuse; 677857de7509SStefano Zampini PetscBool restr, full_restr, have_void; 67799881197aSStefano Zampini PetscErrorCode ierr; 6780fdc09c96SStefano Zampini 6781c8587f34SStefano Zampini PetscFunctionBegin; 6782c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 678368457ee5SStefano 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 */ 6784fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 67855a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 6786fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 6787f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 6788f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 6789f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 6790fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 679151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 679251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 6793dc4bcba2SStefano Zampini PC pc; 6794dc4bcba2SStefano Zampini PetscBool isbddc; 6795dc4bcba2SStefano Zampini 6796dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 6797dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 6798dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 6799dc4bcba2SStefano Zampini if (isbddc) { 680063c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 680163c961adSStefano Zampini } else { 6802727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 680363c961adSStefano Zampini } 6804fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 6805fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 6806fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 6807f4ddd8eeSStefano Zampini } 6808fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 6809fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 6810f4ddd8eeSStefano Zampini } 681170cf5478SStefano Zampini /* reset any subassembling information */ 681257de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 681370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 681457de7509SStefano Zampini } 68156e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 6816fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 6817f4ddd8eeSStefano Zampini } 681857de7509SStefano Zampini /* assemble coarse matrix */ 681957de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 682057de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 682157de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 682257de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 682318a45a71SStefano Zampini } else { 682457de7509SStefano Zampini coarse_mat = NULL; 682557de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 68266e683305SStefano Zampini } 6827e7931f94SStefano Zampini 6828abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 6829abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 6830abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 6831abbbba34SStefano Zampini 6832abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 683322bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 683422bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 683522bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 683622bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 6837e176bc59SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,NULL,&t_coarse_mat_is);CHKERRQ(ierr); 68386e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 68396e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 68406e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6841abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 6842abbbba34SStefano Zampini 684357de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 684457de7509SStefano Zampini im_active = !!(pcis->n); 684557de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 684657de7509SStefano Zampini 684714f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 684857de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 684957de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 685057de7509SStefano Zampini coarse_mat_is = NULL; 685157de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 685257de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 68531ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 685457de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 685557de7509SStefano Zampini if (multilevel_requested) { 685657de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 685757de7509SStefano Zampini restr = PETSC_FALSE; 685857de7509SStefano Zampini full_restr = PETSC_FALSE; 685957de7509SStefano Zampini } else { 686057de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 686157de7509SStefano Zampini restr = PETSC_TRUE; 686257de7509SStefano Zampini full_restr = PETSC_TRUE; 686357de7509SStefano Zampini } 68644b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 686557de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 686657de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 6867a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 686857de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 6869a198735bSStefano Zampini } else { 6870a198735bSStefano Zampini PetscMPIInt size,rank; 6871a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 6872a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 6873a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 6874a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 6875a198735bSStefano Zampini } 687657de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 687757de7509SStefano Zampini PetscInt psum; 687857de7509SStefano Zampini PetscMPIInt size; 687957de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 688057de7509SStefano Zampini else psum = 0; 688157de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 688257de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 688357de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 688457de7509SStefano Zampini } 688557de7509SStefano Zampini /* determine if we can go multilevel */ 688657de7509SStefano Zampini if (multilevel_requested) { 688757de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 688857de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 688957de7509SStefano Zampini } 689057de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 689157de7509SStefano Zampini 6892e4d548c7SStefano Zampini /* dump subassembling pattern */ 6893e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 6894e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 6895e4d548c7SStefano Zampini } 6896e4d548c7SStefano Zampini 68976e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 6898c703fcc7SStefano Zampini if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 68996e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 69006e683305SStefano Zampini const PetscInt *idxs; 69016e683305SStefano Zampini ISLocalToGlobalMapping tmap; 69026e683305SStefano Zampini 69036e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 69040be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 69056e683305SStefano Zampini /* allocate space for temporary storage */ 6906854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 6907854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 69086e683305SStefano Zampini /* allocate for IS array */ 69096e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 69106e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 691127b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 691230368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 6913854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 69146e683305SStefano Zampini /* dofs splitting */ 69156e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 69166e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 69176e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 69186e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 69196e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 69206e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 69216e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 692230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 69236e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 69246e683305SStefano Zampini } 69256e683305SStefano Zampini /* neumann boundaries */ 69266e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 69276e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 69286e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 69296e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 69306e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 69316e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 69326e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 693330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 69346e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 69356e683305SStefano Zampini } 69366e683305SStefano Zampini /* free memory */ 69376e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 69386e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 69396e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 69406e683305SStefano Zampini } else { 69416e683305SStefano Zampini nis = 0; 69426e683305SStefano Zampini nisdofs = 0; 69436e683305SStefano Zampini nisneu = 0; 694430368db7SStefano Zampini nisvert = 0; 69456e683305SStefano Zampini isarray = NULL; 69466e683305SStefano Zampini } 69476e683305SStefano Zampini /* destroy no longer needed map */ 69486e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 69496e683305SStefano Zampini 695057de7509SStefano Zampini /* subassemble */ 695157de7509SStefano Zampini if (multilevel_allowed) { 69521ae86dd6SStefano Zampini Vec vp[1]; 69531ae86dd6SStefano Zampini PetscInt nvecs = 0; 695457de7509SStefano Zampini PetscBool reuse,reuser; 69551ae86dd6SStefano Zampini 695657de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 695757de7509SStefano Zampini else reuse = PETSC_FALSE; 695857de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 69591ae86dd6SStefano Zampini vp[0] = NULL; 69601ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 69611ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 69621ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 69631ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 69641ae86dd6SStefano Zampini nvecs = 1; 69651ae86dd6SStefano Zampini 69661ae86dd6SStefano Zampini if (pcbddc->divudotp) { 6967a198735bSStefano Zampini Mat B,loc_divudotp; 69681ae86dd6SStefano Zampini Vec v,p; 69691ae86dd6SStefano Zampini IS dummy; 69701ae86dd6SStefano Zampini PetscInt np; 69711ae86dd6SStefano Zampini 6972a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 6973a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 69741ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 6975a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 69761ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 69771ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 69781ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 69791ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 69801ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 69811ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 69821ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 69831ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 69841ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 69851ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 69861ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 69871ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 698874e2c79eSStefano Zampini } 69891ae86dd6SStefano Zampini } 69901ae86dd6SStefano Zampini if (reuser) { 69911ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 699274e2c79eSStefano Zampini } else { 69931ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray,nvecs,vp);CHKERRQ(ierr); 69941ae86dd6SStefano Zampini } 69951ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 69961ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 69971ae86dd6SStefano Zampini PetscInt nl; 69981ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 69991ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 70001ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 70011ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 70021ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 70031ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 70041ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 70051ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 7006a198735bSStefano Zampini } else { 7007a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 70081ae86dd6SStefano Zampini } 70091ae86dd6SStefano Zampini } else { 70101ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray,0,NULL);CHKERRQ(ierr); 70116e683305SStefano Zampini } 701257de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 701357de7509SStefano Zampini PetscMPIInt size; 7014f913dca9SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size);CHKERRQ(ierr); 701557de7509SStefano Zampini if (!multilevel_allowed) { 701657de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 70176e683305SStefano Zampini } else { 701857de7509SStefano Zampini Mat A; 7019779c1cceSStefano Zampini 702057de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 702157de7509SStefano Zampini if (coarse_mat_is) { 702257de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 702357de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 702457de7509SStefano Zampini coarse_mat = coarse_mat_is; 702557de7509SStefano Zampini } 702657de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 702757de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 702857de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 7029779c1cceSStefano Zampini } 7030779c1cceSStefano Zampini } 703157de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 703257de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 70336e683305SStefano Zampini 70346e683305SStefano Zampini /* create local to global scatters for coarse problem */ 703568457ee5SStefano Zampini if (compute_vecs) { 70366e683305SStefano Zampini PetscInt lrows; 70376e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 703857de7509SStefano Zampini if (coarse_mat) { 703957de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 70406e683305SStefano Zampini } else { 70416e683305SStefano Zampini lrows = 0; 70426e683305SStefano Zampini } 70436e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 70446e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 70456e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 70466e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 70476e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 70486e683305SStefano Zampini } 70496e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 7050c8587f34SStefano Zampini 7051f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 7052f9eb5b7dSStefano Zampini if (multilevel_allowed) { 7053f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 7054f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 7055f9eb5b7dSStefano Zampini } else { 7056f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 7057f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 7058c8587f34SStefano Zampini } 7059c8587f34SStefano Zampini 70606e683305SStefano Zampini /* print some info if requested */ 70616e683305SStefano Zampini if (pcbddc->dbg_flag) { 70626e683305SStefano Zampini if (!multilevel_allowed) { 70636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 70646e683305SStefano Zampini if (multilevel_requested) { 70656e683305SStefano 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); 70666e683305SStefano Zampini } else if (pcbddc->max_levels) { 70676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 70686e683305SStefano Zampini } 70696e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 70706e683305SStefano Zampini } 70716e683305SStefano Zampini } 70726e683305SStefano Zampini 7073f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 707457de7509SStefano Zampini if (coarse_mat) { 70756a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 70766e683305SStefano Zampini if (pcbddc->dbg_flag) { 707757de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 70786e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 70796e683305SStefano Zampini } 7080f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 7081312be037SStefano Zampini char prefix[256],str_level[16]; 7082e604994aSStefano Zampini size_t len; 708357de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 7084422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 7085c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 7086f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 708757de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7088c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 70896e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 7090c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 7091c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7092e604994aSStefano Zampini /* prefix */ 7093e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 7094e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 7095e604994aSStefano Zampini if (!pcbddc->current_level) { 7096e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 7097e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 7098c8587f34SStefano Zampini } else { 7099e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 7100312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 7101312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 710234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 7103312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 7104e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 7105e604994aSStefano Zampini } 7106e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 71073e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 71083e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 71093e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 71103e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 7111f9eb5b7dSStefano Zampini /* allow user customization */ 7112f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 71133e3c6dadSStefano Zampini } 71143e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 711551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 71163e3c6dadSStefano Zampini if (nisdofs) { 71173e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 71183e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 71193e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 71203e3c6dadSStefano Zampini } 71213e3c6dadSStefano Zampini } 71223e3c6dadSStefano Zampini if (nisneu) { 71233e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 71243e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 7125312be037SStefano Zampini } 712630368db7SStefano Zampini if (nisvert) { 712730368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 712830368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 712930368db7SStefano Zampini } 7130f9eb5b7dSStefano Zampini 7131f9eb5b7dSStefano Zampini /* get some info after set from options */ 7132f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 7133f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 71344f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 71356e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 7136f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7137f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 7138f9eb5b7dSStefano Zampini } 713939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 71404f3a063dSStefano Zampini if (isredundant) { 71414f3a063dSStefano Zampini KSP inner_ksp; 71424f3a063dSStefano Zampini PC inner_pc; 71434f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 71444f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 71454f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 71464f3a063dSStefano Zampini } 7147f9eb5b7dSStefano Zampini 714857de7509SStefano Zampini /* parameters which miss an API */ 714957de7509SStefano Zampini if (isbddc) { 7150720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 7151720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 715257de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 715327b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 715427b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 7155a198735bSStefano Zampini Mat coarsedivudotp_is; 7156a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 7157a198735bSStefano Zampini IS row,col; 7158a198735bSStefano Zampini const PetscInt *gidxs; 7159a198735bSStefano Zampini PetscInt n,st,M,N; 7160a198735bSStefano Zampini 7161a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 7162a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 7163a198735bSStefano Zampini st = st-n; 7164a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 7165a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 7166a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 7167a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7168a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 7169a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7170a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 7171a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 7172a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 7173a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 7174a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 7175a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 7176a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 7177a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 7178a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 7179a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 7180a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 7181a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 7182a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 7183a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 71848ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 7185a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 7186720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 718759e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 7188720d30f9SStefano Zampini } 7189d4d8cf7bSStefano Zampini } 71909881197aSStefano Zampini 71913301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 71925a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 71933301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 71943301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 71953301b35fSStefano Zampini } 71963301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 71973301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 71983301b35fSStefano Zampini } 71993301b35fSStefano Zampini if (pc->pmat->spd_set) { 72003301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 72013301b35fSStefano Zampini } 720227b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 720327b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 720427b6a85dSStefano Zampini } 72056e683305SStefano Zampini /* set operators */ 72065f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 72076e683305SStefano Zampini if (pcbddc->dbg_flag) { 72086e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 72096e683305SStefano Zampini } 72106e683305SStefano Zampini } 72116e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 7212b1ecc7b1SStefano Zampini #if 0 7213b9b85e73SStefano Zampini { 7214b9b85e73SStefano Zampini PetscViewer viewer; 7215b9b85e73SStefano Zampini char filename[256]; 7216b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 7217b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 72186a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 7219b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 7220f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 7221b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 7222b9b85e73SStefano Zampini } 7223b9b85e73SStefano Zampini #endif 7224f9eb5b7dSStefano Zampini 722598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 722698a51de6SStefano Zampini Vec crhs,csol; 722704708bb6SStefano Zampini 7228f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 7229f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 7230f347579bSStefano Zampini if (!csol) { 72312a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 7232f9eb5b7dSStefano Zampini } 7233f347579bSStefano Zampini if (!crhs) { 72342a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 7235f347579bSStefano Zampini } 7236b0f5fe93SStefano Zampini } 72371ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 7238b0f5fe93SStefano Zampini 7239b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 7240b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 7241b0f5fe93SStefano Zampini 7242b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 72434f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 72444f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 72454f1b2e48SStefano Zampini } 7246b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 7247b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 7248b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7249b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7250b0f5fe93SStefano Zampini if (coarse_mat) { 7251b0f5fe93SStefano Zampini Vec nullv; 7252b0f5fe93SStefano Zampini PetscScalar *array,*array2; 7253b0f5fe93SStefano Zampini PetscInt nl; 7254b0f5fe93SStefano Zampini 7255b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 7256b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 7257b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7258b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 7259b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 7260b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 7261b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7262b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 7263b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 7264b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 7265b0f5fe93SStefano Zampini } 7266b0f5fe93SStefano Zampini } 7267b0f5fe93SStefano Zampini 7268b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 7269b0f5fe93SStefano Zampini PetscBool ispreonly; 7270b0f5fe93SStefano Zampini 7271b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7272b0f5fe93SStefano Zampini PetscBool isnull; 7273b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 7274bef83e63SStefano Zampini if (isnull) { 7275b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 7276b0f5fe93SStefano Zampini } 7277bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 7278b0f5fe93SStefano Zampini } 7279b0f5fe93SStefano Zampini /* setup coarse ksp */ 7280b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 7281cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 7282cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 72836e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 7284c8587f34SStefano Zampini KSP check_ksp; 72852b510759SStefano Zampini KSPType check_ksp_type; 7286c8587f34SStefano Zampini PC check_pc; 72876e683305SStefano Zampini Vec check_vec,coarse_vec; 72886a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 72892b510759SStefano Zampini PetscInt its; 72906e683305SStefano Zampini PetscBool compute_eigs; 72916e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 72926e683305SStefano Zampini PetscInt neigs; 72938e185a42SStefano Zampini const char *prefix; 7294c8587f34SStefano Zampini 72952b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 72966e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 7297422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 729823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7299f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 7300e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 7301e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 7302e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 73032b510759SStefano Zampini if (ispreonly) { 73042b510759SStefano Zampini check_ksp_type = KSPPREONLY; 73056e683305SStefano Zampini compute_eigs = PETSC_FALSE; 73062b510759SStefano Zampini } else { 7307cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 73086e683305SStefano Zampini compute_eigs = PETSC_TRUE; 7309c8587f34SStefano Zampini } 7310c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 73116e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 73126e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 73136e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 7314a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 7315a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 7316a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 7317a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 7318c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 7319c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 7320c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 7321c8587f34SStefano Zampini /* create random vec */ 73222701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 7323c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 73246e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 7325c8587f34SStefano Zampini /* solve coarse problem */ 73266e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 7327cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 73286e683305SStefano Zampini if (compute_eigs) { 7329854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 7330854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 73316e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 73321ae86dd6SStefano Zampini if (neigs) { 73336e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 73346e683305SStefano Zampini lambda_min = eigs_r[0]; 73356e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 73362701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 73372701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 7338cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 7339cbcc2c2aSStefano Zampini } 7340c8587f34SStefano Zampini } 7341c8587f34SStefano Zampini } 73421ae86dd6SStefano Zampini } 7343cbcc2c2aSStefano Zampini 7344c8587f34SStefano Zampini /* check coarse problem residual error */ 73456e683305SStefano Zampini if (pcbddc->dbg_flag) { 73466e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 73476e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 73486e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 7349c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 73506e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 73516e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 7352779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 73536e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 73546e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 73556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 73566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 7357b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7358b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 7359b0f5fe93SStefano Zampini } 73606e683305SStefano Zampini if (compute_eigs) { 73616e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 7362deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 7363c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 73646e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 73656e683305SStefano 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); 73666e683305SStefano Zampini for (i=0;i<neigs;i++) { 73676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 7368c8587f34SStefano Zampini } 73696e683305SStefano Zampini } 73706e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 73716e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 73726e683305SStefano Zampini } 7373e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 73742701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 7375c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 73766e683305SStefano Zampini if (compute_eigs) { 73776e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 73786e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 7379c8587f34SStefano Zampini } 73806e683305SStefano Zampini } 73816e683305SStefano Zampini } 7382bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 7383cbcc2c2aSStefano Zampini /* print additional info */ 7384cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 73856e683305SStefano Zampini /* waits until all processes reaches this point */ 73866e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 7387cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 7388cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7389cbcc2c2aSStefano Zampini } 7390cbcc2c2aSStefano Zampini 73912b510759SStefano Zampini /* free memory */ 7392fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 7393c8587f34SStefano Zampini PetscFunctionReturn(0); 7394c8587f34SStefano Zampini } 7395674ae819SStefano Zampini 7396f34684f1SStefano Zampini #undef __FUNCT__ 7397f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 7398f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 7399f34684f1SStefano Zampini { 7400f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7401f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 7402f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 7403dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 7404dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 740573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 7406dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 7407f34684f1SStefano Zampini PetscErrorCode ierr; 7408f34684f1SStefano Zampini 7409f34684f1SStefano Zampini PetscFunctionBegin; 7410f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 74116c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 7412dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 74133bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 7414dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7415dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 74166583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 7417dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 7418dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 7419dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 74206c4ed002SBarry Smith if (local_size != pcbddc->local_primal_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %D != %D",local_size,pcbddc->local_primal_size); 7421dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 7422dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7423dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 7424dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7425dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7426f34684f1SStefano Zampini 7427f34684f1SStefano Zampini /* check numbering */ 7428f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 7429019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 7430dc456d91SStefano Zampini PetscInt i; 7431b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 7432f34684f1SStefano Zampini 7433f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7434f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 7435f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 74361575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7437019a44ceSStefano Zampini /* counter */ 7438019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7439019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 7440019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7441019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7442019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7443019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7444f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 7445f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 7446727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 7447f34684f1SStefano Zampini } 7448f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7449f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7450f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7451e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7452e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7453e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7454e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7455f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7456019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 7457f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 7458019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 74592c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 746075c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 7461b9b85e73SStefano Zampini set_error = PETSC_TRUE; 74622c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 74632c66d082SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d (gid %d) owned by %d processes instead of %d!\n",PetscGlobalRank,i,gi,owned,neigh);CHKERRQ(ierr); 7464f34684f1SStefano Zampini } 7465f34684f1SStefano Zampini } 7466019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 7467b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 7468f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7469f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 7470f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 7471f34684f1SStefano Zampini } 7472f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7473f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7474e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7475e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7476f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 7477f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 7478b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 7479ca8b9ea9SStefano Zampini PetscInt *gidxs; 7480ca8b9ea9SStefano Zampini 7481ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 74823bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 7483f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 7484f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7485f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 7486f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 74874bc2dc4bSStefano 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); 7488f34684f1SStefano Zampini } 7489f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7490ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 7491f34684f1SStefano Zampini } 7492f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 74931575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7494302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 7495f34684f1SStefano Zampini } 74968bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 7497f34684f1SStefano Zampini /* get back data */ 7498f34684f1SStefano Zampini *coarse_size_n = coarse_size; 7499f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 7500674ae819SStefano Zampini PetscFunctionReturn(0); 7501674ae819SStefano Zampini } 7502674ae819SStefano Zampini 7503e456f2a8SStefano Zampini #undef __FUNCT__ 7504e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 7505a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 7506e456f2a8SStefano Zampini { 7507e456f2a8SStefano Zampini IS localis_t; 7508a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 7509e456f2a8SStefano Zampini PetscScalar *vals; 7510e456f2a8SStefano Zampini PetscErrorCode ierr; 7511e456f2a8SStefano Zampini 7512e456f2a8SStefano Zampini PetscFunctionBegin; 7513a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 7514e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 7515854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 7516e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 7517e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 7518a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 7519a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 75201035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 7521a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 75221035eff8SStefano Zampini } 7523a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 7524e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 7525e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 7526a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 7527a7dc3881SStefano Zampini /* now compute set in local ordering */ 7528a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7529a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7530a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 7531a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 7532a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 7533ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 7534e456f2a8SStefano Zampini lsize++; 7535e456f2a8SStefano Zampini } 7536e456f2a8SStefano Zampini } 7537854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 7538a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 7539ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 7540e456f2a8SStefano Zampini idxs[lsize++] = i; 7541e456f2a8SStefano Zampini } 7542e456f2a8SStefano Zampini } 7543a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 7544a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 7545e456f2a8SStefano Zampini *localis = localis_t; 7546e456f2a8SStefano Zampini PetscFunctionReturn(0); 7547e456f2a8SStefano Zampini } 7548906d46d4SStefano Zampini 7549b96c3477SStefano Zampini #undef __FUNCT__ 7550b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 755108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 7552b96c3477SStefano Zampini { 7553a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 7554b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 7555b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 7556a64f4aa4SStefano Zampini Mat S_j; 7557b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 7558b96c3477SStefano Zampini PetscBool free_used_adj; 7559b96c3477SStefano Zampini PetscErrorCode ierr; 7560b96c3477SStefano Zampini 7561b96c3477SStefano Zampini PetscFunctionBegin; 7562b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 7563b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 756408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 7565b96c3477SStefano Zampini used_xadj = NULL; 7566b96c3477SStefano Zampini used_adjncy = NULL; 7567b96c3477SStefano Zampini } else { 756808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 756908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 757008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 757108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 7572b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 7573b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 7574b96c3477SStefano Zampini } else { 75752fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 7576b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 7577b96c3477SStefano Zampini PetscInt nvtxs; 7578b96c3477SStefano Zampini 75792fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 75802fffb893SStefano Zampini if (flg_row) { 7581b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 7582b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 7583b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 7584b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 75852fffb893SStefano Zampini } else { 75862fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 75872fffb893SStefano Zampini used_xadj = NULL; 75882fffb893SStefano Zampini used_adjncy = NULL; 75892fffb893SStefano Zampini } 75902fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 7591b96c3477SStefano Zampini } 7592b96c3477SStefano Zampini } 7593d5574798SStefano Zampini 7594d5574798SStefano Zampini /* setup sub_schurs data */ 7595a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 7596df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 7597df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 7598a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 759991af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 7600a64f4aa4SStefano Zampini } else { 76016816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 76024d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 7603a3df083aSStefano Zampini PetscInt benign_n; 760472b8c272SStefano Zampini Mat change = NULL; 76059d54b7f4SStefano Zampini Vec scaling = NULL; 760672b8c272SStefano Zampini IS change_primal = NULL; 7607a3df083aSStefano Zampini 76085feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 76095feab87aSStefano Zampini PetscInt n_vertices; 76105feab87aSStefano Zampini 76115feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 76122034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 76135feab87aSStefano Zampini } 761404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 761504708bb6SStefano Zampini if (!isseqaij) { 761604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 761704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 761804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 761904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 762004708bb6SStefano Zampini } else { 7621511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 762204708bb6SStefano Zampini } 762304708bb6SStefano Zampini } 7624a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 7625a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 7626ca92afb2SStefano Zampini } else { 7627a3df083aSStefano Zampini benign_n = 0; 7628ca92afb2SStefano Zampini } 7629b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 7630b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 7631b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 763272b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 763322db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 7634b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 763522db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 7636b7ab4a40SStefano Zampini } 7637b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 7638b7ab4a40SStefano Zampini We need to compute the change of basis according to the quadrature weights attached to pmat via MatSetNearNullSpace, and this could not be done (at the moment) without some hacking */ 7639b7ab4a40SStefano Zampini if (need_change) { 764088c03ad3SStefano Zampini PC_IS *pcisf; 764188c03ad3SStefano Zampini PC_BDDC *pcbddcf; 764288c03ad3SStefano Zampini PC pcf; 764388c03ad3SStefano Zampini 7644e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 764588c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 764688c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 764788c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 764888c03ad3SStefano Zampini /* hacks */ 764988c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 765072b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 765172b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 765272b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 765372b8c272SStefano Zampini pcisf->n = pcis->n; 765472b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 765588c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 765688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 765788c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 765888c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 765988c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 766088c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 766172b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 766288c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 766388c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 766472b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 766572b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 766672b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 766772b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 766872b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 766988c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 767072b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 767188c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 767288c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 767388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 767488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 767588c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 767688c03ad3SStefano Zampini pcf->ops->destroy = NULL; 767788c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 767888c03ad3SStefano Zampini } 76799d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 768091af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 768172b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 768272b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 7683ca92afb2SStefano Zampini } 7684d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 7685b96c3477SStefano Zampini 7686b96c3477SStefano Zampini /* free adjacency */ 7687b96c3477SStefano Zampini if (free_used_adj) { 7688b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 7689b96c3477SStefano Zampini } 7690b96c3477SStefano Zampini PetscFunctionReturn(0); 7691b96c3477SStefano Zampini } 7692b96c3477SStefano Zampini 7693b96c3477SStefano Zampini #undef __FUNCT__ 7694b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 769508122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 7696b96c3477SStefano Zampini { 7697b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 7698b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 7699b96c3477SStefano Zampini PCBDDCGraph graph; 7700b96c3477SStefano Zampini PetscErrorCode ierr; 7701b96c3477SStefano Zampini 7702b96c3477SStefano Zampini PetscFunctionBegin; 7703b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 770408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 77053301b35fSStefano Zampini IS verticesIS,verticescomm; 77063301b35fSStefano Zampini PetscInt vsize,*idxs; 7707b96c3477SStefano Zampini 7708b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 77093301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 77103301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 77113301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 77123301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 7713c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 7714b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 7715be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 7716441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 77173301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 7718b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 7719b96c3477SStefano Zampini } else { 7720b96c3477SStefano Zampini graph = pcbddc->mat_graph; 7721b96c3477SStefano Zampini } 7722e4d548c7SStefano Zampini /* print some info */ 7723*5c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 7724e4d548c7SStefano Zampini IS vertices; 7725e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 7726c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 7727e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 7728e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 7729e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7730e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 7731e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 7732e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 7733e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 7734e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7735e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7736c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 7737e4d548c7SStefano Zampini } 7738b96c3477SStefano Zampini 7739b96c3477SStefano Zampini /* sub_schurs init */ 7740b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 7741b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 7742b334f244SStefano Zampini } 77438b6046baSStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 7744a64f4aa4SStefano Zampini 7745b96c3477SStefano Zampini /* free graph struct */ 774608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 7747b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 7748b96c3477SStefano Zampini } 7749b96c3477SStefano Zampini PetscFunctionReturn(0); 7750b96c3477SStefano Zampini } 7751fa34dd3eSStefano Zampini 7752fa34dd3eSStefano Zampini #undef __FUNCT__ 7753fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 7754fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 7755fa34dd3eSStefano Zampini { 7756fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 7757fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 7758fa34dd3eSStefano Zampini PetscErrorCode ierr; 7759fa34dd3eSStefano Zampini 7760fa34dd3eSStefano Zampini PetscFunctionBegin; 7761fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 7762fa34dd3eSStefano Zampini IS zerodiag = NULL; 77634f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 7764fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 77654f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 776675c01103SStefano Zampini PetscReal norm; 7767fa34dd3eSStefano Zampini PetscInt i; 7768fa34dd3eSStefano Zampini 7769fa34dd3eSStefano Zampini /* B0 and B0_B */ 7770fa34dd3eSStefano Zampini if (zerodiag) { 7771fa34dd3eSStefano Zampini IS dummy; 7772fa34dd3eSStefano Zampini 77734f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 77744f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 7775fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 7776fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 7777fa34dd3eSStefano Zampini } 7778fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 7779fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 7780fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 7781fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7782fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7783fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7784fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7785fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 7786fa34dd3eSStefano Zampini /* S_j */ 7787fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 7788fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 7789fa34dd3eSStefano Zampini 7790fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 7791fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 7792fa34dd3eSStefano Zampini /* continuous in primal space */ 7793fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 7794fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7795fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7796fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 77974f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 77984f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 7799fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 7800fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 7801fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7802fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7803fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7804fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7805fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 7806fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 7807fa34dd3eSStefano Zampini 7808fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 7809fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 7810fa34dd3eSStefano Zampini /* local with Schur */ 7811fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 7812fa34dd3eSStefano Zampini if (zerodiag) { 7813fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 78144f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 7815fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 7816fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 7817fa34dd3eSStefano Zampini } 7818fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 7819fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7820fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7821fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7822fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 7823fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 7824fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 7825fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7826fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 7827fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7828fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7829fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7830fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7831fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 7832fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 7833fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 7834fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 7835fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 7836fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7837fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7838fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7839fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7840fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 7841fa34dd3eSStefano Zampini if (zerodiag) { 7842fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 7843fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 78444f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 7845fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 7846fa34dd3eSStefano Zampini } 7847fa34dd3eSStefano Zampini /* BDDC */ 7848fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 7849fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 7850fa34dd3eSStefano Zampini 7851fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 7852fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 7853fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 7854fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 78554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 78564f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 7857fa34dd3eSStefano Zampini } 78584f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 7859fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 7860fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 7861fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 7862fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 7863fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 7864fa34dd3eSStefano Zampini } 7865fa34dd3eSStefano Zampini PetscFunctionReturn(0); 7866fa34dd3eSStefano Zampini } 7867