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; 119b03ebc13SStefano Zampini PetscBT btv,bte,btvc,btb,btvcand,btvi,btee,bter; 120a13144ffSStefano Zampini PetscScalar *vals,*work; 121a13144ffSStefano Zampini PetscReal *rwork; 122a13144ffSStefano Zampini const PetscInt *idxs,*ii,*jj,*iit,*jjt; 123b03ebc13SStefano Zampini PetscInt ne,nv,Lv,order,n,field,rst; 124a13144ffSStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 125b03ebc13SStefano Zampini PetscInt i,j,extmem,cum,maxsize,nee,nquads=2; 126b03ebc13SStefano Zampini PetscInt *extrow,*extrowcum,*marks,*vmarks,*gidxs; 127a13144ffSStefano Zampini PetscInt *sfvleaves,*sfvroots; 128b03ebc13SStefano Zampini PetscInt *corners,*cedges; 129b03ebc13SStefano Zampini #if defined(PETSC_USE_DEBUG) 130b03ebc13SStefano Zampini PetscInt *emarks; 131b03ebc13SStefano Zampini #endif 132c2151214SStefano Zampini PetscBool print,eerr,done,lrc[2],conforming; 133a13144ffSStefano Zampini PetscErrorCode ierr; 134a13144ffSStefano Zampini 135a13144ffSStefano Zampini PetscFunctionBegin; 136a13144ffSStefano Zampini /* test variable order code and print debug info TODO: to be removed */ 137a13144ffSStefano Zampini print = PETSC_FALSE; 138a13144ffSStefano Zampini ierr = PetscOptionsGetBool(NULL,NULL,"-pc_bddc_print_nedelec",&print,NULL);CHKERRQ(ierr); 139a13144ffSStefano Zampini ierr = PetscOptionsGetInt(NULL,NULL,"-pc_bddc_nedelec_order",&pcbddc->nedorder,NULL);CHKERRQ(ierr); 140a13144ffSStefano Zampini 141a13144ffSStefano Zampini /* Return to caller if there are no edges in the decomposition */ 142a13144ffSStefano Zampini ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 143a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&el2g,NULL);CHKERRQ(ierr); 144c2151214SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(el2g,&n);CHKERRQ(ierr); 145a13144ffSStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 146a13144ffSStefano Zampini lrc[0] = PETSC_FALSE; 147c2151214SStefano Zampini for (i=0;i<n;i++) { 148a13144ffSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 149a13144ffSStefano Zampini lrc[0] = PETSC_TRUE; 150a13144ffSStefano Zampini break; 151a13144ffSStefano Zampini } 152a13144ffSStefano Zampini } 153a13144ffSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 154a13144ffSStefano Zampini ierr = MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 155a13144ffSStefano Zampini if (!lrc[1]) PetscFunctionReturn(0); 156a13144ffSStefano Zampini 157c2151214SStefano Zampini /* Get discrete gradient 158c2151214SStefano Zampini If it is defined for a subset of dofs, assumes G is given in global ordering for all the dofs */ 159a13144ffSStefano Zampini G = pcbddc->discretegradient; 160a13144ffSStefano Zampini order = pcbddc->nedorder; 161a13144ffSStefano Zampini conforming = pcbddc->conforming; 162c2151214SStefano Zampini field = pcbddc->nedfield; 163c2151214SStefano 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); 164c2151214SStefano Zampini if (pcbddc->n_ISForDofsLocal && field > -1) { 165c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);CHKERRQ(ierr); 166c2151214SStefano Zampini nedfieldlocal = pcbddc->ISForDofsLocal[field]; 167c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 168c2151214SStefano Zampini } else { 169c2151214SStefano Zampini PetscBool testnedfield = PETSC_FALSE; 170c2151214SStefano Zampini ierr = PetscOptionsGetBool(NULL,NULL,"-pc_bddc_nedelec_field",&testnedfield,NULL);CHKERRQ(ierr); 171c2151214SStefano Zampini if (!testnedfield) { 172c2151214SStefano Zampini ne = n; 173c2151214SStefano Zampini nedfieldlocal = NULL; 174c2151214SStefano Zampini } else { 175c2151214SStefano Zampini /* ierr = ISCreateStride(comm,n,0,1,&nedfieldlocal);CHKERRQ(ierr); */ 176c2151214SStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 177c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 178c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 179c2151214SStefano Zampini for (i=0;i<n;i++) matis->sf_leafdata[i] = 1; 180c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 181c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 182c2151214SStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 183c2151214SStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 184c2151214SStefano Zampini for (i=0,cum=0;i<n;i++) { 185c2151214SStefano Zampini if (matis->sf_leafdata[i] > 1) { 186c2151214SStefano Zampini matis->sf_leafdata[cum++] = i; 187c2151214SStefano Zampini } 188c2151214SStefano Zampini } 189c2151214SStefano Zampini ierr = ISCreateGeneral(comm,cum,matis->sf_leafdata,PETSC_COPY_VALUES,&nedfieldlocal);CHKERRQ(ierr); 190c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 191c2151214SStefano Zampini } 192c2151214SStefano Zampini } 193a13144ffSStefano Zampini 194c2151214SStefano Zampini if (nedfieldlocal) { /* merge with previous code when testing is done */ 195c2151214SStefano Zampini IS is; 196c2151214SStefano Zampini 197c2151214SStefano Zampini /* need to map from the local Nedelec field to local numbering */ 198c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);CHKERRQ(ierr); 199c2151214SStefano Zampini /* need to map from the local Nedelec field to global numbering */ 200c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(el2g,nedfieldlocal,&is);CHKERRQ(ierr); 201c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is,&el2g);CHKERRQ(ierr); 202c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 203c2151214SStefano Zampini } else { 204c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)el2g);CHKERRQ(ierr); 205c2151214SStefano Zampini fl2g = NULL; 206c2151214SStefano Zampini } 207a13144ffSStefano Zampini 208a13144ffSStefano Zampini /* Sanity checks */ 209a13144ffSStefano Zampini if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time"); 210a13144ffSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis"); 211a13144ffSStefano 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); 212a13144ffSStefano Zampini 213a13144ffSStefano Zampini /* Drop connections for interior edges (this modifies G) */ 214c2151214SStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 215c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 216c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 217c2151214SStefano Zampini if (nedfieldlocal) { 218c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 219c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1; 220c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 221c2151214SStefano Zampini } else { 222c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1; 223c2151214SStefano Zampini } 224c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 225c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 226c2151214SStefano Zampini ierr = MatGetOwnershipRange(G,&rst,NULL);CHKERRQ(ierr); 227c2151214SStefano Zampini for (i=0,cum=0;i<pc->pmat->rmap->n;i++) { 228c2151214SStefano Zampini if (matis->sf_rootdata[i] < 2) { 229c2151214SStefano Zampini matis->sf_rootdata[cum++] = i + rst; 230c2151214SStefano Zampini } 231c2151214SStefano Zampini } 232a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);CHKERRQ(ierr); 233a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 234c2151214SStefano Zampini ierr = MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);CHKERRQ(ierr); 235a13144ffSStefano Zampini 236a13144ffSStefano Zampini /* Extract subdomain relevant rows of G */ 237a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(el2g,&idxs);CHKERRQ(ierr); 238a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);CHKERRQ(ierr); 239a13144ffSStefano Zampini ierr = MatGetSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);CHKERRQ(ierr); 240a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);CHKERRQ(ierr); 241a13144ffSStefano Zampini ierr = ISDestroy(&lned);CHKERRQ(ierr); 242a13144ffSStefano Zampini ierr = MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);CHKERRQ(ierr); 243a13144ffSStefano Zampini ierr = MatDestroy(&lGall);CHKERRQ(ierr); 244a13144ffSStefano Zampini ierr = MatISGetLocalMat(lGis,&lG);CHKERRQ(ierr); 245a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)lG);CHKERRQ(ierr); 246a13144ffSStefano Zampini if (print) { 247a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"initial_lG");CHKERRQ(ierr); 248a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 249a13144ffSStefano Zampini } 250a13144ffSStefano Zampini 251a13144ffSStefano Zampini /* SF for nodal communications */ 252c2151214SStefano Zampini ierr = MatGetLocalSize(G,NULL,&Lv);CHKERRQ(ierr); 253a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);CHKERRQ(ierr); 254a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)vl2g);CHKERRQ(ierr); 255a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(vl2g,&nv);CHKERRQ(ierr); 256a13144ffSStefano Zampini ierr = PetscSFCreate(comm,&sfv);CHKERRQ(ierr); 257a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(vl2g,&idxs);CHKERRQ(ierr); 258a13144ffSStefano Zampini ierr = PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);CHKERRQ(ierr); 259a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);CHKERRQ(ierr); 260a13144ffSStefano Zampini ierr = PetscMalloc2(nv,&sfvleaves,Lv,&sfvroots);CHKERRQ(ierr); 261a13144ffSStefano Zampini 262a13144ffSStefano Zampini /* Destroy temporary G created in MATIS format */ 263a13144ffSStefano Zampini ierr = MatDestroy(&lGis);CHKERRQ(ierr); 264a13144ffSStefano Zampini 2650569b399SStefano Zampini /* Save lG */ 2660569b399SStefano Zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);CHKERRQ(ierr); 2670569b399SStefano Zampini 268a13144ffSStefano Zampini /* Analyze the edge-nodes connections (duplicate lG) */ 2694e64d54eSstefano_zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGe);CHKERRQ(ierr); 2704e64d54eSstefano_zampini ierr = MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 271a13144ffSStefano Zampini ierr = PetscBTCreate(nv,&btv);CHKERRQ(ierr); 272a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&bte);CHKERRQ(ierr); 273a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&btb);CHKERRQ(ierr); 274c2151214SStefano Zampini ierr = PetscBTCreate(nv,&btvcand);CHKERRQ(ierr); 275a13144ffSStefano Zampini /* need to import the boundary specification to ensure the 276a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 277a13144ffSStefano Zampini if (pcbddc->DirichletBoundariesLocal) { 278c2151214SStefano Zampini IS is; 279c2151214SStefano Zampini 280c2151214SStefano Zampini if (fl2g) { 281c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);CHKERRQ(ierr); 282c2151214SStefano Zampini } else { 283c2151214SStefano Zampini is = pcbddc->DirichletBoundariesLocal; 284c2151214SStefano Zampini } 285c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 286c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 287a13144ffSStefano Zampini for (i=0;i<cum;i++) { 288a13144ffSStefano Zampini if (idxs[i] >= 0) { 289a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 290a13144ffSStefano Zampini } 291a13144ffSStefano Zampini } 292c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 293c2151214SStefano Zampini if (fl2g) { 294c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 295c2151214SStefano Zampini } 296a13144ffSStefano Zampini } 297a13144ffSStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 298c2151214SStefano Zampini IS is; 299c2151214SStefano Zampini 300c2151214SStefano Zampini if (fl2g) { 301c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);CHKERRQ(ierr); 302c2151214SStefano Zampini } else { 303c2151214SStefano Zampini is = pcbddc->NeumannBoundariesLocal; 304c2151214SStefano Zampini } 305c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 306c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 307a13144ffSStefano Zampini for (i=0;i<cum;i++) { 308a13144ffSStefano Zampini if (idxs[i] >= 0) { 309a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 310a13144ffSStefano Zampini } 311a13144ffSStefano Zampini } 312c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 313c2151214SStefano Zampini if (fl2g) { 314c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 315a13144ffSStefano Zampini } 316c2151214SStefano Zampini } 317c2151214SStefano Zampini 318a13144ffSStefano Zampini /* need to remove coarse faces' dofs to ensure the 319a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 320a13144ffSStefano Zampini ierr = PetscCalloc1(ne,&marks);CHKERRQ(ierr); 32162b0c6f7SStefano Zampini ierr = PetscBTCreate(ne,&btee);CHKERRQ(ierr); 322a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 32362b0c6f7SStefano Zampini for (i=1;i<n_neigh;i++) 32462b0c6f7SStefano Zampini for (j=0;j<n_shared[i];j++) 32562b0c6f7SStefano Zampini marks[shared[i][j]]++; 326a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 32762b0c6f7SStefano Zampini for (i=0;i<ne;i++) { 32862b0c6f7SStefano Zampini if (marks[i] > 1 || (marks[i] == 1 && PetscBTLookup(btb,i))) { 32962b0c6f7SStefano Zampini ierr = PetscBTSet(btee,i);CHKERRQ(ierr); 33062b0c6f7SStefano Zampini } 33162b0c6f7SStefano Zampini } 33262b0c6f7SStefano Zampini 33362b0c6f7SStefano Zampini if (!conforming) { 33462b0c6f7SStefano Zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 33562b0c6f7SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 33662b0c6f7SStefano Zampini } 3374e64d54eSstefano_zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 338dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGe,&vals);CHKERRQ(ierr); 33962b0c6f7SStefano Zampini cum = 0; 340a13144ffSStefano Zampini for (i=0;i<ne;i++) { 341dec27d64SStefano Zampini /* eliminate rows corresponding to edge dofs belonging to coarse faces */ 34262b0c6f7SStefano Zampini if (!PetscBTLookup(btee,i)) { 343a13144ffSStefano Zampini marks[cum++] = i; 344dec27d64SStefano Zampini continue; 345dec27d64SStefano Zampini } 346dec27d64SStefano Zampini /* set badly connected edge dofs as primal */ 34762b0c6f7SStefano Zampini if (!conforming) { 34862b0c6f7SStefano Zampini if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */ 349a13144ffSStefano Zampini marks[cum++] = i; 350a13144ffSStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 351a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 352a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 353a13144ffSStefano Zampini } 35462b0c6f7SStefano Zampini } else { 35562b0c6f7SStefano Zampini /* every edge dofs should be connected trough a certain number of nodal dofs 35662b0c6f7SStefano Zampini to other edge dofs belonging to coarse edges 35762b0c6f7SStefano Zampini - at most 2 endpoints 35862b0c6f7SStefano Zampini - order-1 interior nodal dofs 35962b0c6f7SStefano Zampini - no undefined nodal dofs (nconn < order) 36062b0c6f7SStefano Zampini */ 36162b0c6f7SStefano Zampini PetscInt ends = 0,ints = 0, undef = 0; 36262b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 36362b0c6f7SStefano Zampini PetscInt v = jj[j],k; 36462b0c6f7SStefano Zampini PetscInt nconn = iit[v+1]-iit[v]; 36562b0c6f7SStefano Zampini for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--; 36662b0c6f7SStefano Zampini if (nconn > order) ends++; 36762b0c6f7SStefano Zampini else if (nconn == order) ints++; 36862b0c6f7SStefano Zampini else undef++; 36962b0c6f7SStefano Zampini } 37062b0c6f7SStefano Zampini if (undef || ends > 2 || ints != order -1) { 37162b0c6f7SStefano Zampini marks[cum++] = i; 37262b0c6f7SStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 37362b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 37462b0c6f7SStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 37562b0c6f7SStefano Zampini } 37662b0c6f7SStefano Zampini } 37762b0c6f7SStefano Zampini } 378a13144ffSStefano Zampini } 379dec27d64SStefano Zampini /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */ 380dec27d64SStefano Zampini if (!order && ii[i+1] != ii[i]) { 381dec27d64SStefano Zampini PetscScalar val = 1./(ii[i+1]-ii[i]-1); 382dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vals[j] = val; 383a13144ffSStefano Zampini } 384dec27d64SStefano Zampini } 38562b0c6f7SStefano Zampini ierr = PetscBTDestroy(&btee);CHKERRQ(ierr); 386dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGe,&vals);CHKERRQ(ierr); 3874e64d54eSstefano_zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 38862b0c6f7SStefano Zampini if (!conforming) { 38962b0c6f7SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 39062b0c6f7SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 39162b0c6f7SStefano Zampini } 3924e64d54eSstefano_zampini ierr = MatZeroRows(lGe,cum,marks,0.,NULL,NULL);CHKERRQ(ierr); 393b03ebc13SStefano Zampini /* identify splitpoints and corner candidates */ 3944e64d54eSstefano_zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 395a13144ffSStefano Zampini if (print) { 3964e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");CHKERRQ(ierr); 3974e64d54eSstefano_zampini ierr = MatView(lGe,NULL);CHKERRQ(ierr); 3984e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");CHKERRQ(ierr); 399a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 400a13144ffSStefano Zampini } 401a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 402dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGt,&vals);CHKERRQ(ierr); 403a13144ffSStefano Zampini for (i=0;i<nv;i++) { 404dec27d64SStefano Zampini PetscInt ord = order, test = ii[i+1]-ii[i]; 405b03ebc13SStefano Zampini if (!order) { /* variable order */ 406dec27d64SStefano Zampini PetscReal vorder = 0.; 407dec27d64SStefano Zampini 408dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]); 409dec27d64SStefano Zampini test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON); 410dec27d64SStefano Zampini if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%d)",vorder,test); 411dec27d64SStefano Zampini ord = 1; 412dec27d64SStefano Zampini } 413a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 414dec27d64SStefano 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); 415a13144ffSStefano Zampini #endif 416dec27d64SStefano Zampini if (test >= 3*ord) { /* splitpoints */ 417a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %d\n",i); 418a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 419dec27d64SStefano Zampini } else if (test == ord) { 420b03ebc13SStefano Zampini if (order == 1 || (!order && ii[i+1]-ii[i] == 1)) { 421a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %d\n",i); 422a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 423a13144ffSStefano Zampini } else { 424a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %d\n",i); 425a13144ffSStefano Zampini ierr = PetscBTSet(btvcand,i);CHKERRQ(ierr); 426a13144ffSStefano Zampini } 427a13144ffSStefano Zampini } 428a13144ffSStefano Zampini } 429b03ebc13SStefano Zampini 430b03ebc13SStefano Zampini /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */ 431b03ebc13SStefano Zampini if (order != 1) { 432b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"INSPECTING CANDIDATES\n"); 433b03ebc13SStefano Zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 434b03ebc13SStefano Zampini for (i=0;i<nv;i++) { 435b03ebc13SStefano Zampini if (PetscBTLookup(btvcand,i)) { 436b03ebc13SStefano Zampini PetscBool found = PETSC_FALSE; 437b03ebc13SStefano Zampini for (j=ii[i];j<ii[i+1] && !found;j++) { 438b03ebc13SStefano Zampini PetscInt k,e = jj[j]; 439b03ebc13SStefano Zampini if (PetscBTLookup(bte,e)) continue; 440b03ebc13SStefano Zampini for (k=iit[e];k<iit[e+1];k++) { 441b03ebc13SStefano Zampini PetscInt v = jjt[k]; 442b03ebc13SStefano Zampini if (v != i && PetscBTLookup(btvcand,v)) { 443b03ebc13SStefano Zampini found = PETSC_TRUE; 444b03ebc13SStefano Zampini break; 445b03ebc13SStefano Zampini } 446b03ebc13SStefano Zampini } 447b03ebc13SStefano Zampini } 448b03ebc13SStefano Zampini if (!found) { 449b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d CLEARED\n",i); 450b03ebc13SStefano Zampini ierr = PetscBTClear(btvcand,i);CHKERRQ(ierr); 451b03ebc13SStefano Zampini } else { 452b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d ACCEPTED\n",i); 453b03ebc13SStefano Zampini } 454b03ebc13SStefano Zampini } 455b03ebc13SStefano Zampini } 456b03ebc13SStefano Zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 457b03ebc13SStefano Zampini } 458dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGt,&vals);CHKERRQ(ierr); 459a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 460b03ebc13SStefano Zampini ierr = MatDestroy(&lGe);CHKERRQ(ierr); 461a13144ffSStefano Zampini 462a13144ffSStefano Zampini /* Get the local G^T explicitly */ 4630569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 464a13144ffSStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 4654e64d54eSstefano_zampini ierr = MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 466a13144ffSStefano Zampini 4674e64d54eSstefano_zampini /* Mark interior nodal dofs */ 468a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 4694e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvi);CHKERRQ(ierr); 470a13144ffSStefano Zampini for (i=1;i<n_neigh;i++) { 471a13144ffSStefano Zampini for (j=0;j<n_shared[i];j++) { 4724e64d54eSstefano_zampini ierr = PetscBTSet(btvi,shared[i][j]);CHKERRQ(ierr); 473a13144ffSStefano Zampini } 474a13144ffSStefano Zampini } 475a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 476a13144ffSStefano Zampini 477a13144ffSStefano Zampini /* communicate corners and splitpoints */ 478a13144ffSStefano Zampini ierr = PetscMalloc1(nv,&vmarks);CHKERRQ(ierr); 479a13144ffSStefano Zampini ierr = PetscMemzero(sfvleaves,nv*sizeof(PetscInt));CHKERRQ(ierr); 480a13144ffSStefano Zampini ierr = PetscMemzero(sfvroots,Lv*sizeof(PetscInt));CHKERRQ(ierr); 481a13144ffSStefano Zampini for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1; 482a13144ffSStefano Zampini 483a13144ffSStefano Zampini if (print) { 484a13144ffSStefano Zampini IS tbz; 485a13144ffSStefano Zampini 486a13144ffSStefano Zampini cum = 0; 487a13144ffSStefano Zampini for (i=0;i<nv;i++) 488a13144ffSStefano Zampini if (sfvleaves[i]) 489a13144ffSStefano Zampini vmarks[cum++] = i; 490a13144ffSStefano Zampini 491a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 492a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");CHKERRQ(ierr); 493a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 494a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 495a13144ffSStefano Zampini } 496a13144ffSStefano Zampini 497a13144ffSStefano Zampini ierr = PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 498a13144ffSStefano Zampini ierr = PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 499a13144ffSStefano Zampini ierr = PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 500a13144ffSStefano Zampini ierr = PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 501a13144ffSStefano Zampini 5024e64d54eSstefano_zampini /* Zero rows of lGt corresponding to identified corners 5034e64d54eSstefano_zampini and interior nodal dofs */ 504a13144ffSStefano Zampini cum = 0; 505a13144ffSStefano Zampini for (i=0;i<nv;i++) { 506a13144ffSStefano Zampini if (sfvleaves[i]) { 507a13144ffSStefano Zampini vmarks[cum++] = i; 508a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 509a13144ffSStefano Zampini } 5104e64d54eSstefano_zampini if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i; 511a13144ffSStefano Zampini } 5124e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvi);CHKERRQ(ierr); 513a13144ffSStefano Zampini if (print) { 514a13144ffSStefano Zampini IS tbz; 515a13144ffSStefano Zampini 516a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 5174e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");CHKERRQ(ierr); 518a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 519a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 520a13144ffSStefano Zampini } 521a13144ffSStefano Zampini ierr = MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);CHKERRQ(ierr); 522a13144ffSStefano Zampini ierr = PetscFree(vmarks);CHKERRQ(ierr); 523a13144ffSStefano Zampini ierr = PetscSFDestroy(&sfv);CHKERRQ(ierr); 524a13144ffSStefano Zampini ierr = PetscFree2(sfvleaves,sfvroots);CHKERRQ(ierr); 525a13144ffSStefano Zampini 526a13144ffSStefano Zampini /* Recompute G */ 527a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 528a13144ffSStefano Zampini ierr = MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);CHKERRQ(ierr); 529a13144ffSStefano Zampini if (print) { 530a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"used_lG");CHKERRQ(ierr); 531a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 532a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lGt,"used_lGt");CHKERRQ(ierr); 533a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 534a13144ffSStefano Zampini } 535a13144ffSStefano Zampini 536a13144ffSStefano Zampini /* Get primal dofs (if any) */ 537a13144ffSStefano Zampini cum = 0; 538a13144ffSStefano Zampini for (i=0;i<ne;i++) { 539a13144ffSStefano Zampini if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i; 540a13144ffSStefano Zampini } 541c2151214SStefano Zampini if (fl2g) { 542c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);CHKERRQ(ierr); 543c2151214SStefano Zampini } 544a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 545a13144ffSStefano Zampini if (print) { 546a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");CHKERRQ(ierr); 547a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 548a13144ffSStefano Zampini } 549a13144ffSStefano Zampini ierr = PetscBTDestroy(&bte);CHKERRQ(ierr); 550c2151214SStefano Zampini /* TODO: what if the user passed in some of them ? */ 551a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 552a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 553a13144ffSStefano Zampini 554a13144ffSStefano Zampini /* Compute edge connectivity */ 555a13144ffSStefano Zampini ierr = PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");CHKERRQ(ierr); 556a13144ffSStefano Zampini ierr = MatMatMultSymbolic(lG,lGt,PETSC_DEFAULT,&conn);CHKERRQ(ierr); 557a13144ffSStefano Zampini ierr = MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 558d904f53bSStefano Zampini if (done && i && ii && jj) { /* when lG is empty, don't pass pointers */ 559c2151214SStefano Zampini if (fl2g) { 560c2151214SStefano Zampini PetscBT btf; 561c2151214SStefano Zampini PetscInt *iia,*jja,*iiu,*jju; 562c2151214SStefano Zampini PetscBool rest = PETSC_FALSE,free = PETSC_FALSE; 563c2151214SStefano Zampini 564c2151214SStefano Zampini /* create CSR for all local dofs */ 565c2151214SStefano Zampini ierr = PetscMalloc1(n+1,&iia);CHKERRQ(ierr); 566c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */ 567c2151214SStefano 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); 568c2151214SStefano Zampini iiu = pcbddc->mat_graph->xadj; 569c2151214SStefano Zampini jju = pcbddc->mat_graph->adjncy; 570c2151214SStefano Zampini } else if (pcbddc->use_local_adj) { 571c2151214SStefano Zampini rest = PETSC_TRUE; 572c2151214SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 573c2151214SStefano Zampini } else { 574c2151214SStefano Zampini free = PETSC_TRUE; 575c2151214SStefano Zampini ierr = PetscMalloc2(n+1,&iiu,n,&jju);CHKERRQ(ierr); 576c2151214SStefano Zampini iiu[0] = 0; 577c2151214SStefano Zampini for (i=0;i<n;i++) { 578c2151214SStefano Zampini iiu[i+1] = i+1; 579c2151214SStefano Zampini jju[i] = -1; 580d904f53bSStefano Zampini } 581c2151214SStefano Zampini } 582c2151214SStefano Zampini 583c2151214SStefano Zampini /* import sizes of CSR */ 584c2151214SStefano Zampini iia[0] = 0; 585c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i]; 586c2151214SStefano Zampini 587c2151214SStefano Zampini /* overwrite entries corresponding to the Nedelec field */ 588c2151214SStefano Zampini ierr = PetscBTCreate(n,&btf);CHKERRQ(ierr); 589c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 590c2151214SStefano Zampini for (i=0;i<ne;i++) { 591c2151214SStefano Zampini ierr = PetscBTSet(btf,idxs[i]);CHKERRQ(ierr); 592c2151214SStefano Zampini iia[idxs[i]+1] = ii[i+1]-ii[i]; 593c2151214SStefano Zampini } 594c2151214SStefano Zampini 595c2151214SStefano Zampini /* iia in CSR */ 596c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] += iia[i]; 597c2151214SStefano Zampini 598c2151214SStefano Zampini /* jja in CSR */ 599c2151214SStefano Zampini ierr = PetscMalloc1(iia[n],&jja);CHKERRQ(ierr); 600c2151214SStefano Zampini for (i=0;i<n;i++) 601c2151214SStefano Zampini if (!PetscBTLookup(btf,i)) 602c2151214SStefano Zampini for (j=0;j<iiu[i+1]-iiu[i];j++) 603c2151214SStefano Zampini jja[iia[i]+j] = jju[iiu[i]+j]; 604c2151214SStefano Zampini 605c2151214SStefano Zampini /* map edge dofs connectivity */ 606c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);CHKERRQ(ierr); 607c2151214SStefano Zampini for (i=0;i<ne;i++) { 608c2151214SStefano Zampini PetscInt e = idxs[i]; 609c2151214SStefano Zampini for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j]; 610c2151214SStefano Zampini } 611c2151214SStefano Zampini 612c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 613c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);CHKERRQ(ierr); 614c2151214SStefano Zampini if (rest) { 615c2151214SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 616c2151214SStefano Zampini } 617c2151214SStefano Zampini if (free) { 618c2151214SStefano Zampini ierr = PetscFree2(iiu,jju);CHKERRQ(ierr); 619c2151214SStefano Zampini } 620c2151214SStefano Zampini ierr = PetscBTDestroy(&btf);CHKERRQ(ierr); 621c2151214SStefano Zampini } else { 622c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);CHKERRQ(ierr); 623c2151214SStefano Zampini } 624c2151214SStefano Zampini } 625c2151214SStefano Zampini 626a13144ffSStefano Zampini /* Analyze interface for edge dofs */ 627a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 628a13144ffSStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 629a13144ffSStefano Zampini 630a13144ffSStefano Zampini /* Get coarse edges in the edge space */ 631c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 632a13144ffSStefano Zampini ierr = MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 633a13144ffSStefano Zampini 634c2151214SStefano Zampini if (fl2g) { 635c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 636c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 637c2151214SStefano Zampini for (i=0;i<nee;i++) { 638c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 639c2151214SStefano Zampini } 640c2151214SStefano Zampini } else { 641c2151214SStefano Zampini eedges = alleedges; 642c2151214SStefano Zampini primals = allprimals; 643c2151214SStefano Zampini } 644c2151214SStefano Zampini 645a13144ffSStefano Zampini /* Mark fine edge dofs with their coarse edge id */ 646a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 647c2151214SStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 648c2151214SStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 649c2151214SStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = nee+1; 650c2151214SStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 651c2151214SStefano Zampini if (print) { 652c2151214SStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");CHKERRQ(ierr); 653c2151214SStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 654c2151214SStefano Zampini } 655c2151214SStefano Zampini 656c2151214SStefano Zampini maxsize = 0; 657a13144ffSStefano Zampini for (i=0;i<nee;i++) { 658a13144ffSStefano Zampini PetscInt size,mark = i+1; 659a13144ffSStefano Zampini 660a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 661a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 662a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 663a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 664a13144ffSStefano Zampini maxsize = PetscMax(maxsize,size); 665a13144ffSStefano Zampini } 666a13144ffSStefano Zampini 667a13144ffSStefano Zampini /* Find coarse edge endpoints */ 668a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 669a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 670a13144ffSStefano Zampini for (i=0;i<nee;i++) { 671a13144ffSStefano Zampini PetscInt mark = i+1,size; 672a13144ffSStefano Zampini 673a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 674a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 675a13144ffSStefano Zampini if (print) { 676a13144ffSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %d\n",i); 677a13144ffSStefano Zampini ISView(eedges[i],NULL); 678a13144ffSStefano Zampini } 679a13144ffSStefano Zampini for (j=0;j<size;j++) { 680a13144ffSStefano Zampini PetscInt k, ee = idxs[j]; 681a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," idx %d\n",ee); 682a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 683a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," inspect %d\n",jj[k]); 684a13144ffSStefano Zampini if (PetscBTLookup(btv,jj[k])) { 685a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %d\n",jj[k]); 686a13144ffSStefano Zampini } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */ 687a13144ffSStefano Zampini PetscInt k2; 688a13144ffSStefano Zampini PetscBool corner = PETSC_FALSE; 689a13144ffSStefano Zampini for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) { 690c2151214SStefano 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])); 691c2151214SStefano Zampini /* it's a corner if either is connected with an edge dof belonging to a different cc or 692c2151214SStefano Zampini if the edge dof lie on the natural part of the boundary */ 693c2151214SStefano Zampini if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) { 694a13144ffSStefano Zampini corner = PETSC_TRUE; 695a13144ffSStefano Zampini break; 696a13144ffSStefano Zampini } 697a13144ffSStefano Zampini } 698a13144ffSStefano Zampini if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */ 699a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found %d\n",jj[k]); 700a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 701a13144ffSStefano Zampini } else { 702a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n"); 703a13144ffSStefano Zampini } 704a13144ffSStefano Zampini } 705a13144ffSStefano Zampini } 706a13144ffSStefano Zampini } 707a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 708a13144ffSStefano Zampini } 709a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 710a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 711c2151214SStefano Zampini ierr = PetscBTDestroy(&btb);CHKERRQ(ierr); 712a13144ffSStefano Zampini 713a13144ffSStefano Zampini /* Reset marked primal dofs */ 714a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 715a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 716a13144ffSStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = 0; 717a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 718a13144ffSStefano Zampini 7190569b399SStefano Zampini /* Now use the initial lG */ 7200569b399SStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 7210569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 7220569b399SStefano Zampini lG = lGinit; 7230569b399SStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 7240569b399SStefano Zampini 725a13144ffSStefano Zampini /* Compute extended cols indices */ 726b03ebc13SStefano Zampini ierr = PetscBTCreate(nv,&btvc);CHKERRQ(ierr); 727b03ebc13SStefano Zampini ierr = PetscBTCreate(nee,&bter);CHKERRQ(ierr); 728a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 729a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lG,&i);CHKERRQ(ierr); 730a13144ffSStefano Zampini i *= maxsize; 731b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 732a13144ffSStefano Zampini ierr = PetscMalloc2(i,&extrow,i,&gidxs);CHKERRQ(ierr); 733a13144ffSStefano Zampini eerr = PETSC_FALSE; 734a13144ffSStefano Zampini for (i=0;i<nee;i++) { 735b03ebc13SStefano Zampini PetscInt size,found = 0; 736a13144ffSStefano Zampini 737a13144ffSStefano Zampini cum = 0; 738a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 739a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 740b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 741a13144ffSStefano Zampini for (j=0;j<size;j++) { 742a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 743b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 744b03ebc13SStefano Zampini PetscInt vv = jj[k]; 745b03ebc13SStefano Zampini if (!PetscBTLookup(btv,vv)) extrow[cum++] = vv; 746b03ebc13SStefano Zampini else if (!PetscBTLookupSet(btvc,vv)) found++; 747b03ebc13SStefano Zampini } 748a13144ffSStefano Zampini } 749a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 750a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 751a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 752a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 753a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 754a13144ffSStefano Zampini /* it may happen that endpoints are not defined at this point 755a13144ffSStefano Zampini if it is the case, mark this edge for a second pass */ 756b03ebc13SStefano Zampini if (cum != size -1 || found != 2) { 757b03ebc13SStefano Zampini ierr = PetscBTSet(bter,i);CHKERRQ(ierr); 758a13144ffSStefano Zampini if (print) { 759a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge");CHKERRQ(ierr); 760a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 761a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol");CHKERRQ(ierr); 762a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 763a13144ffSStefano Zampini } 764a13144ffSStefano Zampini eerr = PETSC_TRUE; 765a13144ffSStefano Zampini } 766a13144ffSStefano Zampini } 7674e64d54eSstefano_zampini /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */ 768a13144ffSStefano Zampini ierr = MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 769a13144ffSStefano Zampini if (done) { 770a13144ffSStefano Zampini PetscInt *newprimals; 771a13144ffSStefano Zampini 772a13144ffSStefano Zampini ierr = PetscMalloc1(ne,&newprimals);CHKERRQ(ierr); 773a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 774a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 775a13144ffSStefano Zampini ierr = PetscMemcpy(newprimals,idxs,cum*sizeof(PetscInt));CHKERRQ(ierr); 776a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 7770569b399SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 778b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"DOING SECOND PASS (eerr %d)\n",eerr); 779a13144ffSStefano Zampini for (i=0;i<nee;i++) { 780b03ebc13SStefano Zampini PetscBool has_candidates = PETSC_FALSE; 781b03ebc13SStefano Zampini if (PetscBTLookup(bter,i)) { 782a13144ffSStefano Zampini PetscInt size,mark = i+1; 783a13144ffSStefano Zampini 784a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 785a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 786c2151214SStefano Zampini /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */ 787a13144ffSStefano Zampini for (j=0;j<size;j++) { 788a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 789b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"Inspecting edge dof %d [%d %d)\n",ee,ii[ee],ii[ee+1]); 790a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 791a13144ffSStefano Zampini /* set all candidates located on the edge as corners */ 792a13144ffSStefano Zampini if (PetscBTLookup(btvcand,jj[k])) { 793a13144ffSStefano Zampini PetscInt k2,vv = jj[k]; 794b03ebc13SStefano Zampini has_candidates = PETSC_TRUE; 795b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Candidate set to vertex %d\n",vv); 796a13144ffSStefano Zampini ierr = PetscBTSet(btv,vv);CHKERRQ(ierr); 797a13144ffSStefano Zampini /* set all edge dofs connected to candidate as primals */ 798a13144ffSStefano Zampini for (k2=iit[vv];k2<iit[vv+1];k2++) { 799a13144ffSStefano Zampini if (marks[jjt[k2]] == mark) { 800a13144ffSStefano Zampini PetscInt k3,ee2 = jjt[k2]; 801b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected edge dof set to primal %d\n",ee2); 802a13144ffSStefano Zampini newprimals[cum++] = ee2; 803a13144ffSStefano Zampini /* finally set the new corners */ 804a13144ffSStefano Zampini for (k3=ii[ee2];k3<ii[ee2+1];k3++) { 805b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected nodal dof set to vertex %d\n",jj[k3]); 806a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k3]);CHKERRQ(ierr); 807a13144ffSStefano Zampini } 808a13144ffSStefano Zampini } 809a13144ffSStefano Zampini } 810b03ebc13SStefano Zampini } else { 811b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Not a candidate vertex %d\n",jj[k]); 812a13144ffSStefano Zampini } 813a13144ffSStefano Zampini } 814a13144ffSStefano Zampini } 815b03ebc13SStefano Zampini if (!has_candidates) { /* circular edge */ 816b03ebc13SStefano Zampini PetscInt k, ee = idxs[0],*tmarks; 817b03ebc13SStefano Zampini 818b03ebc13SStefano Zampini ierr = PetscCalloc1(ne,&tmarks);CHKERRQ(ierr); 819b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Circular edge %d\n",i); 820b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 821b03ebc13SStefano Zampini PetscInt k2; 822b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Set to corner %d\n",jj[k]); 823b03ebc13SStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 824b03ebc13SStefano Zampini for (k2=iit[jj[k]];k2<iit[jj[k]+1];k2++) tmarks[jjt[k2]]++; 825b03ebc13SStefano Zampini } 826b03ebc13SStefano Zampini for (j=0;j<size;j++) { 827b03ebc13SStefano Zampini if (tmarks[idxs[j]] > 1) { 828b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Edge dof set to primal %d\n",idxs[j]); 829b03ebc13SStefano Zampini newprimals[cum++] = idxs[j]; 830b03ebc13SStefano Zampini } 831b03ebc13SStefano Zampini } 832b03ebc13SStefano Zampini ierr = PetscFree(tmarks);CHKERRQ(ierr); 833b03ebc13SStefano Zampini } 834a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 835a13144ffSStefano Zampini } 836a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 837a13144ffSStefano Zampini } 838b03ebc13SStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 8390569b399SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 840a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,newprimals);CHKERRQ(ierr); 841c2151214SStefano Zampini if (fl2g) { 842c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);CHKERRQ(ierr); 843c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 844c2151214SStefano Zampini for (i=0;i<nee;i++) { 845c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 846c2151214SStefano Zampini } 847c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 848c2151214SStefano Zampini } 849c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 850a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 851a13144ffSStefano Zampini ierr = PetscFree(newprimals);CHKERRQ(ierr); 852a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 853a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 854a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 855c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 856c2151214SStefano Zampini if (fl2g) { 857c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 858c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 859c2151214SStefano Zampini for (i=0;i<nee;i++) { 860c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 861c2151214SStefano Zampini } 862c2151214SStefano Zampini } else { 863c2151214SStefano Zampini eedges = alleedges; 864c2151214SStefano Zampini primals = allprimals; 865c2151214SStefano Zampini } 866b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 867a13144ffSStefano Zampini 868a13144ffSStefano Zampini /* Mark again */ 869a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 870a13144ffSStefano Zampini for (i=0;i<nee;i++) { 871a13144ffSStefano Zampini PetscInt size,mark = i+1; 872a13144ffSStefano Zampini 873a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 874a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 875a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 876a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 877a13144ffSStefano Zampini } 878a13144ffSStefano Zampini if (print) { 879a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");CHKERRQ(ierr); 880a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 881a13144ffSStefano Zampini } 882a13144ffSStefano Zampini 883a13144ffSStefano Zampini /* Recompute extended cols */ 884a13144ffSStefano Zampini eerr = PETSC_FALSE; 885a13144ffSStefano Zampini for (i=0;i<nee;i++) { 886a13144ffSStefano Zampini PetscInt size; 887a13144ffSStefano Zampini 888a13144ffSStefano Zampini cum = 0; 889a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 890a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 891a13144ffSStefano Zampini for (j=0;j<size;j++) { 892a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 893a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 894a13144ffSStefano Zampini if (!PetscBTLookup(btv,jj[k])) { 895a13144ffSStefano Zampini extrow[cum++] = jj[k]; 896a13144ffSStefano Zampini } 897a13144ffSStefano Zampini } 898a13144ffSStefano Zampini } 899a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 900a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 901a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 902a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 903a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 904a13144ffSStefano Zampini if (cum != size -1) { 905a13144ffSStefano Zampini if (print) { 906a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");CHKERRQ(ierr); 907a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 908a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");CHKERRQ(ierr); 909a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 910a13144ffSStefano Zampini } 911a13144ffSStefano Zampini eerr = PETSC_TRUE; 912a13144ffSStefano Zampini } 913a13144ffSStefano Zampini } 914a13144ffSStefano Zampini } 915a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 916a13144ffSStefano Zampini ierr = PetscFree2(extrow,gidxs);CHKERRQ(ierr); 917b03ebc13SStefano Zampini ierr = PetscBTDestroy(&bter);CHKERRQ(ierr); 918a13144ffSStefano Zampini /* an error should not occur at this point */ 919a13144ffSStefano Zampini if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS"); 920b03ebc13SStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 921a13144ffSStefano Zampini 9224e64d54eSstefano_zampini /* Check the number of endpoints */ 9230569b399SStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 924b03ebc13SStefano Zampini ierr = PetscMalloc1(2*nee,&corners);CHKERRQ(ierr); 925b03ebc13SStefano Zampini ierr = PetscMalloc1(nee,&cedges);CHKERRQ(ierr); 9264e64d54eSstefano_zampini for (i=0;i<nee;i++) { 927b03ebc13SStefano Zampini PetscInt size, found = 0, gc[2]; 9284e64d54eSstefano_zampini 929b03ebc13SStefano Zampini /* init with defaults */ 930b03ebc13SStefano Zampini cedges[i] = corners[i*2] = corners[i*2+1] = -1; 9314e64d54eSstefano_zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 9324e64d54eSstefano_zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 933b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 9344e64d54eSstefano_zampini for (j=0;j<size;j++) { 9354e64d54eSstefano_zampini PetscInt k,ee = idxs[j]; 9364e64d54eSstefano_zampini for (k=ii[ee];k<ii[ee+1];k++) { 9374e64d54eSstefano_zampini PetscInt vv = jj[k]; 9384e64d54eSstefano_zampini if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) { 939b03ebc13SStefano Zampini if (found == 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found more then two corners for edge %d\n",i); 940b03ebc13SStefano Zampini corners[i*2+found++] = vv; 9414e64d54eSstefano_zampini } 9424e64d54eSstefano_zampini } 9434e64d54eSstefano_zampini } 944b03ebc13SStefano Zampini if (found != 2) { 945b03ebc13SStefano Zampini PetscInt e; 946b03ebc13SStefano Zampini if (fl2g) { 947b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,1,idxs,&e);CHKERRQ(ierr); 948b03ebc13SStefano Zampini } else { 949b03ebc13SStefano Zampini e = idxs[0]; 950b03ebc13SStefano Zampini } 951b03ebc13SStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %d corners for edge %d (astart %d, estart %d)\n",found,i,e,idxs[0]); 952b03ebc13SStefano Zampini } 953b03ebc13SStefano Zampini /* WARNING : this depends on how pcbddc->primal_indices_local_idxs is filled up in PCBDDConstraintsSetUp */ 954b03ebc13SStefano Zampini cedges[i] = idxs[0]; 955b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,2,corners+2*i,gc);CHKERRQ(ierr); 956b03ebc13SStefano Zampini if (gc[0] > gc[1]) { 957b03ebc13SStefano Zampini PetscInt swap = corners[2*i]; 958b03ebc13SStefano Zampini corners[2*i] = corners[2*i+1]; 959b03ebc13SStefano Zampini corners[2*i+1] = swap; 960b03ebc13SStefano Zampini } 9614e64d54eSstefano_zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 962b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"EDGE %d: ce %d, corners (%d,%d)\n",i,cedges[i],corners[2*i],corners[2*i+1]); 9634e64d54eSstefano_zampini } 9640569b399SStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 9654e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvc);CHKERRQ(ierr); 9664e64d54eSstefano_zampini 967a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 968a13144ffSStefano Zampini /* Inspects columns of lG (rows of lGt) and make sure the change of basis will 969a13144ffSStefano Zampini not interfere with neighbouring coarse edges */ 970a13144ffSStefano Zampini ierr = PetscMalloc1(nee+1,&emarks);CHKERRQ(ierr); 971a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 972a13144ffSStefano Zampini for (i=0;i<nv;i++) { 973a13144ffSStefano Zampini PetscInt emax = 0,eemax = 0; 974a13144ffSStefano Zampini 975a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 976a13144ffSStefano Zampini ierr = PetscMemzero(emarks,(nee+1)*sizeof(PetscInt));CHKERRQ(ierr); 977a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++; 978a13144ffSStefano Zampini for (j=1;j<nee+1;j++) { 979a13144ffSStefano Zampini if (emax < emarks[j]) { 980a13144ffSStefano Zampini emax = emarks[j]; 981a13144ffSStefano Zampini eemax = j; 982a13144ffSStefano Zampini } 983a13144ffSStefano Zampini } 984a13144ffSStefano Zampini /* not relevant for edges */ 985a13144ffSStefano Zampini if (!eemax) continue; 986a13144ffSStefano Zampini 987a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 988a13144ffSStefano Zampini if (marks[jj[j]] && marks[jj[j]] != eemax) { 989c2151214SStefano 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]); 990a13144ffSStefano Zampini } 991a13144ffSStefano Zampini } 992a13144ffSStefano Zampini } 993a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 994a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 995a13144ffSStefano Zampini #endif 996a13144ffSStefano Zampini 997a13144ffSStefano Zampini /* Compute extended rows indices for edge blocks of the change of basis */ 998a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 999a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);CHKERRQ(ierr); 1000a13144ffSStefano Zampini extmem *= maxsize; 1001a13144ffSStefano Zampini ierr = PetscMalloc1(extmem*nee,&extrow);CHKERRQ(ierr); 1002a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extrows);CHKERRQ(ierr); 1003a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&extrowcum);CHKERRQ(ierr); 1004a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1005a13144ffSStefano Zampini PetscInt mark = 0,size,start; 1006a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1007a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) 1008a13144ffSStefano Zampini if (marks[jj[j]] && !mark) 1009a13144ffSStefano Zampini mark = marks[jj[j]]; 1010a13144ffSStefano Zampini 1011a13144ffSStefano Zampini /* not relevant */ 1012a13144ffSStefano Zampini if (!mark) continue; 1013a13144ffSStefano Zampini 1014a13144ffSStefano Zampini /* import extended row */ 1015a13144ffSStefano Zampini mark--; 1016a13144ffSStefano Zampini start = mark*extmem+extrowcum[mark]; 1017a13144ffSStefano Zampini size = ii[i+1]-ii[i]; 1018a13144ffSStefano Zampini if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %d > %d",extrowcum[mark] + size,extmem); 1019a13144ffSStefano Zampini ierr = PetscMemcpy(extrow+start,jj+ii[i],size*sizeof(PetscInt));CHKERRQ(ierr); 1020a13144ffSStefano Zampini extrowcum[mark] += size; 1021a13144ffSStefano Zampini } 1022a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1023a13144ffSStefano Zampini cum = 0; 1024a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1025a13144ffSStefano Zampini PetscInt size = extrowcum[i],*start = extrow + i*extmem; 1026a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&size,start);CHKERRQ(ierr); 1027a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);CHKERRQ(ierr); 1028a13144ffSStefano Zampini cum = PetscMax(cum,size); 1029a13144ffSStefano Zampini } 1030a13144ffSStefano Zampini ierr = PetscFree(extrowcum);CHKERRQ(ierr); 1031a13144ffSStefano Zampini ierr = PetscFree(marks);CHKERRQ(ierr); 1032a13144ffSStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 1033a13144ffSStefano Zampini ierr = PetscBTDestroy(&btvcand);CHKERRQ(ierr); 1034a13144ffSStefano Zampini 1035a13144ffSStefano Zampini /* Workspace for lapack inner calls and VecSetValues */ 1036a13144ffSStefano Zampini ierr = PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);CHKERRQ(ierr); 1037a13144ffSStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 1038a13144ffSStefano Zampini for (i=0;i<maxsize;i++) vals[i] = 1.; 1039a13144ffSStefano Zampini 1040a13144ffSStefano Zampini /* Create vectors for quadrature rules */ 1041c2151214SStefano Zampini /* TODO preserve other quadratures */ 1042a13144ffSStefano Zampini ierr = PetscMalloc1(nquads,&quads);CHKERRQ(ierr); 1043a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1044a13144ffSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quads[i],NULL);CHKERRQ(ierr); 1045a13144ffSStefano Zampini ierr = VecSetLocalToGlobalMapping(quads[i],el2g);CHKERRQ(ierr); 1046a13144ffSStefano Zampini } 1047a13144ffSStefano Zampini ierr = PCBDDCNullSpaceCreate(comm,PETSC_FALSE,nquads,quads,&nnsp);CHKERRQ(ierr); 1048a13144ffSStefano Zampini 1049a13144ffSStefano Zampini /* Create change of basis matrix (preallocation can be improved) */ 1050a13144ffSStefano Zampini ierr = MatCreate(comm,&T);CHKERRQ(ierr); 1051c2151214SStefano Zampini ierr = MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n, 1052c2151214SStefano Zampini pc->pmat->rmap->N,pc->pmat->rmap->N);CHKERRQ(ierr); 1053a13144ffSStefano Zampini ierr = MatSetType(T,MATAIJ);CHKERRQ(ierr); 1054a13144ffSStefano Zampini ierr = MatSeqAIJSetPreallocation(T,10,NULL);CHKERRQ(ierr); 1055a13144ffSStefano Zampini ierr = MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);CHKERRQ(ierr); 1056a13144ffSStefano Zampini ierr = MatSetLocalToGlobalMapping(T,el2g,el2g);CHKERRQ(ierr); 1057a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1058a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 1059a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_IGNORE_ZERO_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 1060a13144ffSStefano Zampini 1061a13144ffSStefano Zampini /* Defaults to identity */ 1062c2151214SStefano Zampini ierr = MatCreateVecs(pc->pmat,&tvec,NULL);CHKERRQ(ierr); 1063a13144ffSStefano Zampini ierr = VecSet(tvec,1.0);CHKERRQ(ierr); 1064a13144ffSStefano Zampini ierr = MatDiagonalSet(T,tvec,INSERT_VALUES);CHKERRQ(ierr); 1065a13144ffSStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 1066a13144ffSStefano Zampini 1067a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1068a13144ffSStefano Zampini Mat Gins = NULL, GKins = NULL; 1069a13144ffSStefano Zampini 1070a13144ffSStefano Zampini ierr = PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],&Gins,&GKins,work,rwork);CHKERRQ(ierr); 1071a13144ffSStefano Zampini if (Gins && GKins) { 1072a13144ffSStefano Zampini PetscScalar *data; 1073a13144ffSStefano Zampini const PetscInt *rows,*cols; 1074a13144ffSStefano Zampini PetscInt nrh,nch,nrc,ncc; 1075a13144ffSStefano Zampini 1076a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&cols);CHKERRQ(ierr); 1077a13144ffSStefano Zampini /* H1 */ 1078a13144ffSStefano Zampini ierr = ISGetIndices(extrows[i],&rows);CHKERRQ(ierr); 1079a13144ffSStefano Zampini ierr = MatGetSize(Gins,&nrh,&nch);CHKERRQ(ierr); 1080a13144ffSStefano Zampini ierr = MatDenseGetArray(Gins,&data);CHKERRQ(ierr); 1081a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);CHKERRQ(ierr); 1082a13144ffSStefano Zampini ierr = MatDenseRestoreArray(Gins,&data);CHKERRQ(ierr); 1083a13144ffSStefano Zampini ierr = ISRestoreIndices(extrows[i],&rows);CHKERRQ(ierr); 1084a13144ffSStefano Zampini /* complement */ 1085a13144ffSStefano Zampini ierr = MatGetSize(GKins,&nrc,&ncc);CHKERRQ(ierr); 1086a13144ffSStefano Zampini if (ncc > nquads-1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet supported ncc %d nquads %d",ncc,nquads); 1087a13144ffSStefano 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); 1088a13144ffSStefano Zampini ierr = MatDenseGetArray(GKins,&data);CHKERRQ(ierr); 1089a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);CHKERRQ(ierr); 1090a13144ffSStefano Zampini ierr = MatDenseRestoreArray(GKins,&data);CHKERRQ(ierr); 1091a13144ffSStefano Zampini /* Gins kernel quadratures */ 1092a13144ffSStefano Zampini for (j=0;j<ncc;j++) { 1093a13144ffSStefano Zampini ierr = VecSetValueLocal(quads[j],cols[nch+j],1.,INSERT_VALUES);CHKERRQ(ierr); 1094a13144ffSStefano Zampini } 1095a13144ffSStefano Zampini /* H1 average */ 1096a13144ffSStefano Zampini ierr = VecSetValuesLocal(quads[nquads-1],nch,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 1097a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&cols);CHKERRQ(ierr); 1098a13144ffSStefano Zampini } 1099a13144ffSStefano Zampini ierr = ISDestroy(&extrows[i]);CHKERRQ(ierr); 1100a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 1101a13144ffSStefano Zampini ierr = MatDestroy(&Gins);CHKERRQ(ierr); 1102a13144ffSStefano Zampini ierr = MatDestroy(&GKins);CHKERRQ(ierr); 1103a13144ffSStefano Zampini } 1104a13144ffSStefano Zampini 1105a13144ffSStefano Zampini /* Start assembling */ 1106a13144ffSStefano Zampini ierr = MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1107a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1108a13144ffSStefano Zampini ierr = VecAssemblyBegin(quads[i]);CHKERRQ(ierr); 1109a13144ffSStefano Zampini } 1110a13144ffSStefano Zampini 1111a13144ffSStefano Zampini /* Free */ 1112c2151214SStefano Zampini if (fl2g) { 1113c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1114c2151214SStefano Zampini for (i=0;i<nee;i++) { 1115c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1116c2151214SStefano Zampini } 1117c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1118c2151214SStefano Zampini } 1119c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1120c2151214SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1121c2151214SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 1122a13144ffSStefano Zampini ierr = PetscFree(extrow);CHKERRQ(ierr); 1123a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&vl2g);CHKERRQ(ierr); 1124c2151214SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&el2g);CHKERRQ(ierr); 1125c2151214SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&fl2g);CHKERRQ(ierr); 1126a13144ffSStefano Zampini ierr = PetscFree2(work,rwork);CHKERRQ(ierr); 1127a13144ffSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1128b03ebc13SStefano Zampini ierr = PetscFree(corners);CHKERRQ(ierr); 1129b03ebc13SStefano Zampini ierr = PetscFree(cedges);CHKERRQ(ierr); 1130a13144ffSStefano Zampini ierr = PetscFree(extrows);CHKERRQ(ierr); 1131a13144ffSStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 1132a13144ffSStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 1133a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 1134a13144ffSStefano Zampini ierr = MatDestroy(&conn);CHKERRQ(ierr); 1135a13144ffSStefano Zampini 1136a13144ffSStefano Zampini /* Complete assembling */ 1137a13144ffSStefano Zampini ierr = MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1138a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1139a13144ffSStefano Zampini ierr = VecAssemblyEnd(quads[i]);CHKERRQ(ierr); 1140a13144ffSStefano Zampini } 1141a13144ffSStefano Zampini for (i=0;i<nquads;i++) { 1142a13144ffSStefano Zampini ierr = VecDestroy(&quads[i]);CHKERRQ(ierr); 1143a13144ffSStefano Zampini } 1144a13144ffSStefano Zampini ierr = PetscFree(quads);CHKERRQ(ierr); 1145a13144ffSStefano Zampini 1146a13144ffSStefano Zampini /* set change of basis */ 1147a13144ffSStefano Zampini ierr = PCBDDCSetChangeOfBasisMat(pc,T,PETSC_FALSE);CHKERRQ(ierr); 1148a13144ffSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1149a13144ffSStefano Zampini 1150a13144ffSStefano Zampini /* set quadratures */ 1151a13144ffSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,nnsp);CHKERRQ(ierr); 1152a13144ffSStefano Zampini ierr = MatNullSpaceDestroy(&nnsp);CHKERRQ(ierr); 1153a13144ffSStefano Zampini 1154a13144ffSStefano Zampini PetscFunctionReturn(0); 1155a13144ffSStefano Zampini } 1156a13144ffSStefano Zampini 1157d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 1158d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 1159d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 1160d8203eabSStefano Zampini #undef __FUNCT__ 1161d8203eabSStefano Zampini #define __FUNCT__ "PCBDDCNullSpaceCreate" 1162d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 1163d8203eabSStefano Zampini { 1164d8203eabSStefano Zampini PetscErrorCode ierr; 1165d8203eabSStefano Zampini PetscInt i; 1166d8203eabSStefano Zampini 1167d8203eabSStefano Zampini PetscFunctionBegin; 1168d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 1169d8203eabSStefano Zampini PetscInt first,last; 1170d8203eabSStefano Zampini 1171d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 117286fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 1173d8203eabSStefano Zampini if (i>=first && i < last) { 1174d8203eabSStefano Zampini PetscScalar *data; 1175d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1176d8203eabSStefano Zampini if (!has_const) { 1177d8203eabSStefano Zampini data[i-first] = 1.; 1178d8203eabSStefano Zampini } else { 117986fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 118086fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 1181d8203eabSStefano Zampini } 1182d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1183d8203eabSStefano Zampini } 1184d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1185d8203eabSStefano Zampini } 1186d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 1187d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 1188d8203eabSStefano Zampini PetscInt first,last; 1189d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 1190d8203eabSStefano Zampini if (i>=first && i < last) { 1191d8203eabSStefano Zampini PetscScalar *data; 1192d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1193d8203eabSStefano Zampini if (!has_const) { 1194d8203eabSStefano Zampini data[i-first] = 0.; 1195d8203eabSStefano Zampini } else { 119686fa73c5SStefano Zampini data[2*i-first] = 0.; 119786fa73c5SStefano Zampini data[2*i-first+1] = 0.; 1198d8203eabSStefano Zampini } 1199d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1200d8203eabSStefano Zampini } 1201d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1202d8203eabSStefano Zampini } 1203d8203eabSStefano Zampini PetscFunctionReturn(0); 1204d8203eabSStefano Zampini } 1205d8203eabSStefano Zampini 1206669cc0f4SStefano Zampini #undef __FUNCT__ 1207669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 12088ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 1209669cc0f4SStefano Zampini { 1210a198735bSStefano Zampini Mat loc_divudotp; 1211fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 12128ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 1213669cc0f4SStefano Zampini IS *faces,*edges; 1214669cc0f4SStefano Zampini PetscScalar *vals; 1215669cc0f4SStefano Zampini const PetscScalar *array; 1216669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 12171ae86dd6SStefano Zampini PetscMPIInt rank; 1218a198735bSStefano Zampini PetscErrorCode ierr; 1219669cc0f4SStefano Zampini 1220669cc0f4SStefano Zampini PetscFunctionBegin; 1221669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1222669cc0f4SStefano Zampini if (graph->twodim) { 1223669cc0f4SStefano Zampini lmaxneighs = 2; 1224669cc0f4SStefano Zampini } else { 1225669cc0f4SStefano Zampini lmaxneighs = 1; 1226669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1227669cc0f4SStefano Zampini const PetscInt *idxs; 1228669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1229669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 1230669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1231669cc0f4SStefano Zampini } 1232669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 1233669cc0f4SStefano Zampini } 1234669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1235669cc0f4SStefano Zampini maxsize = 0; 1236669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1237669cc0f4SStefano Zampini PetscInt nn; 1238669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1239669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1240669cc0f4SStefano Zampini } 1241669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1242669cc0f4SStefano Zampini PetscInt nn; 1243669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1244669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1245669cc0f4SStefano Zampini } 1246669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 1247669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 1248669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 12498ae0ca82SStefano Zampini if (!transpose) { 12508ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 12518ae0ca82SStefano Zampini } else { 12528ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 12538ae0ca82SStefano Zampini } 1254669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 12551ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 1256d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 1257669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 12588ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 1259669cc0f4SStefano Zampini } 1260d8203eabSStefano Zampini 1261669cc0f4SStefano Zampini /* compute local quad vec */ 1262a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 12638ae0ca82SStefano Zampini if (!transpose) { 1264a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 12658ae0ca82SStefano Zampini } else { 12668ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 12678ae0ca82SStefano Zampini } 1268669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 12698ae0ca82SStefano Zampini if (!transpose) { 1270a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 12718ae0ca82SStefano Zampini } else { 12728ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 12738ae0ca82SStefano Zampini } 1274fa23a32eSStefano Zampini if (vl2l) { 1275fa23a32eSStefano Zampini ierr = VecGetSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1276fa23a32eSStefano Zampini } else { 1277fa23a32eSStefano Zampini vins = v; 1278fa23a32eSStefano Zampini } 1279fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 1280669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 12819a962809SStefano Zampini 12821ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 12831ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 1284669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1285669cc0f4SStefano Zampini const PetscInt *idxs; 1286669cc0f4SStefano Zampini PetscInt idx,nn,j; 1287669cc0f4SStefano Zampini 1288669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 1289669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1290669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 12911ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1292669cc0f4SStefano Zampini idx = -(idx+1); 1293669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1294669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 1295669cc0f4SStefano Zampini } 1296669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1297669cc0f4SStefano Zampini const PetscInt *idxs; 1298669cc0f4SStefano Zampini PetscInt idx,nn,j; 1299669cc0f4SStefano Zampini 1300669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1301669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1302669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 13031ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1304669cc0f4SStefano Zampini idx = -(idx+1); 1305669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1306669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1307669cc0f4SStefano Zampini } 1308c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1309fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 1310fa23a32eSStefano Zampini if (vl2l) { 1311fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1312fa23a32eSStefano Zampini } 1313669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1314669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1315669cc0f4SStefano Zampini 1316669cc0f4SStefano Zampini /* assemble near null space */ 1317669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1318669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 1319669cc0f4SStefano Zampini } 1320669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1321669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 1322669cc0f4SStefano Zampini } 1323669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 1324669cc0f4SStefano Zampini PetscFunctionReturn(0); 1325669cc0f4SStefano Zampini } 1326669cc0f4SStefano Zampini 1327669cc0f4SStefano Zampini 1328a3df083aSStefano Zampini #undef __FUNCT__ 13291f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 13301f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 13311f4df5f7SStefano Zampini { 13321f4df5f7SStefano Zampini PetscErrorCode ierr; 13331f4df5f7SStefano Zampini Vec local,global; 13341f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13351f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 13361f4df5f7SStefano Zampini 13371f4df5f7SStefano Zampini PetscFunctionBegin; 13381f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 13391f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 13401f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 13411f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 13421f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 13431f4df5f7SStefano Zampini PetscInt i; 13441f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 13451f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 13461f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 13471f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 13481f4df5f7SStefano Zampini } 13491f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 13501f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 13511f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 13521f4df5f7SStefano Zampini } 13531f4df5f7SStefano Zampini } else { 1354986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 13551f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1356986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 1357986cdee1SStefano Zampini if (i > 1) { 1358986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 13591f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 13601f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 13611f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 13621f4df5f7SStefano Zampini } 13631f4df5f7SStefano Zampini } 13641f4df5f7SStefano Zampini } 1365986cdee1SStefano Zampini } 13661f4df5f7SStefano Zampini 13671f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 13681f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 13691f4df5f7SStefano Zampini } 13701f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 13711f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 13721f4df5f7SStefano Zampini } 13731f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 13741f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 13751f4df5f7SStefano Zampini } 13761f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 13771f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 13781f4df5f7SStefano Zampini PetscFunctionReturn(0); 13791f4df5f7SStefano Zampini } 13801f4df5f7SStefano Zampini 13811f4df5f7SStefano Zampini #undef __FUNCT__ 13823e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 13833e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 13843e589ea0SStefano Zampini { 13853e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 13863e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 13873e589ea0SStefano Zampini PetscErrorCode ierr; 13883e589ea0SStefano Zampini 13893e589ea0SStefano Zampini PetscFunctionBegin; 13903e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 13913e589ea0SStefano Zampini PetscFunctionReturn(0); 13923e589ea0SStefano Zampini } 13933e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 13943e589ea0SStefano Zampini Vec swap; 13953e589ea0SStefano Zampini 13963e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 13973e589ea0SStefano Zampini swap = pcbddc->work_change; 13983e589ea0SStefano Zampini pcbddc->work_change = r; 13993e589ea0SStefano Zampini r = swap; 14003e589ea0SStefano Zampini } 14013e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 14023e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 14033e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 14043e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 14053e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 14063e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 14073e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1408f913dca9SStefano Zampini pcbddc->work_change = r; 14093e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 14103e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 14113e589ea0SStefano Zampini } 14123e589ea0SStefano Zampini PetscFunctionReturn(0); 14133e589ea0SStefano Zampini } 14143e589ea0SStefano Zampini 14153e589ea0SStefano Zampini #undef __FUNCT__ 1416a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 1417a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1418a3df083aSStefano Zampini { 1419a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1420a3df083aSStefano Zampini PetscErrorCode ierr; 1421a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1422a3df083aSStefano Zampini 1423a3df083aSStefano Zampini PetscFunctionBegin; 1424a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1425a3df083aSStefano Zampini if (transpose) { 1426a3df083aSStefano Zampini apply_right = ctx->apply_left; 1427a3df083aSStefano Zampini apply_left = ctx->apply_right; 1428a3df083aSStefano Zampini } else { 1429a3df083aSStefano Zampini apply_right = ctx->apply_right; 1430a3df083aSStefano Zampini apply_left = ctx->apply_left; 1431a3df083aSStefano Zampini } 1432a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1433a3df083aSStefano Zampini if (apply_right) { 1434a3df083aSStefano Zampini const PetscScalar *ax; 1435a3df083aSStefano Zampini PetscInt nl,i; 1436a3df083aSStefano Zampini 1437a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1438a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1439a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 1440a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1441a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1442a3df083aSStefano Zampini PetscScalar sum,val; 1443a3df083aSStefano Zampini const PetscInt *idxs; 1444a3df083aSStefano Zampini PetscInt nz,j; 1445a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1446a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1447a3df083aSStefano Zampini sum = 0.; 1448a3df083aSStefano Zampini if (ctx->apply_p0) { 1449a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1450a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1451a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1452a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1453a3df083aSStefano Zampini } 1454a3df083aSStefano Zampini } else { 1455a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1456a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1457a3df083aSStefano Zampini } 1458a3df083aSStefano Zampini } 1459a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1460a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1461a3df083aSStefano Zampini } 1462a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1463a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1464a3df083aSStefano Zampini } 1465a3df083aSStefano Zampini if (transpose) { 1466a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1467a3df083aSStefano Zampini } else { 1468a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1469a3df083aSStefano Zampini } 1470a3df083aSStefano Zampini if (reset_x) { 1471a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1472a3df083aSStefano Zampini } 1473a3df083aSStefano Zampini if (apply_left) { 1474a3df083aSStefano Zampini PetscScalar *ay; 1475a3df083aSStefano Zampini PetscInt i; 1476a3df083aSStefano Zampini 1477a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1478a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1479a3df083aSStefano Zampini PetscScalar sum,val; 1480a3df083aSStefano Zampini const PetscInt *idxs; 1481a3df083aSStefano Zampini PetscInt nz,j; 1482a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1483a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1484a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 1485a3df083aSStefano Zampini if (ctx->apply_p0) { 1486a3df083aSStefano Zampini sum = 0.; 1487a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1488a3df083aSStefano Zampini sum += ay[idxs[j]]; 1489a3df083aSStefano Zampini ay[idxs[j]] += val; 1490a3df083aSStefano Zampini } 1491a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 1492a3df083aSStefano Zampini } else { 1493a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1494a3df083aSStefano Zampini ay[idxs[j]] += val; 1495a3df083aSStefano Zampini } 1496a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 1497a3df083aSStefano Zampini } 1498a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1499a3df083aSStefano Zampini } 1500a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 1501a3df083aSStefano Zampini } 1502a3df083aSStefano Zampini PetscFunctionReturn(0); 1503a3df083aSStefano Zampini } 1504a3df083aSStefano Zampini 1505a3df083aSStefano Zampini #undef __FUNCT__ 1506a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 1507a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 1508a3df083aSStefano Zampini { 1509a3df083aSStefano Zampini PetscErrorCode ierr; 1510a3df083aSStefano Zampini 1511a3df083aSStefano Zampini PetscFunctionBegin; 1512a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 1513a3df083aSStefano Zampini PetscFunctionReturn(0); 1514a3df083aSStefano Zampini } 1515a3df083aSStefano Zampini 1516a3df083aSStefano Zampini #undef __FUNCT__ 1517a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 1518a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 1519a3df083aSStefano Zampini { 1520a3df083aSStefano Zampini PetscErrorCode ierr; 1521a3df083aSStefano Zampini 1522a3df083aSStefano Zampini PetscFunctionBegin; 1523a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 1524a3df083aSStefano Zampini PetscFunctionReturn(0); 1525a3df083aSStefano Zampini } 1526a3df083aSStefano Zampini 1527a3df083aSStefano Zampini #undef __FUNCT__ 1528a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 1529a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 1530a3df083aSStefano Zampini { 1531a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1532a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1533a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1534a3df083aSStefano Zampini PetscErrorCode ierr; 1535a3df083aSStefano Zampini 1536a3df083aSStefano Zampini PetscFunctionBegin; 1537a3df083aSStefano Zampini if (!restore) { 15381dd7afcfSStefano Zampini Mat A_IB,A_BI; 1539a3df083aSStefano Zampini PetscScalar *work; 1540b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 1541a3df083aSStefano Zampini 15429a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 15439a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 1544a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 1545a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 1546a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1547a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 1548a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 1549a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 1550a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 1551a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 1552a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 1553a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 1554a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 1555a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 1556059032f7SStefano Zampini if (reuse) { 1557a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 15581dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 1559059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 1560059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 1561059032f7SStefano Zampini PetscInt i; 1562059032f7SStefano Zampini 1563059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 1564059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1565059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1566059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1567059032f7SStefano Zampini } 1568059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 15691dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 1570059032f7SStefano Zampini } 1571a3df083aSStefano Zampini ctx->A = pcis->A_IB; 1572a3df083aSStefano Zampini ctx->work = work; 1573a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 1574a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1575a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1576a3df083aSStefano Zampini pcis->A_IB = A_IB; 1577a3df083aSStefano Zampini 1578a3df083aSStefano Zampini /* A_BI as A_IB^T */ 1579a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 1580a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 1581a3df083aSStefano Zampini pcis->A_BI = A_BI; 1582a3df083aSStefano Zampini } else { 15831dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 15841dd7afcfSStefano Zampini PetscFunctionReturn(0); 15851dd7afcfSStefano Zampini } 1586a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 1587a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 1588a3df083aSStefano Zampini pcis->A_IB = ctx->A; 15891dd7afcfSStefano Zampini ctx->A = NULL; 15901dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 15911dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 15921dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 15931dd7afcfSStefano Zampini if (ctx->free) { 1594059032f7SStefano Zampini PetscInt i; 15951dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1596059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1597059032f7SStefano Zampini } 1598059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1599059032f7SStefano Zampini } 1600a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 1601a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 1602a3df083aSStefano Zampini } 1603a3df083aSStefano Zampini PetscFunctionReturn(0); 1604a3df083aSStefano Zampini } 1605a3df083aSStefano Zampini 1606a3df083aSStefano Zampini /* used just in bddc debug mode */ 1607a3df083aSStefano Zampini #undef __FUNCT__ 1608a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 1609a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 1610a3df083aSStefano Zampini { 1611a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1612a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1613a3df083aSStefano Zampini Mat An; 1614a3df083aSStefano Zampini PetscErrorCode ierr; 1615a3df083aSStefano Zampini 1616a3df083aSStefano Zampini PetscFunctionBegin; 1617a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 1618a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 1619a3df083aSStefano Zampini if (is1) { 1620a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 1621a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 1622a3df083aSStefano Zampini } else { 1623a3df083aSStefano Zampini *B = An; 1624a3df083aSStefano Zampini } 1625a3df083aSStefano Zampini PetscFunctionReturn(0); 1626a3df083aSStefano Zampini } 1627a3df083aSStefano Zampini 16281cf9b237SStefano Zampini /* TODO: add reuse flag */ 16291cf9b237SStefano Zampini #undef __FUNCT__ 16301cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 16311cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 16321cf9b237SStefano Zampini { 16331cf9b237SStefano Zampini Mat Bt; 16341cf9b237SStefano Zampini PetscScalar *a,*bdata; 16351cf9b237SStefano Zampini const PetscInt *ii,*ij; 16361cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 16371cf9b237SStefano Zampini PetscBool flg_row; 16381cf9b237SStefano Zampini PetscErrorCode ierr; 16391cf9b237SStefano Zampini 16401cf9b237SStefano Zampini PetscFunctionBegin; 16411cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 16421cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 16431cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 16441cf9b237SStefano Zampini nnz = n; 16451cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 16461cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 16471cf9b237SStefano Zampini } 16481cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 16491cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 16501cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 16511cf9b237SStefano Zampini nnz = 0; 16521cf9b237SStefano Zampini bii[0] = 0; 16531cf9b237SStefano Zampini for (i=0;i<n;i++) { 16541cf9b237SStefano Zampini PetscInt j; 16551cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 16561cf9b237SStefano Zampini PetscScalar entry = a[j]; 16571cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 16581cf9b237SStefano Zampini bij[nnz] = ij[j]; 16591cf9b237SStefano Zampini bdata[nnz] = entry; 16601cf9b237SStefano Zampini nnz++; 16611cf9b237SStefano Zampini } 16621cf9b237SStefano Zampini } 16631cf9b237SStefano Zampini bii[i+1] = nnz; 16641cf9b237SStefano Zampini } 16651cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 16661cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 16671cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 16681cf9b237SStefano Zampini { 16691cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 16701cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 16711cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 16721cf9b237SStefano Zampini } 16731cf9b237SStefano Zampini *B = Bt; 16741cf9b237SStefano Zampini PetscFunctionReturn(0); 16751cf9b237SStefano Zampini } 16761cf9b237SStefano Zampini 1677674ae819SStefano Zampini #undef __FUNCT__ 16784f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 16794f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 16804f1b2e48SStefano Zampini { 16814f1b2e48SStefano Zampini Mat B; 16824f1b2e48SStefano Zampini IS is_dummy,*cc_n; 16834f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 16844f1b2e48SStefano Zampini PCBDDCGraph graph; 16854f1b2e48SStefano Zampini PetscInt i,n; 16864f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 16874f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 16884f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 16894f1b2e48SStefano Zampini PetscErrorCode ierr; 16904f1b2e48SStefano Zampini 16914f1b2e48SStefano Zampini PetscFunctionBegin; 169263c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 169363c961adSStefano Zampini *ncc = 0; 169463c961adSStefano Zampini *cc = NULL; 169563c961adSStefano Zampini PetscFunctionReturn(0); 169663c961adSStefano Zampini } 16974f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 16984f1b2e48SStefano Zampini if (!isseqaij && filter) { 16991cf9b237SStefano Zampini PetscBool isseqdense; 17001cf9b237SStefano Zampini 17011cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 17021cf9b237SStefano Zampini if (!isseqdense) { 17034f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 17041cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 17051cf9b237SStefano Zampini PetscScalar *array; 17061cf9b237SStefano Zampini PetscReal chop=1.e-6; 17071cf9b237SStefano Zampini 17081cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 17091cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 17101cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 17111cf9b237SStefano Zampini for (i=0;i<n;i++) { 17121cf9b237SStefano Zampini PetscInt j; 17131cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 17141cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 17151cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 17161cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 17171cf9b237SStefano Zampini } 17181cf9b237SStefano Zampini } 17191cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 17209d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 17211cf9b237SStefano Zampini } 17224f1b2e48SStefano Zampini } else { 17234f1b2e48SStefano Zampini B = A; 17244f1b2e48SStefano Zampini } 17254f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 17264f1b2e48SStefano Zampini 17274f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 17284f1b2e48SStefano Zampini if (filter) { 17294f1b2e48SStefano Zampini PetscScalar *data; 17304f1b2e48SStefano Zampini PetscInt j,cum; 17314f1b2e48SStefano Zampini 17324f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 17334f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 17344f1b2e48SStefano Zampini cum = 0; 17354f1b2e48SStefano Zampini for (i=0;i<n;i++) { 17364f1b2e48SStefano Zampini PetscInt t; 17374f1b2e48SStefano Zampini 17384f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 17394f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 17404f1b2e48SStefano Zampini continue; 17414f1b2e48SStefano Zampini } 17424f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 17434f1b2e48SStefano Zampini } 17444f1b2e48SStefano Zampini t = xadj_filtered[i]; 17454f1b2e48SStefano Zampini xadj_filtered[i] = cum; 17464f1b2e48SStefano Zampini cum += t; 17474f1b2e48SStefano Zampini } 17484f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 17494f1b2e48SStefano Zampini } else { 17504f1b2e48SStefano Zampini xadj_filtered = NULL; 17514f1b2e48SStefano Zampini adjncy_filtered = NULL; 17524f1b2e48SStefano Zampini } 17534f1b2e48SStefano Zampini 17544f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 17554f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 17564f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 17574f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 17584f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 1759be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 17604f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 17614f1b2e48SStefano Zampini if (xadj_filtered) { 17624f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 17634f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 17644f1b2e48SStefano Zampini } else { 17654f1b2e48SStefano Zampini graph->xadj = xadj; 17664f1b2e48SStefano Zampini graph->adjncy = adjncy; 17674f1b2e48SStefano Zampini } 17684f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 17694f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 17704f1b2e48SStefano Zampini /* partial clean up */ 17714f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 17724f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 17731cf9b237SStefano Zampini if (A != B) { 17744f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 17754f1b2e48SStefano Zampini } 17764f1b2e48SStefano Zampini 17774f1b2e48SStefano Zampini /* get back data */ 17781cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 17791cf9b237SStefano Zampini if (cc) { 17804f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 17814f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 17824f1b2e48SStefano 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); 17834f1b2e48SStefano Zampini } 17844f1b2e48SStefano Zampini *cc = cc_n; 17851cf9b237SStefano Zampini } 17864f1b2e48SStefano Zampini /* clean up graph */ 17874f1b2e48SStefano Zampini graph->xadj = 0; 17884f1b2e48SStefano Zampini graph->adjncy = 0; 17894f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 17904f1b2e48SStefano Zampini PetscFunctionReturn(0); 17914f1b2e48SStefano Zampini } 17924f1b2e48SStefano Zampini 17934f1b2e48SStefano Zampini #undef __FUNCT__ 17945408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 17955408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 17965408967cSStefano Zampini { 17975408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 17985408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1799dee84bffSStefano Zampini IS dirIS = NULL; 18004f1b2e48SStefano Zampini PetscInt i; 18015408967cSStefano Zampini PetscErrorCode ierr; 18025408967cSStefano Zampini 18035408967cSStefano Zampini PetscFunctionBegin; 1804dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 18055408967cSStefano Zampini if (zerodiag) { 18065408967cSStefano Zampini Mat A; 18075408967cSStefano Zampini Vec vec3_N; 18085408967cSStefano Zampini PetscScalar *vals; 18095408967cSStefano Zampini const PetscInt *idxs; 1810d12d3064SStefano Zampini PetscInt nz,*count; 18115408967cSStefano Zampini 18125408967cSStefano Zampini /* p0 */ 18135408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 18145408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 18155408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 18165408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 18174f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 18185408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 18195408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 18205408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 18215408967cSStefano Zampini /* v_I */ 18225408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 18235408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 18245408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 18255408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 18265408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 18275408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 18285408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 18295408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 18305408967cSStefano Zampini if (dirIS) { 18315408967cSStefano Zampini PetscInt n; 18325408967cSStefano Zampini 18335408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 18345408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 18355408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 18365408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 18375408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 18385408967cSStefano Zampini } 18395408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 18405408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 18415408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 18425408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 1843669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 18445408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 18455408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 18469a962809SStefano 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])); 18475408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 18485408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 1849d12d3064SStefano Zampini 1850d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 1851d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 1852d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 1853d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 1854d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 1855d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 18569a962809SStefano 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]); 1857d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 1858d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 18595408967cSStefano Zampini } 1860dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 18615408967cSStefano Zampini 18625408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 18635408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 18644f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 18655408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 18664f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 18675408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 1868f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1869f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 1870f2a566d8SStefano 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); 1871f2a566d8SStefano Zampini } 18725408967cSStefano Zampini PetscFunctionReturn(0); 18735408967cSStefano Zampini } 18745408967cSStefano Zampini 18755408967cSStefano Zampini #undef __FUNCT__ 1876339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 1877339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 1878339f8db1SStefano Zampini { 1879339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 18804f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 1881b0f5fe93SStefano Zampini PetscInt nz,n; 18821f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 18834f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 1884339f8db1SStefano Zampini PetscErrorCode ierr; 1885339f8db1SStefano Zampini 1886339f8db1SStefano Zampini PetscFunctionBegin; 18879f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 18889f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 1889a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 1890a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 1891a3df083aSStefano Zampini } 1892a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1893a3df083aSStefano Zampini pcbddc->benign_n = 0; 18941ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 18954f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 18964f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 18974f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 18981ae86dd6SStefano Zampini since the local Schur complements are already SPD 18994f1b2e48SStefano Zampini */ 19004f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 19014f1b2e48SStefano Zampini have_null = PETSC_TRUE; 190240fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 19034f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 19044f1b2e48SStefano Zampini 19054f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 19064f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 19074f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 19084f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 1909ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 191040fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 191140fa8d13SStefano Zampini if (!sorted) { 191240fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 191340fa8d13SStefano Zampini } 191440fa8d13SStefano Zampini } else { 191540fa8d13SStefano Zampini pressures = NULL; 191640fa8d13SStefano Zampini } 191797d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 191897d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 191927b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 192097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 1921339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 1922339f8db1SStefano Zampini if (!sorted) { 1923339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 1924339f8db1SStefano Zampini } 1925339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 19264f1b2e48SStefano Zampini if (!nz) { 19274f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 19284f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 192940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 193040fa8d13SStefano Zampini } 19314f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 19324f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 19334f1b2e48SStefano Zampini zerodiag_subs = NULL; 19344f1b2e48SStefano Zampini pcbddc->benign_n = 0; 19351f4df5f7SStefano Zampini n_interior_dofs = 0; 19361f4df5f7SStefano Zampini interior_dofs = NULL; 19371f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 19381f4df5f7SStefano Zampini PetscInt n,i,j; 19391f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 19401f4df5f7SStefano Zampini PetscInt *iwork; 19411f4df5f7SStefano Zampini 19421f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 19431f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 19441f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 19451f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 194690648384SStefano Zampini for (i=1;i<n_neigh;i++) 19471f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 19481f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 19491f4df5f7SStefano Zampini for (i=0;i<n;i++) 19501f4df5f7SStefano Zampini if (!iwork[i]) 19511f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 19521f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 19531f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 19541f4df5f7SStefano Zampini } 19554f1b2e48SStefano Zampini if (has_null_pressures) { 19564f1b2e48SStefano Zampini IS *subs; 19571f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 19581f4df5f7SStefano Zampini const PetscInt *idxs; 19591f4df5f7SStefano Zampini PetscScalar *array; 19601f4df5f7SStefano Zampini Vec *work; 19611f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 19624f1b2e48SStefano Zampini 19634f1b2e48SStefano Zampini subs = pcbddc->local_subs; 19644f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 19651f4df5f7SStefano 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) */ 19661f4df5f7SStefano Zampini if (pcbddc->current_level) { 19671f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 19681f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 19691f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 19701f4df5f7SStefano Zampini /* work[0] = 1_p */ 19711f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 19721f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 19731f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 19741f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 19751f4df5f7SStefano Zampini /* work[0] = 1_v */ 19761f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 19771f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 19781f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 19791f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 19801f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 19811f4df5f7SStefano Zampini } 19824f1b2e48SStefano Zampini if (nsubs > 1) { 19834f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 19844f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 19854f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 19864f1b2e48SStefano Zampini IS t_zerodiag_subs; 19874f1b2e48SStefano Zampini PetscInt nl; 19884f1b2e48SStefano Zampini 19894f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 19904f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 19914f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 19924f1b2e48SStefano Zampini if (nl) { 19934f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 19944f1b2e48SStefano Zampini 19951f4df5f7SStefano Zampini if (pcbddc->current_level) { 19961f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 19971f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 19981f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 19991f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 20001f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 20011f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 20021f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 20031f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 20041f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 20051f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 20061f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 20071f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 20081f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 20091f4df5f7SStefano Zampini valid = PETSC_FALSE; 20101f4df5f7SStefano Zampini break; 20111f4df5f7SStefano Zampini } 20121f4df5f7SStefano Zampini } 20131f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 20141f4df5f7SStefano Zampini } 20151f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 20161f4df5f7SStefano Zampini IS t_bc; 20171f4df5f7SStefano Zampini PetscInt nzb; 20181f4df5f7SStefano Zampini 20191f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 20201f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 20211f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 20221f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 20231f4df5f7SStefano Zampini } 20241f4df5f7SStefano Zampini if (valid && pressures) { 20254f1b2e48SStefano Zampini IS t_pressure_subs; 20264f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 20274f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 20284f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 20294f1b2e48SStefano Zampini } 20304f1b2e48SStefano Zampini if (valid) { 20314f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 20324f1b2e48SStefano Zampini pcbddc->benign_n++; 20334f1b2e48SStefano Zampini } else { 20344f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 20354f1b2e48SStefano Zampini } 20364f1b2e48SStefano Zampini } 20374f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 20384f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 20394f1b2e48SStefano Zampini } 20404f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 20414f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 20421f4df5f7SStefano Zampini 20431f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 20441f4df5f7SStefano Zampini PetscInt nzb; 20451f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 20461f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 20471f4df5f7SStefano Zampini } 20481f4df5f7SStefano Zampini if (valid && pressures) { 20494f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 20504f1b2e48SStefano Zampini } 20511f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 20521f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 20531f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 20541f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 20551f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 20561f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 20571f4df5f7SStefano Zampini valid = PETSC_FALSE; 20581f4df5f7SStefano Zampini break; 20591f4df5f7SStefano Zampini } 20601f4df5f7SStefano Zampini } 20611f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 20621f4df5f7SStefano Zampini } 20634f1b2e48SStefano Zampini if (valid) { 20644f1b2e48SStefano Zampini pcbddc->benign_n = 1; 2065ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 20664f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 20674f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 20684f1b2e48SStefano Zampini } 20694f1b2e48SStefano Zampini } 20701f4df5f7SStefano Zampini if (pcbddc->current_level) { 20711f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 20724f1b2e48SStefano Zampini } 20731f4df5f7SStefano Zampini } 20741f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 20754f1b2e48SStefano Zampini 20764f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2077b9b0e38cSStefano Zampini PetscInt n; 2078b9b0e38cSStefano Zampini 20794f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 20804f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2081b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2082b9b0e38cSStefano Zampini if (n) { 20834f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 20844f1b2e48SStefano Zampini have_null = PETSC_FALSE; 20854f1b2e48SStefano Zampini } 2086b9b0e38cSStefano Zampini } 20874f1b2e48SStefano Zampini 20884f1b2e48SStefano Zampini /* final check for null pressures */ 20894f1b2e48SStefano Zampini if (zerodiag && pressures) { 20904f1b2e48SStefano Zampini PetscInt nz,np; 20914f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 20924f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 20934f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 20944f1b2e48SStefano Zampini } 20954f1b2e48SStefano Zampini 20964f1b2e48SStefano Zampini if (recompute_zerodiag) { 20974f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 20984f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 20994f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 21004f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 21014f1b2e48SStefano Zampini } else { 21024f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 21034f1b2e48SStefano Zampini 21044f1b2e48SStefano Zampini nzn = 0; 21054f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 21064f1b2e48SStefano Zampini PetscInt ns; 21074f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 21084f1b2e48SStefano Zampini nzn += ns; 21094f1b2e48SStefano Zampini } 21104f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 21114f1b2e48SStefano Zampini nzn = 0; 21124f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 21134f1b2e48SStefano Zampini PetscInt ns,*idxs; 21144f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 21154f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 21164f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 21174f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 21184f1b2e48SStefano Zampini nzn += ns; 21194f1b2e48SStefano Zampini } 21204f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 21214f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 21224f1b2e48SStefano Zampini } 21234f1b2e48SStefano Zampini have_null = PETSC_FALSE; 21244f1b2e48SStefano Zampini } 21254f1b2e48SStefano Zampini 2126669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2127a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2128a198735bSStefano Zampini Mat A,loc_divudotp; 2129a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2130a198735bSStefano Zampini IS row,col,isused = NULL; 2131a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2132a198735bSStefano Zampini 21331f4df5f7SStefano Zampini if (pressures) { 21341f4df5f7SStefano Zampini isused = pressures; 21351f4df5f7SStefano Zampini } else { 21361f4df5f7SStefano Zampini isused = zerodiag; 21371f4df5f7SStefano Zampini } 2138a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2139669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 21401ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 21411ae86dd6SStefano 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"); 2142a198735bSStefano Zampini n_isused = 0; 2143a198735bSStefano Zampini if (isused) { 2144a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2145a198735bSStefano Zampini } 2146a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2147a198735bSStefano Zampini st = st-n_isused; 21481ae86dd6SStefano Zampini if (n) { 2149a198735bSStefano Zampini const PetscInt *gidxs; 2150a198735bSStefano Zampini 2151a198735bSStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2152a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2153a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2154a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2155a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2156a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 21571ae86dd6SStefano Zampini } else { 2158a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2159a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2160a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2161a198735bSStefano Zampini } 2162a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2163a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2164a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2165a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2166a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2167a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2168a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2169a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2170a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2171a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2172a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2173a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2174a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2175a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 21761ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 21771ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 21781ae86dd6SStefano Zampini } 2179b3afcdbeSStefano Zampini 2180b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 21814f1b2e48SStefano Zampini if (has_null_pressures) { 21824f1b2e48SStefano Zampini IS zerodiagc; 21834f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 21844f1b2e48SStefano Zampini PetscInt i,s,*nnz; 21854f1b2e48SStefano Zampini 21864f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2187339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 2188339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 2189339f8db1SStefano Zampini /* local change of basis for pressures */ 2190339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 219197d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2192339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2193339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2194339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 21954f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 21964f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 21974f1b2e48SStefano Zampini PetscInt nzs,j; 21984f1b2e48SStefano Zampini 21994f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 22004f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 22014f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 22024f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 22034f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 22044f1b2e48SStefano Zampini } 2205339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2206339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2207339f8db1SStefano Zampini /* set identity on velocities */ 2208339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 2209339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 2210339f8db1SStefano Zampini } 22114f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 22124f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 22139f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 22144f1b2e48SStefano 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); 2215339f8db1SStefano Zampini /* set change on pressures */ 22164f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 22174f1b2e48SStefano Zampini PetscScalar *array; 22184f1b2e48SStefano Zampini PetscInt nzs; 22194f1b2e48SStefano Zampini 22204f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 22214f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 22224f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2223339f8db1SStefano Zampini PetscScalar vals[2]; 2224339f8db1SStefano Zampini PetscInt cols[2]; 2225339f8db1SStefano Zampini 2226339f8db1SStefano Zampini cols[0] = idxs[i]; 22274f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2228339f8db1SStefano Zampini vals[0] = 1.; 2229b0f5fe93SStefano Zampini vals[1] = 1.; 22304f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2231339f8db1SStefano Zampini } 22324f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 22334f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 22344f1b2e48SStefano Zampini array[nzs-1] = 1.; 22354f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 22364f1b2e48SStefano Zampini /* store local idxs for p0 */ 22374f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 22384f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2239339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 22404f1b2e48SStefano Zampini } 2241339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2242339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2243a3df083aSStefano Zampini /* project if needed */ 2244a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 22451dd7afcfSStefano Zampini Mat M; 22461dd7afcfSStefano Zampini 22471dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2248339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 22491dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 22501dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2251a3df083aSStefano Zampini } 22524f1b2e48SStefano Zampini /* store global idxs for p0 */ 22534f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2254339f8db1SStefano Zampini } 2255ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 22564f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 2257b0f5fe93SStefano Zampini 2258b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 2259b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 226027b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 226127b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2262339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2263339f8db1SStefano Zampini PetscFunctionReturn(0); 2264339f8db1SStefano Zampini } 2265339f8db1SStefano Zampini 2266339f8db1SStefano Zampini #undef __FUNCT__ 2267015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 2268015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 2269efc2fbd9SStefano Zampini { 2270efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2271de9d7bd0SStefano Zampini PetscScalar *array; 2272efc2fbd9SStefano Zampini PetscErrorCode ierr; 2273efc2fbd9SStefano Zampini 2274efc2fbd9SStefano Zampini PetscFunctionBegin; 2275efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 2276efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 22774f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2278efc2fbd9SStefano Zampini } 2279de9d7bd0SStefano Zampini if (get) { 2280efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 22814f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 22824f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 2283efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 2284de9d7bd0SStefano Zampini } else { 2285de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 2286de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2287de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2288de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 2289efc2fbd9SStefano Zampini } 2290efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2291efc2fbd9SStefano Zampini } 2292efc2fbd9SStefano Zampini 2293efc2fbd9SStefano Zampini #undef __FUNCT__ 2294c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 2295c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 2296c263805aSStefano Zampini { 2297c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2298c263805aSStefano Zampini PetscErrorCode ierr; 2299c263805aSStefano Zampini 2300c263805aSStefano Zampini PetscFunctionBegin; 2301c263805aSStefano Zampini /* TODO: add error checking 2302c263805aSStefano Zampini - avoid nested pop (or push) calls. 2303c263805aSStefano Zampini - cannot push before pop. 23041c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 2305c263805aSStefano Zampini */ 23064f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2307efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2308efc2fbd9SStefano Zampini } 2309c263805aSStefano Zampini if (pop) { 2310a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 23114f1b2e48SStefano Zampini IS is_p0; 23124f1b2e48SStefano Zampini MatReuse reuse; 2313c263805aSStefano Zampini 2314c263805aSStefano Zampini /* extract B_0 */ 23154f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 23164f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 23174f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 23184f1b2e48SStefano Zampini } 23194f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 23204f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 2321c263805aSStefano Zampini /* remove rows and cols from local problem */ 2322c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 232397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 23244f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 23254f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 2326a3df083aSStefano Zampini } else { 2327a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2328a3df083aSStefano Zampini PetscScalar *vals; 2329a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 2330a3df083aSStefano Zampini 2331a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 2332a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 2333a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 23340b5adadeSStefano Zampini PetscInt *nnz; 2335a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 2336a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 2337a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2338331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 2339331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2340331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 2341331e053bSStefano Zampini nnz[i] = n - nnz[i]; 2342331e053bSStefano Zampini } 2343331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 2344331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2345331e053bSStefano Zampini } 2346a3df083aSStefano Zampini 2347a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2348a3df083aSStefano Zampini PetscScalar *array; 2349a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 2350a3df083aSStefano Zampini 2351a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 2352a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2353a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2354a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 2355a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2356a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 2357a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 2358a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 2359a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 2360a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 2361a3df083aSStefano Zampini cum = 0; 2362a3df083aSStefano Zampini for (j=0;j<n;j++) { 236322db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 2364a3df083aSStefano Zampini vals[cum] = array[j]; 2365a3df083aSStefano Zampini idxs_ins[cum] = j; 2366a3df083aSStefano Zampini cum++; 2367a3df083aSStefano Zampini } 2368a3df083aSStefano Zampini } 2369a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 2370a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 2371a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2372a3df083aSStefano Zampini } 2373a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2374a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2375a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 2376a3df083aSStefano Zampini } 2377c263805aSStefano Zampini } else { /* push */ 2378a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 23794f1b2e48SStefano Zampini PetscInt i; 23804f1b2e48SStefano Zampini 23814f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 23824f1b2e48SStefano Zampini PetscScalar *B0_vals; 23834f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 23844f1b2e48SStefano Zampini 23854f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 23864f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 23877b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 23884f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 23894f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 23904f1b2e48SStefano Zampini } 2391c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2392c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2393a3df083aSStefano Zampini } else { 2394a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 2395a3df083aSStefano Zampini } 2396c263805aSStefano Zampini } 2397c263805aSStefano Zampini PetscFunctionReturn(0); 2398c263805aSStefano Zampini } 2399c263805aSStefano Zampini 2400c263805aSStefano Zampini #undef __FUNCT__ 2401b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 240208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 2403b1b3d7a2SStefano Zampini { 2404b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 240508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 240608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 240708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 240808122e43SStefano Zampini PetscScalar *work,lwork; 240908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 241008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 241108122e43SStefano Zampini PetscReal *eigs,thresh; 24121b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 2413f6f667cfSStefano Zampini PetscBool allocated_S_St; 241408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 241508122e43SStefano Zampini PetscReal *rwork; 241608122e43SStefano Zampini #endif 2417b1b3d7a2SStefano Zampini PetscErrorCode ierr; 2418b1b3d7a2SStefano Zampini 2419b1b3d7a2SStefano Zampini PetscFunctionBegin; 2420b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 2421af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 2422af25d912SStefano 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); 242306a4e24aSStefano Zampini 2424fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2425fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2426fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2427fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 24281575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2429fd14bc51SStefano Zampini } 2430fd14bc51SStefano Zampini 2431e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 2432e496cd5dSStefano 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); 2433e496cd5dSStefano Zampini } 2434e496cd5dSStefano Zampini 243508122e43SStefano Zampini /* max size of subsets */ 243608122e43SStefano Zampini mss = 0; 243708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 243808122e43SStefano Zampini PetscInt subset_size; 2439862806e4SStefano Zampini 244008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 244108122e43SStefano Zampini mss = PetscMax(mss,subset_size); 244208122e43SStefano Zampini } 244308122e43SStefano Zampini 244408122e43SStefano Zampini /* min/max and threshold */ 244508122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 2446f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 244708122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 2448f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 2449f6f667cfSStefano Zampini if (nmin) { 2450f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 2451f6f667cfSStefano Zampini } 245208122e43SStefano Zampini 245308122e43SStefano Zampini /* allocate lapack workspace */ 245408122e43SStefano Zampini cum = cum2 = 0; 245508122e43SStefano Zampini maxneigs = 0; 245608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 245708122e43SStefano Zampini PetscInt n,subset_size; 2458f6f667cfSStefano Zampini 245908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 246008122e43SStefano Zampini n = PetscMin(subset_size,nmax); 24619162d606SStefano Zampini cum += subset_size; 24629162d606SStefano Zampini cum2 += subset_size*n; 246308122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 246408122e43SStefano Zampini } 246508122e43SStefano Zampini if (mss) { 24669ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 246708122e43SStefano Zampini PetscBLASInt B_itype = 1; 246808122e43SStefano Zampini PetscBLASInt B_N = mss; 24694c6709b3SStefano Zampini PetscReal zero = 0.0; 24704c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 247108122e43SStefano Zampini 247208122e43SStefano Zampini B_lwork = -1; 247308122e43SStefano Zampini S = NULL; 247408122e43SStefano Zampini St = NULL; 2475a58a30b4SStefano Zampini eigs = NULL; 2476a58a30b4SStefano Zampini eigv = NULL; 2477a58a30b4SStefano Zampini B_iwork = NULL; 2478a58a30b4SStefano Zampini B_ifail = NULL; 2479d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2480d1710679SStefano Zampini rwork = NULL; 2481d1710679SStefano Zampini #endif 24828bec7fa6SStefano Zampini thresh = 1.0; 248308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 248408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 248508122e43SStefano 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)); 248608122e43SStefano Zampini #else 248708122e43SStefano 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)); 248808122e43SStefano Zampini #endif 248908122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 249008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 249108122e43SStefano Zampini } else { 249208122e43SStefano Zampini /* TODO */ 249308122e43SStefano Zampini } 249408122e43SStefano Zampini } else { 249508122e43SStefano Zampini lwork = 0; 249608122e43SStefano Zampini } 249708122e43SStefano Zampini 249808122e43SStefano Zampini nv = 0; 2499d62866d3SStefano 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) */ 2500d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 250108122e43SStefano Zampini } 25024c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 2503f6f667cfSStefano Zampini if (allocated_S_St) { 2504f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 2505f6f667cfSStefano Zampini } 2506f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 250708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 250808122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 250908122e43SStefano Zampini #endif 25109162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 25119162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 25129162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 251308122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 25149162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 251508122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 251608122e43SStefano Zampini 251708122e43SStefano Zampini maxneigs = 0; 251872b8c272SStefano Zampini cum = cumarray = 0; 25199162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 25209162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 2521d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 252208122e43SStefano Zampini const PetscInt *idxs; 252308122e43SStefano Zampini 2524d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 252508122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 252608122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 252708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 252808122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 25299162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 25309162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 253108122e43SStefano Zampini } 2532d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 253308122e43SStefano Zampini } 253408122e43SStefano Zampini 253508122e43SStefano Zampini if (mss) { /* multilevel */ 253608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 253708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 253808122e43SStefano Zampini } 253908122e43SStefano Zampini 2540ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 254108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 254208122e43SStefano Zampini const PetscInt *idxs; 25439d54b7f4SStefano Zampini PetscReal upper,lower; 2544862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 254508122e43SStefano Zampini PetscBLASInt B_N; 2546aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 254708122e43SStefano Zampini 25489d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 25499d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 25509d54b7f4SStefano Zampini lower = thresh; 25519d54b7f4SStefano Zampini } else { 25529d54b7f4SStefano Zampini upper = 1./thresh; 25539d54b7f4SStefano Zampini lower = 0.; 25549d54b7f4SStefano Zampini } 2555862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 2556ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 2557f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 2558f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 25599ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 2560aff50787SStefano Zampini PetscInt j,k; 2561aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 2562aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 2563aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 256408122e43SStefano Zampini } 256508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 2566aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 2567aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 2568aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 2569aff50787SStefano Zampini } 257008122e43SStefano Zampini } 257108122e43SStefano Zampini } else { 257208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 257308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 257408122e43SStefano Zampini } 25758bec7fa6SStefano Zampini } else { 2576f6f667cfSStefano Zampini S = Sarray + cumarray; 2577f6f667cfSStefano Zampini St = Starray + cumarray; 25788bec7fa6SStefano Zampini } 2579aff50787SStefano Zampini /* see if we can save some work */ 2580b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 2581aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 2582aff50787SStefano Zampini } 2583aff50787SStefano Zampini 2584b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 2585aff50787SStefano Zampini B_neigs = 0; 2586aff50787SStefano Zampini } else { 25879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 258808122e43SStefano Zampini PetscBLASInt B_itype = 1; 2589f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 25904c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 25919552c7c7SStefano Zampini PetscInt nmin_s; 2592b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 259308122e43SStefano Zampini 2594fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 25958bec7fa6SStefano 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]]); 2596fd14bc51SStefano Zampini } 2597d16cbb6bSStefano Zampini 2598b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 2599b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 2600b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 2601b7ab4a40SStefano Zampini } 2602b7ab4a40SStefano Zampini 260308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2604b7ab4a40SStefano Zampini if (compute_range) { 2605d16cbb6bSStefano Zampini 2606d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 260708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 26089d54b7f4SStefano 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)); 260908122e43SStefano Zampini #else 26109d54b7f4SStefano 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)); 261108122e43SStefano Zampini #endif 2612b7ab4a40SStefano Zampini } else if (!same_data) { 2613d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 2614d16cbb6bSStefano Zampini B_IL = 1; 2615d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 26169d54b7f4SStefano 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)); 2617d16cbb6bSStefano Zampini #else 26189d54b7f4SStefano 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)); 2619d16cbb6bSStefano Zampini #endif 2620b03ebc13SStefano Zampini } else { /* same_data is true, so just get the adaptive functional requested by the user */ 2621b7ab4a40SStefano Zampini PetscInt k; 2622b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 2623b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 2624b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 2625b7ab4a40SStefano Zampini nmin = nmax; 2626b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 2627b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 2628b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 2629b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 2630b7ab4a40SStefano Zampini } 2631d16cbb6bSStefano Zampini } 263208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 263308122e43SStefano Zampini if (B_ierr) { 26346c4ed002SBarry 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); 26356c4ed002SBarry 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); 26366c4ed002SBarry 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); 263708122e43SStefano Zampini } 263808122e43SStefano Zampini 263908122e43SStefano Zampini if (B_neigs > nmax) { 2640fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2641fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 2642fd14bc51SStefano Zampini } 26439d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 264408122e43SStefano Zampini B_neigs = nmax; 264508122e43SStefano Zampini } 264608122e43SStefano Zampini 26479552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 26489552c7c7SStefano Zampini if (B_neigs < nmin_s) { 264908122e43SStefano Zampini PetscBLASInt B_neigs2; 265008122e43SStefano Zampini 26519d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2652f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 26539d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 26549d54b7f4SStefano Zampini } else { 26559d54b7f4SStefano Zampini B_IL = B_neigs + 1; 26569d54b7f4SStefano Zampini B_IU = nmin_s; 26579d54b7f4SStefano Zampini } 2658fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2659fd14bc51SStefano 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); 2660fd14bc51SStefano Zampini } 26619ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 26621ae86dd6SStefano Zampini PetscInt j,k; 266308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 26641ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 26651ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 26661ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 266708122e43SStefano Zampini } 266808122e43SStefano Zampini } 266908122e43SStefano Zampini } else { 267008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 267108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 267208122e43SStefano Zampini } 267308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 267408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 26759d54b7f4SStefano 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)); 267608122e43SStefano Zampini #else 26779d54b7f4SStefano 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)); 267808122e43SStefano Zampini #endif 267908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 268008122e43SStefano Zampini B_neigs += B_neigs2; 268108122e43SStefano Zampini } 268208122e43SStefano Zampini if (B_ierr) { 26836c4ed002SBarry 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); 26846c4ed002SBarry 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); 26856c4ed002SBarry 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); 268608122e43SStefano Zampini } 2687fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2688ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 268908122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 269008122e43SStefano Zampini if (eigs[j] == 0.0) { 2691ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 269208122e43SStefano Zampini } else { 26939d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2694ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 26959d54b7f4SStefano Zampini } else { 26969d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 26979d54b7f4SStefano Zampini } 2698fd14bc51SStefano Zampini } 269908122e43SStefano Zampini } 270008122e43SStefano Zampini } 270108122e43SStefano Zampini } else { 270208122e43SStefano Zampini /* TODO */ 270308122e43SStefano Zampini } 2704aff50787SStefano Zampini } 27056c3e6151SStefano Zampini /* change the basis back to the original one */ 27066c3e6151SStefano Zampini if (sub_schurs->change) { 270772b8c272SStefano Zampini Mat change,phi,phit; 27086c3e6151SStefano Zampini 27096c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 27106c3e6151SStefano Zampini PetscInt ii; 27116c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 27126c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 27136c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 2714684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 2715684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 2716684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 2717684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 2718684229deSStefano Zampini #else 27196c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 2720684229deSStefano Zampini #endif 27216c3e6151SStefano Zampini } 27226c3e6151SStefano Zampini } 27236c3e6151SStefano Zampini } 272472b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 27256c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 272672b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 27276c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 27286c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 27296c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 27306c3e6151SStefano Zampini } 27318bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 27328bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 27339162d606SStefano Zampini if (B_neigs) { 27349162d606SStefano 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); 2735fd14bc51SStefano Zampini 2736fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 27379552c7c7SStefano Zampini PetscInt ii; 27389552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 2739ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 27409552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2741ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 2742ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2743ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2744ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 2745ac47001eSStefano Zampini #else 2746ac47001eSStefano 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); 2747ac47001eSStefano Zampini #endif 27489552c7c7SStefano Zampini } 27499552c7c7SStefano Zampini } 2750fd14bc51SStefano Zampini } 27519162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 27529162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 27539162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 27549162d606SStefano Zampini cum++; 275508122e43SStefano Zampini } 275608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 275708122e43SStefano Zampini /* shift for next computation */ 275808122e43SStefano Zampini cumarray += subset_size*subset_size; 275908122e43SStefano Zampini } 2760fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2761fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2762fd14bc51SStefano Zampini } 276308122e43SStefano Zampini 276408122e43SStefano Zampini if (mss) { 276508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 276608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 2767f6f667cfSStefano Zampini /* destroy matrices (junk) */ 2768f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 2769f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 277008122e43SStefano Zampini } 2771f6f667cfSStefano Zampini if (allocated_S_St) { 2772f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 2773f6f667cfSStefano Zampini } 2774f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 277508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 277608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 277708122e43SStefano Zampini #endif 277808122e43SStefano Zampini if (pcbddc->dbg_flag) { 27791b968477SStefano Zampini PetscInt maxneigs_r; 2780b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 27819b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 278208122e43SStefano Zampini } 278308122e43SStefano Zampini PetscFunctionReturn(0); 278408122e43SStefano Zampini } 2785b1b3d7a2SStefano Zampini 2786674ae819SStefano Zampini #undef __FUNCT__ 2787c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 2788c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 2789c8587f34SStefano Zampini { 27908629588bSStefano Zampini PetscScalar *coarse_submat_vals; 2791c8587f34SStefano Zampini PetscErrorCode ierr; 2792c8587f34SStefano Zampini 2793c8587f34SStefano Zampini PetscFunctionBegin; 2794f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 27955e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 2796c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 2797c8587f34SStefano Zampini 2798684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 27990fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 2800684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 2801c8587f34SStefano Zampini 28028629588bSStefano Zampini /* 28038629588bSStefano Zampini Setup local correction and local part of coarse basis. 28048629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 28058629588bSStefano Zampini */ 280647f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 28078629588bSStefano Zampini 28088629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 28098629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 28108629588bSStefano Zampini 28118629588bSStefano Zampini /* free */ 28128629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 2813c8587f34SStefano Zampini PetscFunctionReturn(0); 2814c8587f34SStefano Zampini } 2815c8587f34SStefano Zampini 2816c8587f34SStefano Zampini #undef __FUNCT__ 2817674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 2818674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 2819674ae819SStefano Zampini { 2820674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2821674ae819SStefano Zampini PetscErrorCode ierr; 2822674ae819SStefano Zampini 2823674ae819SStefano Zampini PetscFunctionBegin; 2824674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 282530368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 2826674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 2827785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 2828674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 2829f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2830f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2831785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 283263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 283363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 2834674ae819SStefano Zampini PetscFunctionReturn(0); 2835674ae819SStefano Zampini } 2836674ae819SStefano Zampini 2837674ae819SStefano Zampini #undef __FUNCT__ 2838674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 2839674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 2840674ae819SStefano Zampini { 2841674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 28424f1b2e48SStefano Zampini PetscInt i; 2843674ae819SStefano Zampini PetscErrorCode ierr; 2844674ae819SStefano Zampini 2845674ae819SStefano Zampini PetscFunctionBegin; 2846a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 2847b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2848674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 284916909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 28501dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 2851674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2852669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 2853fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 2854a13144ffSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2855674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 28564f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 28574f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 28584f1b2e48SStefano Zampini } 28594f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 2860b334f244SStefano Zampini if (pcbddc->sub_schurs) { 2861b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 2862b334f244SStefano Zampini } 2863c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 28648af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 2865674ae819SStefano Zampini PetscFunctionReturn(0); 2866674ae819SStefano Zampini } 2867674ae819SStefano Zampini 2868674ae819SStefano Zampini #undef __FUNCT__ 2869674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 2870674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 2871674ae819SStefano Zampini { 2872674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2873674ae819SStefano Zampini PetscErrorCode ierr; 2874674ae819SStefano Zampini 2875674ae819SStefano Zampini PetscFunctionBegin; 2876674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 287758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 2878ca92afb2SStefano Zampini PetscScalar *array; 287906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 288006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 288158da7f69SStefano Zampini } 2882674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 2883674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 288415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 288515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 2886674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 2887674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 2888674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 288906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 2890674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 2891674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 28928ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2893674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2894674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2895674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 2896f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 2897f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 2898f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 2899f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2900727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 29010e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2902f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 290370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 290481d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 29050369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 29061dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 29074f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 29088b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 2909ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 2910ca92afb2SStefano Zampini PetscInt i; 2911ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2912ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 2913ca92afb2SStefano Zampini } 2914ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 2915ca92afb2SStefano Zampini } 29164f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 2917674ae819SStefano Zampini PetscFunctionReturn(0); 2918674ae819SStefano Zampini } 2919674ae819SStefano Zampini 2920674ae819SStefano Zampini #undef __FUNCT__ 2921f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 2922f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 29236bfb1811SStefano Zampini { 29246bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 29256bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 29266bfb1811SStefano Zampini VecType impVecType; 29274f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 29286bfb1811SStefano Zampini PetscErrorCode ierr; 29296bfb1811SStefano Zampini 29306bfb1811SStefano Zampini PetscFunctionBegin; 29316c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 2932e7b262bdSStefano Zampini /* get sizes */ 29334f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 2934b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 29356bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 2936e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 2937e7b262bdSStefano Zampini /* R nodes */ 2938e7b262bdSStefano Zampini old_size = -1; 2939e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 2940e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 2941e7b262bdSStefano Zampini } 2942e7b262bdSStefano Zampini if (n_R != old_size) { 2943e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 2944e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 29456bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 29466bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 29476bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 29486bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 2949e7b262bdSStefano Zampini } 2950e7b262bdSStefano Zampini /* local primal dofs */ 2951e7b262bdSStefano Zampini old_size = -1; 2952e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 2953e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 2954e7b262bdSStefano Zampini } 2955e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 2956e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 295783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 2958e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 29596bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 2960e7b262bdSStefano Zampini } 2961e7b262bdSStefano Zampini /* local explicit constraints */ 2962e7b262bdSStefano Zampini old_size = -1; 2963e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 2964e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 2965e7b262bdSStefano Zampini } 2966e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 2967e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 296883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 296983b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 297083b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 297183b7ccabSStefano Zampini } 29726bfb1811SStefano Zampini PetscFunctionReturn(0); 29736bfb1811SStefano Zampini } 29746bfb1811SStefano Zampini 29756bfb1811SStefano Zampini #undef __FUNCT__ 297647f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 297747f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 297888ebb749SStefano Zampini { 297925084f0cSStefano Zampini PetscErrorCode ierr; 298025084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 298188ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 298288ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2983d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 298425084f0cSStefano Zampini /* submatrices of local problem */ 298580677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 298606656605SStefano Zampini /* submatrices of local coarse problem */ 298706656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 298825084f0cSStefano Zampini /* working matrices */ 298906656605SStefano Zampini Mat C_CR; 299025084f0cSStefano Zampini /* additional working stuff */ 299106656605SStefano Zampini PC pc_R; 29924f1b2e48SStefano Zampini Mat F; 29935cbda25cSStefano Zampini Vec dummy_vec; 2994a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 299525084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 299606656605SStefano Zampini PetscScalar *work; 299706656605SStefano Zampini PetscInt *idx_V_B; 2998ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 299906656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 3000ffd830a3SStefano Zampini 300125084f0cSStefano Zampini /* some shortcuts to scalars */ 300206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 300388ebb749SStefano Zampini 300488ebb749SStefano Zampini PetscFunctionBegin; 30059a962809SStefano 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"); 3006ffd830a3SStefano Zampini 3007ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 3008b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 30094f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 3010b371cd4fSStefano Zampini n_B = pcis->n_B; 3011b371cd4fSStefano Zampini n_D = pcis->n - n_B; 301288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 301388ebb749SStefano Zampini 301488ebb749SStefano Zampini /* vertices in boundary numbering */ 3015785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 30160e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 30176c4ed002SBarry 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); 301888ebb749SStefano Zampini 301906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 3020019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 302106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 302206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 302306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 302406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 302506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 302606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 302706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 302806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 302906656605SStefano Zampini 303006656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 303106656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 303206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 303306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 303406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 3035ffd830a3SStefano Zampini lda_rhs = n_R; 3036a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 303706656605SStefano Zampini if (isLU || isILU || isCHOL) { 303806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 3039b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 3040df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3041d62866d3SStefano Zampini MatFactorType type; 3042d62866d3SStefano Zampini 3043df4d28bfSStefano Zampini F = reuse_solver->F; 30446816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 3045d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 3046ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 304722db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 304806656605SStefano Zampini } else { 304906656605SStefano Zampini F = NULL; 305006656605SStefano Zampini } 305106656605SStefano Zampini 3052ffd830a3SStefano Zampini /* allocate workspace */ 3053ffd830a3SStefano Zampini n = 0; 3054ffd830a3SStefano Zampini if (n_constraints) { 3055ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 3056ffd830a3SStefano Zampini } 3057ffd830a3SStefano Zampini if (n_vertices) { 3058ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 3059ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 3060ffd830a3SStefano Zampini } 3061ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 3062ffd830a3SStefano Zampini 30635cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 30645cbda25cSStefano Zampini dummy_vec = NULL; 30655cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 30665cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 30675cbda25cSStefano Zampini } 30685cbda25cSStefano Zampini 306988ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 307088ebb749SStefano Zampini if (n_constraints) { 307172b8c272SStefano Zampini Mat M1,M2,M3,C_B; 307206656605SStefano Zampini IS is_aux; 307380677318SStefano Zampini PetscScalar *array,*array2; 307406656605SStefano Zampini 3075f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 307680677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 307788ebb749SStefano Zampini 307825084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 307925084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 30808ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 308172b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 308288ebb749SStefano Zampini 308380677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 308480677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 3085ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 308688ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 308706656605SStefano Zampini const PetscScalar *row_cmat_values; 308806656605SStefano Zampini const PetscInt *row_cmat_indices; 308906656605SStefano Zampini PetscInt size_of_constraint,j; 309088ebb749SStefano Zampini 309106656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 309206656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 3093ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 309406656605SStefano Zampini } 309506656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 309606656605SStefano Zampini } 3097ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 309806656605SStefano Zampini if (F) { 309906656605SStefano Zampini Mat B; 310006656605SStefano Zampini 3101ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 3102a3df083aSStefano Zampini if (need_benign_correction) { 3103df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3104a3df083aSStefano Zampini 310572b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 310672b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 3107a3df083aSStefano Zampini } 310880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 3109a3df083aSStefano Zampini if (need_benign_correction) { 3110a3df083aSStefano Zampini PetscScalar *marr; 3111df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3112a3df083aSStefano Zampini 3113a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 31145cbda25cSStefano Zampini if (lda_rhs != n_R) { 31155cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 31165cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 31175cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 31185cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 31195cbda25cSStefano Zampini } 31205cbda25cSStefano Zampini } else { 3121a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 3122a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 31235cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3124a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3125a3df083aSStefano Zampini } 31265cbda25cSStefano Zampini } 3127a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 3128a3df083aSStefano Zampini } 312906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 313006656605SStefano Zampini } else { 313180677318SStefano Zampini PetscScalar *marr; 313280677318SStefano Zampini 313380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 313406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 3135ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3136ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 313706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 313806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 313906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 314006656605SStefano Zampini } 314180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 314206656605SStefano Zampini } 314380677318SStefano Zampini if (!pcbddc->switch_static) { 314480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 314580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 314680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 314780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 3148ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 314980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 315080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 315380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 315480677318SStefano Zampini } 315580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 315680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 315772b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 315880677318SStefano Zampini } else { 3159ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3160ffd830a3SStefano Zampini IS dummy; 3161ffd830a3SStefano Zampini 3162ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 316372b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 3164ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 3165ffd830a3SStefano Zampini } else { 316680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 316780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 3168ffd830a3SStefano Zampini } 316925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 317080677318SStefano Zampini } 317180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 317280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 317380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 317406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 317506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 317680677318SStefano Zampini if (isCHOL) { 317780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 317880677318SStefano Zampini } else { 317925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 318080677318SStefano Zampini } 318180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 318206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 318325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 318425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 318525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 318680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 318772b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 318872b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 318906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 319006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 3191f4ddd8eeSStefano Zampini } 3192fc227af8SStefano Zampini 3193fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 319488ebb749SStefano Zampini if (n_vertices) { 319506656605SStefano Zampini IS is_aux; 31963a50541eSStefano Zampini 3197b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 31986816873aSStefano Zampini IS tis; 31996816873aSStefano Zampini 32006816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 32016816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 32026816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 32036816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 32046816873aSStefano Zampini } else { 32053a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 32066816873aSStefano Zampini } 32079577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 32089577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 320904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 321025084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 321188ebb749SStefano Zampini } 321288ebb749SStefano Zampini 321388ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 3214f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 321506656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 321606656605SStefano Zampini if (pcbddc->coarse_phi_D) { 321706656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 321806656605SStefano Zampini } 3219f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 322006656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 322106656605SStefano Zampini PetscScalar *marray; 322206656605SStefano Zampini 322306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 322406656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 3225f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3226f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3227f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3228f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3229f4ddd8eeSStefano Zampini } 3230f4ddd8eeSStefano Zampini } 323106656605SStefano Zampini 3232f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 323306656605SStefano Zampini PetscScalar *marray; 323488ebb749SStefano Zampini 323506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 32368eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 323706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 323888ebb749SStefano Zampini } 32393301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 324006656605SStefano Zampini n *= 2; 324188ebb749SStefano Zampini } 324206656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 324306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 324406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 32458eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 324606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 324706656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 324888ebb749SStefano Zampini } 32493301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 325006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 32518eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 325206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 325306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 325488ebb749SStefano Zampini } 325588ebb749SStefano Zampini } else { 3256c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 3257c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 32581b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3259c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 3260c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 3261c0553b1fSStefano Zampini } 326288ebb749SStefano Zampini } 326306656605SStefano Zampini } 3264019a44ceSStefano Zampini 326506656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 32664f1b2e48SStefano Zampini p0_lidx_I = NULL; 32674f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 3268d12edf2fSStefano Zampini const PetscInt *idxs; 3269d12edf2fSStefano Zampini 3270d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 32714f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 32724f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 32734f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 32744f1b2e48SStefano Zampini } 3275d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 3276d12edf2fSStefano Zampini } 3277d16cbb6bSStefano Zampini 327806656605SStefano Zampini /* vertices */ 327906656605SStefano Zampini if (n_vertices) { 328016f15bc4SStefano Zampini 3281af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 328204708bb6SStefano Zampini 328316f15bc4SStefano Zampini if (n_R) { 328414393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 328506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 328616f15bc4SStefano Zampini PetscScalar *x,*y; 328704708bb6SStefano Zampini PetscBool isseqaij; 328806656605SStefano Zampini 328921eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 329014393ed6SStefano Zampini if (need_benign_correction) { 329114393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 329214393ed6SStefano Zampini IS is_p0; 329314393ed6SStefano Zampini PetscInt *idxs_p0,n; 329414393ed6SStefano Zampini 329514393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 329614393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 329714393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 3298af25d912SStefano 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); 329914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 330014393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 330114393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 330214393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 330314393ed6SStefano Zampini } 330414393ed6SStefano Zampini 3305ffd830a3SStefano Zampini if (lda_rhs == n_R) { 3306af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3307ffd830a3SStefano Zampini } else { 3308ca92afb2SStefano Zampini PetscScalar *av,*array; 3309ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 3310ca92afb2SStefano Zampini PetscInt n; 3311ca92afb2SStefano Zampini PetscBool flg_row; 3312ffd830a3SStefano Zampini 3313ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 3314ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 33159d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3316ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3317ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 3318ca92afb2SStefano Zampini for (i=0;i<n;i++) { 3319ca92afb2SStefano Zampini PetscInt j; 3320ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 3321ffd830a3SStefano Zampini } 3322ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3323ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3324ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 3325ffd830a3SStefano Zampini } 3326ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3327a3df083aSStefano Zampini if (need_benign_correction) { 3328df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3329a3df083aSStefano Zampini PetscScalar *marr; 3330a3df083aSStefano Zampini 3331a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 333214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 333314393ed6SStefano Zampini 333414393ed6SStefano Zampini | 0 0 0 | (V) 333514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 333614393ed6SStefano Zampini | 0 0 -1 | (p0) 333714393ed6SStefano Zampini 333814393ed6SStefano Zampini */ 3339df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 334014393ed6SStefano Zampini const PetscScalar *vals; 334114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 334214393ed6SStefano Zampini PetscInt n,j,nz; 334314393ed6SStefano Zampini 3344df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3345df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 334614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 334714393ed6SStefano Zampini for (j=0;j<n;j++) { 334814393ed6SStefano Zampini PetscScalar val = vals[j]; 334914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 335014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 335114393ed6SStefano Zampini } 335214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3353df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 335414393ed6SStefano Zampini } 335572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 335672b8c272SStefano Zampini } 335772b8c272SStefano Zampini if (F) { 335814393ed6SStefano Zampini /* need to correct the rhs */ 335972b8c272SStefano Zampini if (need_benign_correction) { 336072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 336172b8c272SStefano Zampini PetscScalar *marr; 336272b8c272SStefano Zampini 336372b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 33645cbda25cSStefano Zampini if (lda_rhs != n_R) { 33655cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 33665cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 33675cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 33685cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 33695cbda25cSStefano Zampini } 33705cbda25cSStefano Zampini } else { 3371a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3372a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 33735cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3374a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3375a3df083aSStefano Zampini } 33765cbda25cSStefano Zampini } 3377a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 3378a3df083aSStefano Zampini } 337906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 338014393ed6SStefano Zampini /* need to correct the solution */ 3381a3df083aSStefano Zampini if (need_benign_correction) { 3382df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3383a3df083aSStefano Zampini PetscScalar *marr; 3384a3df083aSStefano Zampini 3385a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 33865cbda25cSStefano Zampini if (lda_rhs != n_R) { 33875cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 33885cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 33895cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 33905cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 33915cbda25cSStefano Zampini } 33925cbda25cSStefano Zampini } else { 3393a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3394a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 33955cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3396a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3397a3df083aSStefano Zampini } 33985cbda25cSStefano Zampini } 3399a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 3400a3df083aSStefano Zampini } 340106656605SStefano Zampini } else { 340206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 340306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 3404ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 3405ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 340606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 340706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 340806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 340906656605SStefano Zampini } 341006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 341106656605SStefano Zampini } 341280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3413ffd830a3SStefano Zampini /* S_VV and S_CV */ 341406656605SStefano Zampini if (n_constraints) { 341506656605SStefano Zampini Mat B; 341680677318SStefano Zampini 3417ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 341880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 3419ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3420ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 342180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 342380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 342480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 342580677318SStefano Zampini } 3426ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 342780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 342880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3429ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 343080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 343106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 3432ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 3433ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 343406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 343506656605SStefano Zampini } 343604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 343704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 3438511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 343904708bb6SStefano Zampini } 3440ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3441ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 3442ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3443ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 3444ffd830a3SStefano Zampini } 344506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 344614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 344714393ed6SStefano Zampini if (need_benign_correction) { 3448df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 344914393ed6SStefano Zampini PetscScalar *marr,*sums; 345014393ed6SStefano Zampini 345114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 3452f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 3453df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 345414393ed6SStefano Zampini const PetscScalar *vals; 345514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 345614393ed6SStefano Zampini PetscInt n,j,nz; 345714393ed6SStefano Zampini 3458df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3459df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 346014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 346114393ed6SStefano Zampini PetscInt k; 346214393ed6SStefano Zampini sums[j] = 0.; 346314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 346414393ed6SStefano Zampini } 346514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 346614393ed6SStefano Zampini for (j=0;j<n;j++) { 346714393ed6SStefano Zampini PetscScalar val = vals[j]; 346814393ed6SStefano Zampini PetscInt k; 346914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 347014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 347114393ed6SStefano Zampini } 347214393ed6SStefano Zampini } 347314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3474df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 347514393ed6SStefano Zampini } 347614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 3477f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 347814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 347914393ed6SStefano Zampini } 348080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 348106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 348206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 348306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 348406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 348506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 348606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 348706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3488d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 3489019a44ceSStefano Zampini } else { 3490d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3491d16cbb6bSStefano Zampini } 349221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 3493d16cbb6bSStefano Zampini 349406656605SStefano Zampini /* coarse basis functions */ 349506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 349616f15bc4SStefano Zampini PetscScalar *y; 349716f15bc4SStefano Zampini 3498ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 349906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 350006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 350106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 350206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 350306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 350406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 350506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 350606656605SStefano Zampini 350706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 35084f1b2e48SStefano Zampini PetscInt j; 35094f1b2e48SStefano Zampini 351006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 351106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 351206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 351306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 351406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 35154f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 351606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 351706656605SStefano Zampini } 351806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 351906656605SStefano Zampini } 352004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 352104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 352206656605SStefano Zampini } 35235cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 352406656605SStefano Zampini 352506656605SStefano Zampini if (n_constraints) { 352606656605SStefano Zampini Mat B; 352706656605SStefano Zampini 3528ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 352906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 353080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 353106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 353206656605SStefano Zampini if (n_vertices) { 353380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 353480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 353580677318SStefano Zampini } else { 353680677318SStefano Zampini Mat S_VCt; 353780677318SStefano Zampini 3538ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3539ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 354072b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 3541ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 3542ffd830a3SStefano Zampini } 354380677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 354480677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 354580677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 354680677318SStefano Zampini } 354706656605SStefano Zampini } 354806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 354906656605SStefano Zampini /* coarse basis functions */ 355006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 355106656605SStefano Zampini PetscScalar *y; 355206656605SStefano Zampini 3553ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 355406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 355506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 355606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 355706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 355806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 355906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 356006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 35614f1b2e48SStefano Zampini PetscInt j; 35624f1b2e48SStefano Zampini 356306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 356406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 356506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 356606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 356706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 35684f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 356906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 357006656605SStefano Zampini } 357106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 357206656605SStefano Zampini } 357306656605SStefano Zampini } 357480677318SStefano Zampini if (n_constraints) { 357580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 357680677318SStefano Zampini } 35774f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 357872b8c272SStefano Zampini 357972b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 358072b8c272SStefano Zampini if (pcbddc->benign_n) { 358172b8c272SStefano Zampini Mat B0_B,B0_BPHI; 358272b8c272SStefano Zampini IS is_dummy; 358372b8c272SStefano Zampini PetscScalar *data; 358472b8c272SStefano Zampini PetscInt j; 358572b8c272SStefano Zampini 358672b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 358772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 358872b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 358972b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 359086c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 359172b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 359272b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 359372b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 359472b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 359572b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 359672b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 359772b8c272SStefano Zampini } 359872b8c272SStefano Zampini } 359972b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 360072b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 360172b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 360272b8c272SStefano Zampini } 3603019a44ceSStefano Zampini 360406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 36053301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 3606ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 3607ffd830a3SStefano Zampini PetscScalar *marray; 360806656605SStefano Zampini 360906656605SStefano Zampini if (n_constraints) { 3610ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 361106656605SStefano Zampini 3612af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 361306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 3614ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 361516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 361606656605SStefano Zampini if (n_vertices) { 3617ffd830a3SStefano Zampini Mat S_VCT; 361806656605SStefano Zampini 361906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 3620ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 362116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 362206656605SStefano Zampini } 3623ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 36245b782168SStefano Zampini } else { 36255b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 362606656605SStefano Zampini } 362716f15bc4SStefano Zampini if (n_vertices && n_R) { 3628ffd830a3SStefano Zampini PetscScalar *av,*marray; 3629ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 3630ffd830a3SStefano Zampini PetscInt n; 3631ffd830a3SStefano Zampini PetscBool flg_row; 363206656605SStefano Zampini 3633ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 3634af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 3635ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3636ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 3637ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 3638ffd830a3SStefano Zampini for (i=0;i<n;i++) { 3639ffd830a3SStefano Zampini PetscInt j; 3640ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 3641ffd830a3SStefano Zampini } 3642ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 3643ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3644ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 364506656605SStefano Zampini } 364606656605SStefano Zampini 3647ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 3648ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 3649ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 3650ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 3651ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 365206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 365306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 365406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 365506656605SStefano Zampini } 3656ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 36575b782168SStefano Zampini if (B_C) { 3658ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 3659ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 3660ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 3661ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 3662ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 3663ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3664ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 366506656605SStefano Zampini } 3666ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 36675b782168SStefano Zampini } 366806656605SStefano Zampini /* coarse basis functions */ 366906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 367006656605SStefano Zampini PetscScalar *y; 367106656605SStefano Zampini 3672ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 367306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 367406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 367506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 367606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 367706656605SStefano Zampini if (i<n_vertices) { 367806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 367906656605SStefano Zampini } 368006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 368106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 368206656605SStefano Zampini 368306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 368406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 368506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 368606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 368706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 368806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 368906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 369006656605SStefano Zampini } 369106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 369206656605SStefano Zampini } 3693ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 3694ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 369506656605SStefano Zampini } 3696d62866d3SStefano Zampini /* free memory */ 369788ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 369806656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 369906656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 370006656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 370106656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 3702d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3703d62866d3SStefano Zampini if (n_vertices) { 3704d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 3705d62866d3SStefano Zampini } 3706d62866d3SStefano Zampini if (n_constraints) { 3707d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 3708d62866d3SStefano Zampini } 370988ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 371088ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 371188ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 3712d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 371388ebb749SStefano Zampini Mat coarse_sub_mat; 371425084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 371588ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 371688ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 371788ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 37188bec7fa6SStefano Zampini Mat C_B,CPHI; 37198bec7fa6SStefano Zampini IS is_dummy; 37208bec7fa6SStefano Zampini Vec mones; 372188ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 372288ebb749SStefano Zampini PetscReal real_value; 372388ebb749SStefano Zampini 3724a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 3725a3df083aSStefano Zampini Mat A; 3726a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 3727a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 3728a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 3729a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 3730a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 3731a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 3732a3df083aSStefano Zampini } else { 373388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 373488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 373588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 373688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 3737a3df083aSStefano Zampini } 373888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 373988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 3740ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 374188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 374288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 374388ebb749SStefano Zampini } 374488ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 374588ebb749SStefano Zampini 374625084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 37473301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 374825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3749ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 375088ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 375188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 375288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 375388ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 375488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 375588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 375688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 375788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 375888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 375988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 376088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 376188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 376288ebb749SStefano Zampini } else { 376388ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 376488ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 376588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 376688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 376788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 376888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 376988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 377088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 377188ebb749SStefano Zampini } 377288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 377388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 377488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 3775511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 37764f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3777fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 3778d12edf2fSStefano Zampini PetscScalar *data,*data2; 37794f1b2e48SStefano Zampini PetscInt j; 3780d12edf2fSStefano Zampini 37814f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 3782fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 3783d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 378486c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 3785d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 3786d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 37874f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 37884f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 3789d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 37904f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 37914f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 37924f1b2e48SStefano Zampini } 3793d12edf2fSStefano Zampini } 3794d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 3795d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 3796d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 3797d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3798d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 3799d12edf2fSStefano Zampini } 3800d12edf2fSStefano Zampini #if 0 3801d12edf2fSStefano Zampini { 3802d12edf2fSStefano Zampini PetscViewer viewer; 3803d12edf2fSStefano Zampini char filename[256]; 3804ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 3805d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 3806d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3807ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 3808ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 3809ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 3810d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 381172b8c272SStefano Zampini if (save_change) { 381272b8c272SStefano Zampini Mat phi_B; 381372b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 381472b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 381572b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 381672b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 381772b8c272SStefano Zampini } else { 3818ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 3819ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 382072b8c272SStefano Zampini } 3821ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 3822ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 3823ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 3824ffd830a3SStefano Zampini } 3825ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 3826ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 3827ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 3828ffd830a3SStefano Zampini } 382972b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 3830ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 3831ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 3832ffd830a3SStefano Zampini } 3833d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 3834d12edf2fSStefano Zampini } 3835d12edf2fSStefano Zampini #endif 383681d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 38378bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 38381575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 383906656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 38408bec7fa6SStefano Zampini 38418bec7fa6SStefano Zampini /* check constraints */ 3842a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 3843a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 38444f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 38458bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 3846a00504b5SStefano Zampini } else { 3847a00504b5SStefano Zampini PetscScalar *data; 3848a00504b5SStefano Zampini Mat tmat; 3849a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 3850a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 3851a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 3852a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 3853a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3854a00504b5SStefano Zampini } 38558bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 38568bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 38578bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 38588bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 3859bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 3860ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 3861bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 3862bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 3863bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 3864bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 3865bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 386688ebb749SStefano Zampini } 38678bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 38688bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 38698bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 38708bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 387125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 387288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 387388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 387488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 387588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 387688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 387788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 387888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 387988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 388088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 388188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 3882ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 388388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 388488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 388588ebb749SStefano Zampini } 388688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 388788ebb749SStefano Zampini } 38888629588bSStefano Zampini /* get back data */ 38898629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 389088ebb749SStefano Zampini PetscFunctionReturn(0); 389188ebb749SStefano Zampini } 389288ebb749SStefano Zampini 389388ebb749SStefano Zampini #undef __FUNCT__ 3894d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 3895d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 3896aa0d41d4SStefano Zampini { 3897d65f70fdSStefano Zampini Mat *work_mat; 3898d65f70fdSStefano Zampini IS isrow_s,iscol_s; 3899d65f70fdSStefano Zampini PetscBool rsorted,csorted; 3900c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 3901aa0d41d4SStefano Zampini PetscErrorCode ierr; 3902aa0d41d4SStefano Zampini 3903aa0d41d4SStefano Zampini PetscFunctionBegin; 3904d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 3905d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 3906d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 3907d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3908aa0d41d4SStefano Zampini 3909d65f70fdSStefano Zampini if (!rsorted) { 3910906d46d4SStefano Zampini const PetscInt *idxs; 3911906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 3912aa0d41d4SStefano Zampini 3913d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 3914d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 3915d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 3916d65f70fdSStefano Zampini idxs_perm_r[i] = i; 3917aa0d41d4SStefano Zampini } 3918d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 3919d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 3920d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 3921d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 3922aa0d41d4SStefano Zampini } 3923d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 3924d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 3925d65f70fdSStefano Zampini } else { 3926d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 3927d65f70fdSStefano Zampini isrow_s = isrow; 3928aa0d41d4SStefano Zampini } 3929906d46d4SStefano Zampini 3930d65f70fdSStefano Zampini if (!csorted) { 3931d65f70fdSStefano Zampini if (isrow == iscol) { 3932d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 3933d65f70fdSStefano Zampini iscol_s = isrow_s; 3934d65f70fdSStefano Zampini } else { 3935d65f70fdSStefano Zampini const PetscInt *idxs; 3936d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 3937906d46d4SStefano Zampini 3938d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 3939d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 3940d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 3941d65f70fdSStefano Zampini idxs_perm_c[i] = i; 3942d65f70fdSStefano Zampini } 3943d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 3944d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 3945d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 3946d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 3947d65f70fdSStefano Zampini } 3948d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 3949d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 3950d65f70fdSStefano Zampini } 3951d65f70fdSStefano Zampini } else { 3952d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 3953d65f70fdSStefano Zampini iscol_s = iscol; 3954d65f70fdSStefano Zampini } 3955d65f70fdSStefano Zampini 3956d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 3957d65f70fdSStefano Zampini 3958d65f70fdSStefano Zampini if (!rsorted || !csorted) { 3959906d46d4SStefano Zampini Mat new_mat; 3960d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 3961906d46d4SStefano Zampini 3962d65f70fdSStefano Zampini if (!rsorted) { 3963d65f70fdSStefano Zampini PetscInt *idxs_r,i; 3964d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 3965d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 3966d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 3967906d46d4SStefano Zampini } 3968d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 3969d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 3970d65f70fdSStefano Zampini } else { 3971d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 3972906d46d4SStefano Zampini } 3973d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 3974d65f70fdSStefano Zampini 3975d65f70fdSStefano Zampini if (!csorted) { 3976d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 3977d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 3978d65f70fdSStefano Zampini is_perm_c = is_perm_r; 3979d65f70fdSStefano Zampini } else { 3980d65f70fdSStefano Zampini PetscInt *idxs_c,i; 3981f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 3982d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 3983d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 3984d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 3985d65f70fdSStefano Zampini } 3986d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 3987d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 3988d65f70fdSStefano Zampini } 3989d65f70fdSStefano Zampini } else { 3990d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 3991d65f70fdSStefano Zampini } 3992d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 3993d65f70fdSStefano Zampini 3994d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 3995d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 3996d65f70fdSStefano Zampini work_mat[0] = new_mat; 3997d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 3998d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 3999d65f70fdSStefano Zampini } 4000d65f70fdSStefano Zampini 4001d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 4002d65f70fdSStefano Zampini *B = work_mat[0]; 4003d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 4004d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 4005d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 4006d65f70fdSStefano Zampini PetscFunctionReturn(0); 4007d65f70fdSStefano Zampini } 4008d65f70fdSStefano Zampini 4009d65f70fdSStefano Zampini #undef __FUNCT__ 40105e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 40115e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 4012aa0d41d4SStefano Zampini { 4013aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 40145e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4015d65f70fdSStefano Zampini Mat new_mat; 40165e8657edSStefano Zampini IS is_local,is_global; 4017d65f70fdSStefano Zampini PetscInt local_size; 4018d65f70fdSStefano Zampini PetscBool isseqaij; 4019aa0d41d4SStefano Zampini PetscErrorCode ierr; 4020aa0d41d4SStefano Zampini 4021aa0d41d4SStefano Zampini PetscFunctionBegin; 4022aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 40235e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 40245e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 4025b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 4026aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 4027d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 4028aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 4029906d46d4SStefano Zampini 4030906d46d4SStefano Zampini /* check */ 4031906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4032906d46d4SStefano Zampini Vec x,x_change; 4033906d46d4SStefano Zampini PetscReal error; 4034906d46d4SStefano Zampini 40355e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 4036906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 40375e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 4038e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4039e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4040d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 4041*88428137SStefano Zampini if (!pcbddc->change_interior) { 4042*88428137SStefano Zampini const PetscScalar *x,*y,*v; 4043*88428137SStefano Zampini PetscReal lerror = 0.; 4044*88428137SStefano Zampini PetscInt i; 4045*88428137SStefano Zampini 4046*88428137SStefano Zampini ierr = VecGetArrayRead(matis->x,&x);CHKERRQ(ierr); 4047*88428137SStefano Zampini ierr = VecGetArrayRead(matis->y,&y);CHKERRQ(ierr); 4048*88428137SStefano Zampini ierr = VecGetArrayRead(matis->counter,&v);CHKERRQ(ierr); 4049*88428137SStefano Zampini for (i=0;i<local_size;i++) 4050*88428137SStefano Zampini if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror) 4051*88428137SStefano Zampini lerror = PetscAbsScalar(x[i]-y[i]); 4052*88428137SStefano Zampini ierr = VecRestoreArrayRead(matis->x,&x);CHKERRQ(ierr); 4053*88428137SStefano Zampini ierr = VecRestoreArrayRead(matis->y,&y);CHKERRQ(ierr); 4054*88428137SStefano Zampini ierr = VecRestoreArrayRead(matis->counter,&v);CHKERRQ(ierr); 4055*88428137SStefano Zampini ierr = MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4056*88428137SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on I: %1.6e\n",error);CHKERRQ(ierr); 4057*88428137SStefano Zampini } 4058e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4059e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4060906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4061906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4062906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4063906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 4064906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4065906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4066906d46d4SStefano Zampini } 4067906d46d4SStefano Zampini 406822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 40699b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 407022d5777bSStefano Zampini if (isseqaij) { 4071a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4072a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 4073aa0d41d4SStefano Zampini } else { 4074a00504b5SStefano Zampini Mat work_mat; 40751cf9b237SStefano Zampini 4076a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4077aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4078a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 40791d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 4080aa0d41d4SStefano Zampini } 40813301b35fSStefano Zampini if (matis->A->symmetric_set) { 40823301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 4083e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 40843301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 4085e496cd5dSStefano Zampini #endif 40863301b35fSStefano Zampini } 4087d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 4088aa0d41d4SStefano Zampini PetscFunctionReturn(0); 4089aa0d41d4SStefano Zampini } 4090aa0d41d4SStefano Zampini 4091aa0d41d4SStefano Zampini #undef __FUNCT__ 4092a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 40938ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 4094a64d13efSStefano Zampini { 4095a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4096a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4097d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 409853892102SStefano Zampini PetscInt *idx_R_local=NULL; 40993a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 41003a50541eSStefano Zampini PetscInt vbs,bs; 41016816873aSStefano Zampini PetscBT bitmask=NULL; 4102a64d13efSStefano Zampini PetscErrorCode ierr; 4103a64d13efSStefano Zampini 4104a64d13efSStefano Zampini PetscFunctionBegin; 4105b23d619eSStefano Zampini /* 4106b23d619eSStefano Zampini No need to setup local scatters if 4107b23d619eSStefano Zampini - primal space is unchanged 4108b23d619eSStefano Zampini AND 4109b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 4110b23d619eSStefano Zampini AND 4111b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 4112b23d619eSStefano Zampini */ 4113b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 4114f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 4115f4ddd8eeSStefano Zampini } 4116f4ddd8eeSStefano Zampini /* destroy old objects */ 4117f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 4118f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 4119f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 4120a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 4121b371cd4fSStefano Zampini n_B = pcis->n_B; 4122b371cd4fSStefano Zampini n_D = pcis->n - n_B; 4123b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 41243a50541eSStefano Zampini 4125a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 41266816873aSStefano Zampini 412753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 4128b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4129854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 4130a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 4131a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 41320e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 4133a64d13efSStefano Zampini } 4134a64d13efSStefano Zampini 4135a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 41364641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 41376816873aSStefano Zampini idx_R_local[n_R++] = i; 4138a64d13efSStefano Zampini } 4139a64d13efSStefano Zampini } 4140df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 4141df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 41426816873aSStefano Zampini 4143df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4144df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 41456816873aSStefano Zampini } 41463a50541eSStefano Zampini 41473a50541eSStefano Zampini /* Block code */ 41483a50541eSStefano Zampini vbs = 1; 41493a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 41503a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 41513a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 41523a50541eSStefano Zampini PetscInt *vary; 4153b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4154785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 41553a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 4156d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 4157d3df7717SStefano 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 */ 41580e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 4159d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 41603a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 41613a50541eSStefano Zampini is_blocked = PETSC_FALSE; 41623a50541eSStefano Zampini break; 41633a50541eSStefano Zampini } 41643a50541eSStefano Zampini } 4165d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 4166d3df7717SStefano Zampini } else { 4167d3df7717SStefano Zampini /* Verify directly the R set */ 4168d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 4169d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 4170d3df7717SStefano Zampini for (j=1; j<bs; j++) { 4171d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 4172d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 4173d3df7717SStefano Zampini break; 4174d3df7717SStefano Zampini } 4175d3df7717SStefano Zampini } 4176d3df7717SStefano Zampini } 4177d3df7717SStefano Zampini } 41783a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 41793a50541eSStefano Zampini vbs = bs; 41803a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 41813a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 41823a50541eSStefano Zampini } 41833a50541eSStefano Zampini } 41843a50541eSStefano Zampini } 41853a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 4186b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4187df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 418853892102SStefano Zampini 4189df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4190df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 419153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 4192df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 419353892102SStefano Zampini } else { 41943a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 419553892102SStefano Zampini } 4196a64d13efSStefano Zampini 4197a64d13efSStefano Zampini /* print some info if requested */ 4198a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 4199a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4200a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42011575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4202a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 4203a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 42044f1b2e48SStefano 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); 4205a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4206a64d13efSStefano Zampini } 4207a64d13efSStefano Zampini 4208a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 4209b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 42106816873aSStefano Zampini IS is_aux1,is_aux2; 42116816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 42126816873aSStefano Zampini 42133a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4214854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 4215854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 4216a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 42174641a718SStefano Zampini for (i=0; i<n_D; i++) { 42184641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 42194641a718SStefano Zampini } 4220a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4221a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 42224641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 42234641a718SStefano Zampini aux_array1[j++] = i; 4224a64d13efSStefano Zampini } 4225a64d13efSStefano Zampini } 4226a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4227a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4228a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 42294641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 42304641a718SStefano Zampini aux_array2[j++] = i; 4231a64d13efSStefano Zampini } 4232a64d13efSStefano Zampini } 4233a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4234a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 4235a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 4236a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4237a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 4238a64d13efSStefano Zampini 42398eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4240785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 4241a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 42424641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 42434641a718SStefano Zampini aux_array1[j++] = i; 4244a64d13efSStefano Zampini } 4245a64d13efSStefano Zampini } 4246a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4247a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 4248a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4249a64d13efSStefano Zampini } 42504641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 42513a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4252d62866d3SStefano Zampini } else { 4253df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 42546816873aSStefano Zampini IS tis; 42556816873aSStefano Zampini PetscInt schur_size; 42566816873aSStefano Zampini 4257df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 42586816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 4259df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 42606816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 42616816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 42626816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 42636816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 42646816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 4265d62866d3SStefano Zampini } 4266d62866d3SStefano Zampini } 4267a64d13efSStefano Zampini PetscFunctionReturn(0); 4268a64d13efSStefano Zampini } 4269a64d13efSStefano Zampini 4270304d26faSStefano Zampini 4271304d26faSStefano Zampini #undef __FUNCT__ 4272304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 4273684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 4274304d26faSStefano Zampini { 4275304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4276304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4277304d26faSStefano Zampini PC pc_temp; 4278304d26faSStefano Zampini Mat A_RR; 4279f4ddd8eeSStefano Zampini MatReuse reuse; 4280304d26faSStefano Zampini PetscScalar m_one = -1.0; 4281304d26faSStefano Zampini PetscReal value; 428204708bb6SStefano Zampini PetscInt n_D,n_R; 4283c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 4284304d26faSStefano Zampini PetscErrorCode ierr; 4285e604994aSStefano Zampini /* prefixes stuff */ 4286312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 4287e604994aSStefano Zampini size_t len; 4288304d26faSStefano Zampini 4289304d26faSStefano Zampini PetscFunctionBegin; 4290304d26faSStefano Zampini 4291e604994aSStefano Zampini /* compute prefixes */ 4292e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 4293e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 4294e604994aSStefano Zampini if (!pcbddc->current_level) { 4295e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4296e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4297e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4298e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4299e604994aSStefano Zampini } else { 4300e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4301312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4302e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4303e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 4304312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4305312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 430634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 430734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4308e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4309e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4310e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 4311e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 4312e604994aSStefano Zampini } 4313e604994aSStefano Zampini 4314304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 4315684f6988SStefano Zampini if (dirichlet) { 4316d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4317450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 43189a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 4319450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 4320a3df083aSStefano Zampini Mat A_IIn; 4321a3df083aSStefano Zampini 4322a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 4323a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 4324a3df083aSStefano Zampini pcis->A_II = A_IIn; 4325a3df083aSStefano Zampini } 4326450f8f5eSStefano Zampini } 43273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 43283301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 4329964fefecSStefano Zampini } 4330ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 4331964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 4332304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 4333304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 4334304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 4335304d26faSStefano Zampini /* default */ 4336304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 4337e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 43389577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 4339304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 43409577ea80SStefano Zampini if (issbaij) { 43419577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 43429577ea80SStefano Zampini } else { 4343304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 43449577ea80SStefano Zampini } 4345304d26faSStefano Zampini /* Allow user's customization */ 4346304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 4347304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4348304d26faSStefano Zampini } 4349d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 4350b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4351df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4352d62866d3SStefano Zampini 4353df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 4354d5574798SStefano Zampini } 4355304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4356304d26faSStefano Zampini if (!n_D) { 4357304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 4358304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4359304d26faSStefano Zampini } 4360304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 4361304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 4362304d26faSStefano Zampini /* set ksp_D into pcis data */ 4363304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 4364304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 4365304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 4366684f6988SStefano Zampini } 4367304d26faSStefano Zampini 4368304d26faSStefano Zampini /* NEUMANN PROBLEM */ 4369684f6988SStefano Zampini A_RR = 0; 4370684f6988SStefano Zampini if (neumann) { 4371d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 437204708bb6SStefano Zampini PetscInt ibs,mbs; 437304708bb6SStefano Zampini PetscBool issbaij; 437404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4375f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 43768ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 4377f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 4378f4ddd8eeSStefano Zampini PetscInt nn_R; 437981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 4380f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4381f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 4382f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 4383f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 4384f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4385f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4386f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 4387727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 4388f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4389f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4390f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 4391f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 4392f4ddd8eeSStefano Zampini } 4393f4ddd8eeSStefano Zampini } 4394f4ddd8eeSStefano Zampini /* last check */ 4395d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 4396f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4397f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4398f4ddd8eeSStefano Zampini } 4399f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 4400f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4401f4ddd8eeSStefano Zampini } 4402a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 4403af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 4404af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 440504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 440604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 440704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 440804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 440904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4410af732b37SStefano Zampini } else { 4411511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 44126816873aSStefano Zampini } 441304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 441404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 441504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 441604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 441704708bb6SStefano Zampini } else { 4418511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 441904708bb6SStefano Zampini } 442004708bb6SStefano Zampini } 4421a00504b5SStefano Zampini /* extract A_RR */ 4422b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4423a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4424a00504b5SStefano Zampini 4425a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 442616e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4427a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 442816e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 442916e386b8SStefano Zampini } else { 4430a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 4431a00504b5SStefano Zampini } 4432a00504b5SStefano Zampini } else { 4433a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4434a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 4435a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4436a00504b5SStefano Zampini } 4437a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 4438f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 443916e386b8SStefano Zampini } 44403301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 44413301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 44426816873aSStefano Zampini } 4443f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 4444304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 4445304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 4446304d26faSStefano Zampini /* default */ 4447304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 4448e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 4449304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 44509577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 44519577ea80SStefano Zampini if (issbaij) { 44529577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 44539577ea80SStefano Zampini } else { 4454304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 44559577ea80SStefano Zampini } 4456304d26faSStefano Zampini /* Allow user's customization */ 4457304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 4458304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4459304d26faSStefano Zampini } 4460304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4461304d26faSStefano Zampini if (!n_R) { 4462304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 4463304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4464304d26faSStefano Zampini } 44655cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 4466df4d28bfSStefano Zampini /* Reuse solver if it is present */ 4467b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4468df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4469d62866d3SStefano Zampini 4470df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 4471d62866d3SStefano Zampini } 4472304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 4473304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 4474684f6988SStefano Zampini } 4475304d26faSStefano Zampini 4476684f6988SStefano Zampini if (pcbddc->dbg_flag) { 4477684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44781575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4479684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4480684f6988SStefano Zampini } 4481c7017625SStefano Zampini 4482c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 4483c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 4484c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 4485c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 4486c7017625SStefano Zampini } 4487c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 4488c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 4489c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 4490c7017625SStefano Zampini } 4491c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 4492c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 4493c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 4494c7017625SStefano Zampini } 4495c7017625SStefano Zampini 4496c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 4497c7017625SStefano Zampini if (pcbddc->dbg_flag) { 4498684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 44990fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 45000fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 45010fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 45020fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 45030fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 4504e604994aSStefano 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); 4505c7017625SStefano Zampini if (check_corr[0]) { 4506c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 4507c7017625SStefano Zampini } 4508304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4509304d26faSStefano Zampini } 4510684f6988SStefano Zampini if (neumann) { /* Neumann */ 45110fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 45120fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 45130fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 45140fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 45150fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 4516e604994aSStefano 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); 4517c7017625SStefano Zampini if (check_corr[1]) { 4518c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 4519c7017625SStefano Zampini } 4520304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4521304d26faSStefano Zampini } 4522684f6988SStefano Zampini } 45235cbda25cSStefano Zampini /* free Neumann problem's matrix */ 45245cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4525304d26faSStefano Zampini PetscFunctionReturn(0); 4526304d26faSStefano Zampini } 4527304d26faSStefano Zampini 4528304d26faSStefano Zampini #undef __FUNCT__ 4529ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 453080677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 4531674ae819SStefano Zampini { 4532674ae819SStefano Zampini PetscErrorCode ierr; 4533674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4534be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4535b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 4536674ae819SStefano Zampini 4537674ae819SStefano Zampini PetscFunctionBegin; 4538b334f244SStefano Zampini if (!reuse_solver) { 453980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 454020c7b377SStefano Zampini } 454180677318SStefano Zampini if (!pcbddc->switch_static) { 454280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 454380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 454480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 454520c7b377SStefano Zampini } 4546b334f244SStefano Zampini if (!reuse_solver) { 454780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 454880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 454920c7b377SStefano Zampini } else { 4550df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4551be83ff47SStefano Zampini 4552df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4553df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 455420c7b377SStefano Zampini } 4555be83ff47SStefano Zampini } else { 455680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 455780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 455880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 455980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 456080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 456180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 456280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 456380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 456480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4565674ae819SStefano Zampini } 4566674ae819SStefano Zampini } 4567b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 456880677318SStefano Zampini if (applytranspose) { 456980677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 457080677318SStefano Zampini } else { 457180677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 457280677318SStefano Zampini } 4573be83ff47SStefano Zampini } else { 4574df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4575be83ff47SStefano Zampini 4576be83ff47SStefano Zampini if (applytranspose) { 4577df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4578be83ff47SStefano Zampini } else { 4579df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4580be83ff47SStefano Zampini } 4581be83ff47SStefano Zampini } 458280677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 458380677318SStefano Zampini if (!pcbddc->switch_static) { 4584b334f244SStefano Zampini if (!reuse_solver) { 458580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 458680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4587be83ff47SStefano Zampini } else { 4588df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4589be83ff47SStefano Zampini 4590df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4591df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4592be83ff47SStefano Zampini } 459380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 459480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 459580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 459680677318SStefano Zampini } 459780677318SStefano Zampini } else { 459880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 459980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 460080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 460180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 460280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 460380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 460480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 460580677318SStefano Zampini } 460680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 460780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 460880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 460980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4610674ae819SStefano Zampini } 4611674ae819SStefano Zampini PetscFunctionReturn(0); 4612674ae819SStefano Zampini } 4613674ae819SStefano Zampini 4614dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 4615674ae819SStefano Zampini #undef __FUNCT__ 4616674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 4617dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 4618674ae819SStefano Zampini { 4619674ae819SStefano Zampini PetscErrorCode ierr; 4620674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4621674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 4622674ae819SStefano Zampini const PetscScalar zero = 0.0; 4623674ae819SStefano Zampini 4624674ae819SStefano Zampini PetscFunctionBegin; 4625dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 46264fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 4627dc359a40SStefano Zampini if (applytranspose) { 4628674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 46298eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 4630dc359a40SStefano Zampini } else { 4631674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 4632674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 463315aaf578SStefano Zampini } 46344fee134fSStefano Zampini } else { 46354fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 46364fee134fSStefano Zampini } 4637efc2fbd9SStefano Zampini 4638efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 46394f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4640efc2fbd9SStefano Zampini PetscScalar *array; 46414f1b2e48SStefano Zampini PetscInt j; 4642efc2fbd9SStefano Zampini 4643efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 46444f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 4645efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 4646efc2fbd9SStefano Zampini } 4647efc2fbd9SStefano Zampini 464812edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 464912edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 465012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 465112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 465212edc857SStefano Zampini 46539f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 465412edc857SStefano Zampini if (pcbddc->coarse_ksp) { 465551694757SStefano Zampini Mat coarse_mat; 4656964fefecSStefano Zampini Vec rhs,sol; 465751694757SStefano Zampini MatNullSpace nullsp; 465827b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 4659964fefecSStefano Zampini 466027b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 466127b6a85dSStefano Zampini PC coarse_pc; 466227b6a85dSStefano Zampini 466327b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 466427b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 466527b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 466627b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 466727b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 466827b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 46693bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 467027b6a85dSStefano Zampini } 467127b6a85dSStefano Zampini } 4672964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 4673964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 467451694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 467551694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 467651694757SStefano Zampini if (nullsp) { 467751694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 467851694757SStefano Zampini } 467912edc857SStefano Zampini if (applytranspose) { 46809a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 4681964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 46822701bc32SStefano Zampini } else { 46831f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 46842701bc32SStefano Zampini PC coarse_pc; 46852701bc32SStefano Zampini 46862701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 46872701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 46883e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 46892701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 469012edc857SStefano Zampini } else { 4691964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 469212edc857SStefano Zampini } 46932701bc32SStefano Zampini } 46941d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 469527b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 469627b6a85dSStefano Zampini PC coarse_pc; 469727b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 469827b6a85dSStefano Zampini 469927b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 470027b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 470127b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 47023bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 470327b6a85dSStefano Zampini } 470451694757SStefano Zampini if (nullsp) { 470551694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 470651694757SStefano Zampini } 470712edc857SStefano Zampini } 4708674ae819SStefano Zampini 4709674ae819SStefano Zampini /* Local solution on R nodes */ 47104fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 471180677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 47129f00e9b4SStefano Zampini } 47139f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 47149f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 471512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4716674ae819SStefano Zampini 47174fee134fSStefano Zampini /* Sum contributions from the two levels */ 47184fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 4719dc359a40SStefano Zampini if (applytranspose) { 4720dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 4721dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 4722dc359a40SStefano Zampini } else { 4723674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 47248eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 4725dc359a40SStefano Zampini } 4726efc2fbd9SStefano Zampini /* store p0 */ 47274f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4728efc2fbd9SStefano Zampini PetscScalar *array; 47294f1b2e48SStefano Zampini PetscInt j; 4730efc2fbd9SStefano Zampini 4731efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 47324f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 4733efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 4734efc2fbd9SStefano Zampini } 47354fee134fSStefano Zampini } else { /* expand the coarse solution */ 47364fee134fSStefano Zampini if (applytranspose) { 47374fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 47384fee134fSStefano Zampini } else { 47394fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 47404fee134fSStefano Zampini } 47414fee134fSStefano Zampini } 4742674ae819SStefano Zampini PetscFunctionReturn(0); 4743674ae819SStefano Zampini } 4744674ae819SStefano Zampini 4745674ae819SStefano Zampini #undef __FUNCT__ 4746674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 474712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 4748674ae819SStefano Zampini { 4749674ae819SStefano Zampini PetscErrorCode ierr; 4750674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 475158da7f69SStefano Zampini PetscScalar *array; 475212edc857SStefano Zampini Vec from,to; 4753674ae819SStefano Zampini 4754674ae819SStefano Zampini PetscFunctionBegin; 475512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 475612edc857SStefano Zampini from = pcbddc->coarse_vec; 475712edc857SStefano Zampini to = pcbddc->vec1_P; 475812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 475912edc857SStefano Zampini Vec tvec; 476058da7f69SStefano Zampini 476158da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 476258da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 476312edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 476458da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 476558da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 476658da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 476712edc857SStefano Zampini } 476812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 476912edc857SStefano Zampini from = pcbddc->vec1_P; 477012edc857SStefano Zampini to = pcbddc->coarse_vec; 477112edc857SStefano Zampini } 477212edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 4773674ae819SStefano Zampini PetscFunctionReturn(0); 4774674ae819SStefano Zampini } 4775674ae819SStefano Zampini 4776674ae819SStefano Zampini #undef __FUNCT__ 4777674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 477812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 4779674ae819SStefano Zampini { 4780674ae819SStefano Zampini PetscErrorCode ierr; 4781674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 478258da7f69SStefano Zampini PetscScalar *array; 478312edc857SStefano Zampini Vec from,to; 4784674ae819SStefano Zampini 4785674ae819SStefano Zampini PetscFunctionBegin; 478612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 478712edc857SStefano Zampini from = pcbddc->coarse_vec; 478812edc857SStefano Zampini to = pcbddc->vec1_P; 478912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 479012edc857SStefano Zampini from = pcbddc->vec1_P; 479112edc857SStefano Zampini to = pcbddc->coarse_vec; 479212edc857SStefano Zampini } 479312edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 479412edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 479512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 479612edc857SStefano Zampini Vec tvec; 479758da7f69SStefano Zampini 479812edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 479958da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 480058da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 480158da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 480258da7f69SStefano Zampini } 480358da7f69SStefano Zampini } else { 480458da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 480558da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 480612edc857SStefano Zampini } 480712edc857SStefano Zampini } 4808674ae819SStefano Zampini PetscFunctionReturn(0); 4809674ae819SStefano Zampini } 4810674ae819SStefano Zampini 4811984c4197SStefano Zampini /* uncomment for testing purposes */ 4812984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 4813674ae819SStefano Zampini #undef __FUNCT__ 4814674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 4815674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 4816674ae819SStefano Zampini { 4817674ae819SStefano Zampini PetscErrorCode ierr; 4818674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4819674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4820674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4821984c4197SStefano Zampini /* one and zero */ 4822984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 4823984c4197SStefano Zampini /* space to store constraints and their local indices */ 48249162d606SStefano Zampini PetscScalar *constraints_data; 48259162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 48269162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 48279162d606SStefano Zampini PetscInt *constraints_n; 4828984c4197SStefano Zampini /* iterators */ 4829b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 4830984c4197SStefano Zampini /* BLAS integers */ 4831e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 4832e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 4833c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 4834727cdba6SStefano Zampini /* reuse */ 48350e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 48360e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 4837984c4197SStefano Zampini /* change of basis */ 4838b3d85658SStefano Zampini PetscBool qr_needed; 48399162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 4840984c4197SStefano Zampini /* auxiliary stuff */ 484164efe560SStefano Zampini PetscInt *nnz,*is_indices; 48428a0068c3SStefano Zampini PetscInt ncc; 4843984c4197SStefano Zampini /* some quantities */ 484445a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 4845a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 4846984c4197SStefano Zampini 4847674ae819SStefano Zampini PetscFunctionBegin; 48488e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 48498e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 48508e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 485116909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 4852088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 4853088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 48540e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 48550e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 48560e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 48570e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 48580e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 4859088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4860cf5a6209SStefano Zampini 4861cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 48629162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 4863cf5a6209SStefano Zampini MatNullSpace nearnullsp; 4864cf5a6209SStefano Zampini const Vec *nearnullvecs; 4865cf5a6209SStefano Zampini Vec *localnearnullsp; 4866cf5a6209SStefano Zampini PetscScalar *array; 4867cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 4868cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 4869674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 4870b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 4871674ae819SStefano Zampini PetscScalar *work; 4872674ae819SStefano Zampini PetscReal *singular_vals; 4873674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4874674ae819SStefano Zampini PetscReal *rwork; 4875674ae819SStefano Zampini #endif 4876674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4877674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 4878674ae819SStefano Zampini #else 4879964fefecSStefano Zampini PetscBLASInt dummy_int=1; 4880964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 4881674ae819SStefano Zampini #endif 4882674ae819SStefano Zampini 4883674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 4884d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 4885e4d548c7SStefano Zampini /* print some info */ 48865c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 4887e4d548c7SStefano Zampini PetscInt nv; 4888e4d548c7SStefano Zampini 4889c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4890e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 4891e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4892e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4893e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 4894e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 4895e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 4896e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4897e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4898e4d548c7SStefano Zampini } 4899e4d548c7SStefano Zampini 4900d06fc5fdSStefano Zampini /* free unneeded index sets */ 4901d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 4902d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 4903674ae819SStefano Zampini } 4904d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 4905d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 4906d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 4907d06fc5fdSStefano Zampini } 4908d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 4909d06fc5fdSStefano Zampini n_ISForEdges = 0; 4910d06fc5fdSStefano Zampini } 4911d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 4912d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 4913d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 4914d06fc5fdSStefano Zampini } 4915d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 4916d06fc5fdSStefano Zampini n_ISForFaces = 0; 4917d06fc5fdSStefano Zampini } 491870022509SStefano Zampini 4919674ae819SStefano Zampini /* check if near null space is attached to global mat */ 4920674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 4921674ae819SStefano Zampini if (nearnullsp) { 4922674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 4923f4ddd8eeSStefano Zampini /* remove any stored info */ 4924f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 4925f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 4926f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 4927f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 4928f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 4929473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 4930f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 4931f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 4932f4ddd8eeSStefano Zampini } 4933984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 4934984c4197SStefano Zampini nnsp_size = 0; 4935674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 4936674ae819SStefano Zampini } 4937984c4197SStefano Zampini /* get max number of constraints on a single cc */ 4938984c4197SStefano Zampini max_constraints = nnsp_size; 4939984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 4940984c4197SStefano Zampini 4941674ae819SStefano Zampini /* 4942674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 49439162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 49449162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 49459162d606SStefano Zampini There can be multiple constraints per connected component 4946674ae819SStefano Zampini */ 4947674ae819SStefano Zampini n_vertices = 0; 4948674ae819SStefano Zampini if (ISForVertices) { 4949674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 4950674ae819SStefano Zampini } 49519162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 49529162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 49539162d606SStefano Zampini 49549162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 49559162d606SStefano Zampini total_counts *= max_constraints; 4956674ae819SStefano Zampini total_counts += n_vertices; 49574641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 49589162d606SStefano Zampini 4959674ae819SStefano Zampini total_counts = 0; 4960674ae819SStefano Zampini max_size_of_constraint = 0; 4961674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 49629162d606SStefano Zampini IS used_is; 4963674ae819SStefano Zampini if (i<n_ISForEdges) { 49649162d606SStefano Zampini used_is = ISForEdges[i]; 4965674ae819SStefano Zampini } else { 49669162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 4967674ae819SStefano Zampini } 49689162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 4969674ae819SStefano Zampini total_counts += j; 4970674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 4971674ae819SStefano Zampini } 49729162d606SStefano 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); 49739162d606SStefano Zampini 4974984c4197SStefano Zampini /* get local part of global near null space vectors */ 4975785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 4976984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 4977984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 4978e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4979e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4980984c4197SStefano Zampini } 4981674ae819SStefano Zampini 4982242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 4983242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 4984a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 4985242a89d7SStefano Zampini 4986984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 4987a773dcb8SStefano Zampini if (!skip_lapack) { 4988674ae819SStefano Zampini PetscScalar temp_work; 4989911cabfeSStefano Zampini 4990674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 4991984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 4992785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 4993785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 4994785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 4995674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 4996785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 4997674ae819SStefano Zampini #endif 4998674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 4999c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 5000c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 5001674ae819SStefano Zampini lwork = -1; 5002674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5003674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5004c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 5005674ae819SStefano Zampini #else 5006c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 5007674ae819SStefano Zampini #endif 5008674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5009984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 5010674ae819SStefano Zampini #else /* on missing GESVD */ 5011674ae819SStefano Zampini /* SVD */ 5012674ae819SStefano Zampini PetscInt max_n,min_n; 5013674ae819SStefano Zampini max_n = max_size_of_constraint; 5014984c4197SStefano Zampini min_n = max_constraints; 5015984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 5016674ae819SStefano Zampini min_n = max_size_of_constraint; 5017984c4197SStefano Zampini max_n = max_constraints; 5018674ae819SStefano Zampini } 5019785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 5020674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5021785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 5022674ae819SStefano Zampini #endif 5023674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5024674ae819SStefano Zampini lwork = -1; 5025e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 5026e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 5027b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 5028674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5029674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 50309162d606SStefano 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)); 5031674ae819SStefano Zampini #else 50329162d606SStefano 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)); 5033674ae819SStefano Zampini #endif 5034674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5035984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 5036984c4197SStefano Zampini #endif /* on missing GESVD */ 5037674ae819SStefano Zampini /* Allocate optimal workspace */ 5038674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 5039854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 5040674ae819SStefano Zampini } 5041674ae819SStefano Zampini /* Now we can loop on constraining sets */ 5042674ae819SStefano Zampini total_counts = 0; 50439162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 50449162d606SStefano Zampini constraints_data_ptr[0] = 0; 5045674ae819SStefano Zampini /* vertices */ 50469162d606SStefano Zampini if (n_vertices) { 5047674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 50489162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 5049674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 50509162d606SStefano Zampini constraints_n[total_counts] = 1; 50519162d606SStefano Zampini constraints_data[total_counts] = 1.0; 50529162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 50539162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 5054674ae819SStefano Zampini total_counts++; 5055674ae819SStefano Zampini } 5056674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5057674ae819SStefano Zampini n_vertices = total_counts; 5058674ae819SStefano Zampini } 5059984c4197SStefano Zampini 5060674ae819SStefano Zampini /* edges and faces */ 50619162d606SStefano Zampini total_counts_cc = total_counts; 5062911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 50639162d606SStefano Zampini IS used_is; 50649162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 50659162d606SStefano Zampini 5066911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 50679162d606SStefano Zampini used_is = ISForEdges[ncc]; 5068984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 5069674ae819SStefano Zampini } else { 50709162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 5071984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 5072674ae819SStefano Zampini } 5073674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 50749162d606SStefano Zampini 50759162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 50769162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5077984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 5078984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 5079674ae819SStefano Zampini if (nnsp_has_cnst) { 50805b08dc53SStefano Zampini PetscScalar quad_value; 50819162d606SStefano Zampini 50829162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 50839162d606SStefano Zampini idxs_copied = PETSC_TRUE; 50849162d606SStefano Zampini 5085a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 5086674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 5087a773dcb8SStefano Zampini } else { 5088a773dcb8SStefano Zampini quad_value = 1.0; 5089a773dcb8SStefano Zampini } 5090674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 50919162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 5092674ae819SStefano Zampini } 50939162d606SStefano Zampini temp_constraints++; 5094674ae819SStefano Zampini total_counts++; 5095674ae819SStefano Zampini } 5096674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 5097984c4197SStefano Zampini PetscReal real_value; 50989162d606SStefano Zampini PetscScalar *ptr_to_data; 50999162d606SStefano Zampini 5100984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 51019162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 5102674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 51039162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 5104674ae819SStefano Zampini } 5105984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 5106984c4197SStefano Zampini /* check if array is null on the connected component */ 5107e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 51089162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 51095b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 5110674ae819SStefano Zampini temp_constraints++; 5111674ae819SStefano Zampini total_counts++; 51129162d606SStefano Zampini if (!idxs_copied) { 51139162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 51149162d606SStefano Zampini idxs_copied = PETSC_TRUE; 5115674ae819SStefano Zampini } 5116674ae819SStefano Zampini } 51179162d606SStefano Zampini } 51189162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 511945a1bb75SStefano Zampini valid_constraints = temp_constraints; 5120eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 5121a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 51229162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 51239162d606SStefano Zampini 51249162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5125a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 51269162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 5127a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 51289162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 5129a773dcb8SStefano Zampini } else { /* perform SVD */ 5130984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 51319162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5132674ae819SStefano Zampini 5133674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5134984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 5135984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 5136984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 5137984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 5138984c4197SStefano Zampini from that computed using LAPACKgesvd 5139984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 5140984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 5141984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 5142674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 5143e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5144984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5145674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 5146674ae819SStefano Zampini for (k=0;k<j+1;k++) { 51479162d606SStefano 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)); 5148674ae819SStefano Zampini } 5149674ae819SStefano Zampini } 5150e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 5151e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5152e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 5153674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5154c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 5155674ae819SStefano Zampini #else 5156c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 5157674ae819SStefano Zampini #endif 5158674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5159984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 5160984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 5161674ae819SStefano Zampini j = 0; 5162984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 5163674ae819SStefano Zampini total_counts = total_counts-j; 516445a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 5165e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 5166c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5167c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5168c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 5169c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5170c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 5171c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5172674ae819SStefano Zampini if (j<temp_constraints) { 5173984c4197SStefano Zampini PetscInt ii; 5174984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 5175674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 51769162d606SStefano 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)); 5177674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5178984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 5179674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 51809162d606SStefano 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]; 5181674ae819SStefano Zampini } 5182674ae819SStefano Zampini } 5183674ae819SStefano Zampini } 5184674ae819SStefano Zampini #else /* on missing GESVD */ 5185e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5186e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5187b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5188674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5189674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 51909162d606SStefano 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)); 5191674ae819SStefano Zampini #else 51929162d606SStefano 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)); 5193674ae819SStefano Zampini #endif 5194984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 5195674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5196984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 5197e310c8b4SStefano Zampini k = temp_constraints; 5198e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 5199674ae819SStefano Zampini j = 0; 5200e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 520145a1bb75SStefano Zampini valid_constraints = k-j; 5202911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 5203984c4197SStefano Zampini #endif /* on missing GESVD */ 5204674ae819SStefano Zampini } 5205a773dcb8SStefano Zampini } 52069162d606SStefano Zampini /* update pointers information */ 52079162d606SStefano Zampini if (valid_constraints) { 52089162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 52099162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 52109162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 52119162d606SStefano Zampini /* set change_of_basis flag */ 521245a1bb75SStefano Zampini if (boolforchange) { 5213b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 52149162d606SStefano Zampini } 5215b3d85658SStefano Zampini total_counts_cc++; 521645a1bb75SStefano Zampini } 521745a1bb75SStefano Zampini } 5218984c4197SStefano Zampini /* free workspace */ 52198f1c130eSStefano Zampini if (!skip_lapack) { 5220984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 5221984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5222984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 5223984c4197SStefano Zampini #endif 5224984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 5225984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5226984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 5227984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 5228984c4197SStefano Zampini #endif 5229984c4197SStefano Zampini } 5230984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5231984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 5232984c4197SStefano Zampini } 5233984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 5234cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 5235cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5236cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5237cf5a6209SStefano Zampini } 5238cf5a6209SStefano Zampini if (n_ISForFaces) { 5239cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5240cf5a6209SStefano Zampini } 5241cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5242cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5243cf5a6209SStefano Zampini } 5244cf5a6209SStefano Zampini if (n_ISForEdges) { 5245cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5246cf5a6209SStefano Zampini } 5247cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 524808122e43SStefano Zampini } else { 524908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5250984c4197SStefano Zampini 525108122e43SStefano Zampini total_counts = 0; 525208122e43SStefano Zampini n_vertices = 0; 5253d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 5254d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 525508122e43SStefano Zampini } 525608122e43SStefano Zampini max_constraints = 0; 52579162d606SStefano Zampini total_counts_cc = 0; 525808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 525908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 52609162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 526108122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 526208122e43SStefano Zampini } 52639162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 52649162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 52659162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 52669162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 526774d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 52689162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 52699162d606SStefano Zampini total_counts_cc = 0; 52709162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 52719162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 52729162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 527308122e43SStefano Zampini } 527408122e43SStefano Zampini } 52759162d606SStefano Zampini #if 0 52769162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 52779162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 52789162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 52799162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 52809162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 52819162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 52829162d606SStefano Zampini } 52839162d606SStefano Zampini printf("\n"); 52849162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 52859162d606SStefano Zampini } 52861b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 52878bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 52881b968477SStefano Zampini } 52891b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 52908bec7fa6SStefano 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]); 52911b968477SStefano Zampini } 529208122e43SStefano Zampini #endif 529308122e43SStefano Zampini 52948bec7fa6SStefano Zampini max_size_of_constraint = 0; 52959162d606SStefano 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]); 52969162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 529708122e43SStefano Zampini /* Change of basis */ 5298b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 529908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 530008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 530108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 5302b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 530308122e43SStefano Zampini } 530408122e43SStefano Zampini } 530508122e43SStefano Zampini } 530608122e43SStefano Zampini } 5307984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 53084f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 530908122e43SStefano Zampini 53109162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 53119162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 53126c4ed002SBarry 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); 5313674ae819SStefano Zampini 5314674ae819SStefano Zampini /* Create constraint matrix */ 5315674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 531616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 5317984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 5318984c4197SStefano Zampini 5319984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 5320a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 5321a717540cSStefano Zampini qr_needed = PETSC_FALSE; 532274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 5323984c4197SStefano Zampini total_primal_vertices=0; 5324b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 53259162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 53269162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 532772b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 53289162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 5329b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 533064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 53319162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 53329162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 5333a717540cSStefano Zampini } 5334b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 533591af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 5336a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 5337a717540cSStefano Zampini qr_needed = PETSC_TRUE; 5338a717540cSStefano Zampini } 5339fa434743SStefano Zampini } else { 5340b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 5341fa434743SStefano Zampini } 5342a717540cSStefano Zampini } 5343b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 5344b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 5345674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 534670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5347b3d85658SStefano Zampini 53484f1b2e48SStefano 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); 53490e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 53500e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 5351984c4197SStefano Zampini 5352984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 535374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 5354785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 5355984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 535674d5cdf7SStefano Zampini 5357984c4197SStefano Zampini j = total_primal_vertices; 535874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 5359b3d85658SStefano Zampini cum = total_primal_vertices; 53609162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 53614641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 5362b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 5363b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 5364b3d85658SStefano Zampini cum++; 53659162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 536674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 536774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 536874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 536974d5cdf7SStefano Zampini } 53709162d606SStefano Zampini j += constraints_n[i]; 5371674ae819SStefano Zampini } 5372674ae819SStefano Zampini } 5373674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 5374674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 5375088faed8SStefano Zampini 5376674ae819SStefano Zampini /* set values in constraint matrix */ 5377984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 53780e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5379674ae819SStefano Zampini } 5380984c4197SStefano Zampini total_counts = total_primal_vertices; 53819162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 53824641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 53839162d606SStefano Zampini PetscInt *cols; 53849162d606SStefano Zampini 53859162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 53869162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 53879162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 53889162d606SStefano Zampini PetscInt row = total_counts+k; 53899162d606SStefano Zampini PetscScalar *vals; 53909162d606SStefano Zampini 53919162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 53929162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 53939162d606SStefano Zampini } 53949162d606SStefano Zampini total_counts += constraints_n[i]; 5395674ae819SStefano Zampini } 5396674ae819SStefano Zampini } 5397674ae819SStefano Zampini /* assembling */ 5398674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5399674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5400088faed8SStefano Zampini 5401984c4197SStefano Zampini /* 54026a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5403984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 5404f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5405984c4197SStefano Zampini */ 5406674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 5407674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 5408026de310SStefano Zampini /* dual and primal dofs on a single cc */ 5409984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 5410984c4197SStefano Zampini /* working stuff for GEQRF */ 541181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 5412984c4197SStefano Zampini PetscBLASInt lqr_work; 5413984c4197SStefano Zampini /* working stuff for UNGQR */ 5414984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 5415984c4197SStefano Zampini PetscBLASInt lgqr_work; 5416984c4197SStefano Zampini /* working stuff for TRTRS */ 5417984c4197SStefano Zampini PetscScalar *trs_rhs; 54183f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 5419984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 5420984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 5421984c4197SStefano Zampini PetscScalar *start_vals; 5422984c4197SStefano Zampini /* working stuff for values insertion */ 54234641a718SStefano Zampini PetscBT is_primal; 542464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 5425906d46d4SStefano Zampini /* matrix sizes */ 5426906d46d4SStefano Zampini PetscInt global_size,local_size; 5427906d46d4SStefano Zampini /* temporary change of basis */ 5428906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 5429cf5a6209SStefano Zampini /* extra space for debugging */ 5430cf5a6209SStefano Zampini PetscScalar *dbg_work; 5431984c4197SStefano Zampini 5432906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 5433906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 543416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 5435bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 5436906d46d4SStefano Zampini /* nonzeros for local mat */ 5437bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 54381dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5439bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 54401dd7afcfSStefano Zampini } else { 54411dd7afcfSStefano Zampini const PetscInt *ii; 54421dd7afcfSStefano Zampini PetscInt n; 54431dd7afcfSStefano Zampini PetscBool flg_row; 54441dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 54451dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 54461dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 54471dd7afcfSStefano Zampini } 54489162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 5449a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 54509162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 5451a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 54529162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 5453a717540cSStefano Zampini } else { 54549162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 54559162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 5456a717540cSStefano Zampini } 5457a717540cSStefano Zampini } 5458a717540cSStefano Zampini } 5459906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 5460bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 54611dd7afcfSStefano Zampini /* Set interior change in the matrix */ 54621dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5463bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 5464906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 5465a717540cSStefano Zampini } 54661dd7afcfSStefano Zampini } else { 54671dd7afcfSStefano Zampini const PetscInt *ii,*jj; 54681dd7afcfSStefano Zampini PetscScalar *aa; 54691dd7afcfSStefano Zampini PetscInt n; 54701dd7afcfSStefano Zampini PetscBool flg_row; 54711dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 54721dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 54731dd7afcfSStefano Zampini for (i=0;i<n;i++) { 54741dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 54751dd7afcfSStefano Zampini } 54761dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 54771dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 54781dd7afcfSStefano Zampini } 5479a717540cSStefano Zampini 5480a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5481a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5482a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5483a717540cSStefano Zampini } 5484a717540cSStefano Zampini 5485a717540cSStefano Zampini 5486a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 5487a717540cSStefano Zampini /* 5488a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 5489a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 5490a717540cSStefano Zampini 5491a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 5492a717540cSStefano Zampini 5493a6b551f4SStefano 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) 5494a6b551f4SStefano Zampini 5495a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 5496a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 5497a717540cSStefano Zampini | ... | 5498a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 5499a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 5500a717540cSStefano Zampini 5501a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 5502a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 5503a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 5504a6b551f4SStefano Zampini 5505a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 5506a717540cSStefano Zampini */ 5507a717540cSStefano Zampini if (qr_needed) { 5508984c4197SStefano Zampini /* space to store Q */ 5509854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 55104e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 55114e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 5512984c4197SStefano Zampini /* first we issue queries for optimal work */ 55133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 55143f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 55153f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5516984c4197SStefano Zampini lqr_work = -1; 55173f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 5518984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 5519984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 5520785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 5521984c4197SStefano Zampini lgqr_work = -1; 55223f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 55233f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 55243f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 55253f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 55263f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 55273f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 5528984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 5529984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 5530785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 5531984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 5532785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 5533a717540cSStefano Zampini /* allocating workspace for check */ 5534a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5535cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 5536a717540cSStefano Zampini } 5537a717540cSStefano Zampini } 5538984c4197SStefano Zampini /* array to store whether a node is primal or not */ 55394641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 5540473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 55410e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 55426c4ed002SBarry 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); 554339e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 554439e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 554539e2fb2aSStefano Zampini } 554639e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 5547984c4197SStefano Zampini 5548a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 55499162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 55509162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 55514641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 5552984c4197SStefano Zampini /* get constraint info */ 55539162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 5554984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 5555984c4197SStefano Zampini 5556984c4197SStefano Zampini if (pcbddc->dbg_flag) { 55579162d606SStefano 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); 5558674ae819SStefano Zampini } 5559984c4197SStefano Zampini 5560fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 5561a717540cSStefano Zampini 5562a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 5563a717540cSStefano Zampini if (pcbddc->dbg_flag) { 55649162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5565a717540cSStefano Zampini } 5566984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 55679162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5568984c4197SStefano Zampini 5569984c4197SStefano Zampini /* compute QR decomposition of constraints */ 55703f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 55713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 55723f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5573674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 55743f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 5575984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 5576674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5577984c4197SStefano Zampini 5578984c4197SStefano Zampini /* explictly compute R^-T */ 5579984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 5580984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 55813f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 55823f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 55833f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 55843f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 5585984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 55863f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 5587984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 5588984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5589984c4197SStefano Zampini 5590a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 55913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 55923f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 55933f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 55943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5595984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 55963f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 5597984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 5598984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5599984c4197SStefano Zampini 5600984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 5601984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 5602984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 56033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 56043f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 56053f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 56063f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 56073f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 56083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5609984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 56109162d606SStefano 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)); 5611984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 56129162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5613984c4197SStefano Zampini 5614984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 56159162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 5616984c4197SStefano Zampini /* insert cols for primal dofs */ 5617984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 5618984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 56199162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 5620906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 5621984c4197SStefano Zampini } 5622984c4197SStefano Zampini /* insert cols for dual dofs */ 5623984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 56249162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 5625984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 56269162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 5627906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 5628984c4197SStefano Zampini j++; 5629674ae819SStefano Zampini } 5630674ae819SStefano Zampini } 5631984c4197SStefano Zampini 5632984c4197SStefano Zampini /* check change of basis */ 5633984c4197SStefano Zampini if (pcbddc->dbg_flag) { 5634984c4197SStefano Zampini PetscInt ii,jj; 5635984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 5636c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 5637c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5638c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 5639c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5640c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 5641c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 5642984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5643cf5a6209SStefano 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)); 5644984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5645984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 5646984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 5647cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 5648cf5a6209SStefano 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; 5649674ae819SStefano Zampini } 5650674ae819SStefano Zampini } 5651984c4197SStefano Zampini if (!valid_qr) { 565222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 5653984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 5654984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 5655cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 5656cf5a6209SStefano 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])); 5657674ae819SStefano Zampini } 5658cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 5659cf5a6209SStefano 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])); 5660984c4197SStefano Zampini } 5661984c4197SStefano Zampini } 5662984c4197SStefano Zampini } 5663674ae819SStefano Zampini } else { 566422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 5665674ae819SStefano Zampini } 5666674ae819SStefano Zampini } 5667a717540cSStefano Zampini } else { /* simple transformation block */ 5668a717540cSStefano Zampini PetscInt row,col; 5669a6b551f4SStefano Zampini PetscScalar val,norm; 5670a6b551f4SStefano Zampini 5671a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 56729162d606SStefano 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)); 5673a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 56749162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 56759162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 5676bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 56779162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 5678906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 56799162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 5680a717540cSStefano Zampini } else { 5681a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 56829162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 5683a717540cSStefano Zampini if (row != col) { 56849162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 5685a717540cSStefano Zampini } else { 56869162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 5687a717540cSStefano Zampini } 5688906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 5689a717540cSStefano Zampini } 5690a717540cSStefano Zampini } 5691a717540cSStefano Zampini } 569298a51de6SStefano Zampini if (pcbddc->dbg_flag) { 569322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 5694a717540cSStefano Zampini } 5695674ae819SStefano Zampini } 5696984c4197SStefano Zampini } else { 5697984c4197SStefano Zampini if (pcbddc->dbg_flag) { 56989162d606SStefano 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); 5699674ae819SStefano Zampini } 5700674ae819SStefano Zampini } 5701674ae819SStefano Zampini } 5702a717540cSStefano Zampini 5703a717540cSStefano Zampini /* free workspace */ 5704a717540cSStefano Zampini if (qr_needed) { 5705984c4197SStefano Zampini if (pcbddc->dbg_flag) { 5706cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 5707984c4197SStefano Zampini } 5708984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 5709984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 5710984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 5711984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 5712984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 5713674ae819SStefano Zampini } 5714a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 5715906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5716906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5717906d46d4SStefano Zampini 5718906d46d4SStefano Zampini /* assembling of global change of variable */ 571988c03ad3SStefano Zampini if (!pcbddc->fake_change) { 5720bbb9e6c6SStefano Zampini Mat tmat; 572116f15bc4SStefano Zampini PetscInt bs; 572216f15bc4SStefano Zampini 5723906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 5724906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 5725bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 5726bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 5727bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 5728bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 572916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 573016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 5731906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 5732bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 5733bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 5734bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 5735bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5736bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5737e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5738e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5739bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 5740bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 574188c03ad3SStefano Zampini 5742906d46d4SStefano Zampini /* check */ 5743906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 5744906d46d4SStefano Zampini PetscReal error; 5745906d46d4SStefano Zampini Vec x,x_change; 5746906d46d4SStefano Zampini 5747906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 5748906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 5749906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 5750906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 5751e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5752e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5753bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 5754e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5755e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5756906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 5757906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 5758906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 5759906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5760bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 5761906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 5762906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 5763906d46d4SStefano Zampini } 5764b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 5765b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 5766b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5767bf3a8328SStefano Zampini 57689a962809SStefano 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); 5769b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 5770ac632422SStefano Zampini Mat S_new,tmat; 5771bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 5772bbb9e6c6SStefano Zampini 5773bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 57746816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 5775bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 5776bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 5777bf3a8328SStefano Zampini IS is_V; 5778b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 5779b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 5780b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 5781b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 5782b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 5783bf3a8328SStefano Zampini } 5784bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 5785ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 5786b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 5787ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 5788bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 5789bf3a8328SStefano Zampini const PetscScalar *array; 5790bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 5791bf3a8328SStefano Zampini PetscInt i,n_V; 5792bf3a8328SStefano Zampini 5793b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 5794b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 5795b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 5796b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 5797b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 5798b087196eSStefano Zampini for (i=0;i<n_V;i++) { 5799b087196eSStefano Zampini PetscScalar val; 5800b087196eSStefano Zampini PetscInt idx; 5801b087196eSStefano Zampini 5802b087196eSStefano Zampini idx = idxs_V[i]; 5803b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 5804b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 5805b087196eSStefano Zampini } 5806b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5807b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5808bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 5809bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 5810bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 5811bf3a8328SStefano Zampini } 5812ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 5813ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 5814ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 5815ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 5816b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 5817ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 5818bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 5819b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 5820bf3a8328SStefano Zampini } 5821ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 5822ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 5823ac632422SStefano Zampini } 5824b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 582588c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 5826b96c3477SStefano Zampini } 5827c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 5828b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 5829c9db6a07SStefano Zampini PetscInt i; 5830c9db6a07SStefano Zampini 5831c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5832c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 5833c9db6a07SStefano Zampini } 5834c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 5835c9db6a07SStefano Zampini } 5836b96c3477SStefano Zampini } 583716909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 583816909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 583916909a7fSStefano Zampini } else { 5840906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 584116909a7fSStefano Zampini } 58421dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 584327b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 584472b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 584572b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 584672b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 584772b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 584872b8c272SStefano Zampini } 58491dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 585027b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 5851b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 5852b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 5853906d46d4SStefano Zampini } else { 58541dd7afcfSStefano Zampini Mat benign_global = NULL; 585527b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 58561dd7afcfSStefano Zampini Mat tmat; 58571dd7afcfSStefano Zampini 58581dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 58591dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 58601dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 58611dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 58621dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 58631dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 58641dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 58651dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 58661dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 58671dd7afcfSStefano Zampini if (pcbddc->benign_change) { 58681dd7afcfSStefano Zampini Mat M; 58691dd7afcfSStefano Zampini 58701dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 58711dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 58721dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 58731dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 5874906d46d4SStefano Zampini } else { 58751dd7afcfSStefano Zampini Mat eye; 58761dd7afcfSStefano Zampini PetscScalar *array; 58771dd7afcfSStefano Zampini 58781dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 58791dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 58801dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 58811dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 5882906d46d4SStefano Zampini } 58831dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 58841dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 58851dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 58861dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 58871dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 58881dd7afcfSStefano Zampini } 58891dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 58901dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 58911dd7afcfSStefano Zampini } 58921dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 58931dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 58941dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 589527b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 58961dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 58971dd7afcfSStefano Zampini } 58981dd7afcfSStefano Zampini } 589916909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 590016909a7fSStefano Zampini IS is_global; 590116909a7fSStefano Zampini const PetscInt *gidxs; 590216909a7fSStefano Zampini 590316909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 590416909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 590516909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 590616909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 590716909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 590816909a7fSStefano Zampini } 59091dd7afcfSStefano Zampini } 59101dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 59111dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 5912b9b85e73SStefano Zampini } 5913a717540cSStefano Zampini 591472b8c272SStefano Zampini if (!pcbddc->fake_change) { 59154f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 59164f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 59174f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 59184f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 5919019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 5920019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 5921019a44ceSStefano Zampini pcbddc->local_primal_size++; 5922019a44ceSStefano Zampini } 5923019a44ceSStefano Zampini 5924019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 5925727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 5926727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 59279f47a83aSStefano 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); 5928c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 59290e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 59309f47a83aSStefano 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); 5931727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 5932727cdba6SStefano Zampini } 59330e6343abSStefano Zampini } 5934727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 5935b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 593672b8c272SStefano Zampini } 593772b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 5938727cdba6SStefano Zampini 5939a717540cSStefano Zampini /* flush dbg viewer */ 5940b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 5941b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5942b8ffe317SStefano Zampini } 5943a717540cSStefano Zampini 5944e310c8b4SStefano Zampini /* free workspace */ 5945a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 59464641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 594708122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 59489162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 59499162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 595008122e43SStefano Zampini } else { 59519162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 59529162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 59539162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 595408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 595508122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 59569162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 59579162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 595808122e43SStefano Zampini } 5959674ae819SStefano Zampini PetscFunctionReturn(0); 5960674ae819SStefano Zampini } 5961674ae819SStefano Zampini 5962674ae819SStefano Zampini #undef __FUNCT__ 5963674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 5964674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 5965674ae819SStefano Zampini { 596671582508SStefano Zampini ISLocalToGlobalMapping map; 5967674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5968674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 596914f95afaSStefano Zampini PetscInt ierr,i,N; 5970674ae819SStefano Zampini 5971674ae819SStefano Zampini PetscFunctionBegin; 59728af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 5973b03ebc13SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 59748e61c736SStefano Zampini /* Reset previously computed graph */ 59758e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 5976674ae819SStefano Zampini /* Init local Graph struct */ 59777fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 597871582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 5979be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 5980674ae819SStefano Zampini 5981575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 59829a962809SStefano 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); 59839577ea80SStefano Zampini 5984674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 5985d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 59864d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 59874d379d7bSStefano Zampini PetscInt nvtxs; 5988e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 5989674ae819SStefano Zampini 59902fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 59912fffb893SStefano Zampini if (flg_row) { 59924d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 5993b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 59942fffb893SStefano Zampini } 59952fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 5996674ae819SStefano Zampini } 59979b28b941SStefano Zampini if (pcbddc->dbg_flag) { 59989b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5999674ae819SStefano Zampini } 6000674ae819SStefano Zampini 6001674ae819SStefano Zampini /* Setup of Graph */ 60024b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 600314f95afaSStefano 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); 6004674ae819SStefano Zampini 60054f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 60064f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 60074f1b2e48SStefano Zampini PetscInt *local_subs; 60084f1b2e48SStefano Zampini 60094f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 60104f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 60114f1b2e48SStefano Zampini const PetscInt *idxs; 60124f1b2e48SStefano Zampini PetscInt nl,j; 60134f1b2e48SStefano Zampini 60144f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 60154f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 601671582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 60174f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 60184f1b2e48SStefano Zampini } 60194f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 60204f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 60214f1b2e48SStefano Zampini } 60228af8fcf9SStefano Zampini } 60234f1b2e48SStefano Zampini 6024cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 6025674ae819SStefano Zampini /* Graph's connected components analysis */ 6026674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 602771582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 60288af8fcf9SStefano Zampini } 6029674ae819SStefano Zampini PetscFunctionReturn(0); 6030674ae819SStefano Zampini } 6031674ae819SStefano Zampini 60329a7d3425SStefano Zampini #undef __FUNCT__ 60339a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 60349a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 60359a7d3425SStefano Zampini { 60369a7d3425SStefano Zampini PetscInt i,j; 60379a7d3425SStefano Zampini PetscScalar *alphas; 60389a7d3425SStefano Zampini PetscErrorCode ierr; 60399a7d3425SStefano Zampini 60409a7d3425SStefano Zampini PetscFunctionBegin; 6041785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 60429a7d3425SStefano Zampini for (i=0;i<n;i++) { 60439a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 6044669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 6045669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 6046669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 60479a7d3425SStefano Zampini } 60489a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 60499a7d3425SStefano Zampini PetscFunctionReturn(0); 60509a7d3425SStefano Zampini } 60519a7d3425SStefano Zampini 6052e7931f94SStefano Zampini #undef __FUNCT__ 605370cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 605457de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 6055e7931f94SStefano Zampini { 605657de7509SStefano Zampini Mat A; 6057e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 6058e7931f94SStefano Zampini PetscMPIInt size,rank,color; 605952e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 606052e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 606127b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 606257de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 606327b6a85dSStefano Zampini PetscInt xadj_count, *count; 606427b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 606527b6a85dSStefano Zampini PetscSubcomm psubcomm; 606627b6a85dSStefano Zampini MPI_Comm subcomm; 606752e5ac9dSStefano Zampini PetscErrorCode ierr; 6068a57a6d2fSStefano Zampini 6069e7931f94SStefano Zampini PetscFunctionBegin; 607057de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 607157de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 607257de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 607357de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 607457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 607557de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 607657de7509SStefano Zampini 607757de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 607857de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 607957de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 608057de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 608157de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 608257de7509SStefano Zampini im_active = !!(n); 608357de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 608457de7509SStefano Zampini void_procs = size - active_procs; 608557de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 608657de7509SStefano Zampini if (void_procs) { 608757de7509SStefano Zampini PetscInt ncand; 608857de7509SStefano Zampini 608957de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 609057de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 609157de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 609257de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 609357de7509SStefano Zampini if (!procs_candidates[i]) { 609457de7509SStefano Zampini procs_candidates[ncand++] = i; 609557de7509SStefano Zampini } 609657de7509SStefano Zampini } 609757de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 609857de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 609957de7509SStefano Zampini } 610057de7509SStefano Zampini 610114f0bfb9SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix 610214f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 610314f0bfb9SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1) { 610414f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 610514f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 610614f0bfb9SStefano Zampini else dest = rank; 610757de7509SStefano Zampini if (im_active) { 610857de7509SStefano Zampini issize = 1; 610957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 611014f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 611157de7509SStefano Zampini } else { 611214f0bfb9SStefano Zampini isidx = dest; 611357de7509SStefano Zampini } 611457de7509SStefano Zampini } else { 611557de7509SStefano Zampini issize = 0; 611657de7509SStefano Zampini isidx = -1; 611757de7509SStefano Zampini } 611857de7509SStefano Zampini *n_subdomains = active_procs; 611957de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 6120daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 612157de7509SStefano Zampini PetscFunctionReturn(0); 612257de7509SStefano Zampini } 6123c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 6124c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 612527b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 6126e7931f94SStefano Zampini 6127e7931f94SStefano Zampini /* Get info on mapping */ 61283bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 61293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6130e7931f94SStefano Zampini 6131e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 6132785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 6133e7931f94SStefano Zampini xadj[0] = 0; 6134e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 6135785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 6136785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 613727b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 613827b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 613927b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 614027b6a85dSStefano Zampini count[shared[i][j]] += 1; 6141e7931f94SStefano Zampini 614227b6a85dSStefano Zampini xadj_count = 0; 61432b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 614427b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 614527b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 6146d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 6147d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 6148d023bfaeSStefano Zampini xadj_count++; 614927b6a85dSStefano Zampini break; 615027b6a85dSStefano Zampini } 6151e7931f94SStefano Zampini } 6152e7931f94SStefano Zampini } 6153d023bfaeSStefano Zampini xadj[1] = xadj_count; 615427b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 61553bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6156e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 6157e7931f94SStefano Zampini 61583837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 6159e7931f94SStefano Zampini 616027b6a85dSStefano Zampini /* Restrict work on active processes only */ 616127b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 616227b6a85dSStefano Zampini if (void_procs) { 616327b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 616427b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 616527b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 616627b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 616727b6a85dSStefano Zampini } else { 616827b6a85dSStefano Zampini psubcomm = NULL; 616927b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 617027b6a85dSStefano Zampini } 617127b6a85dSStefano Zampini 617227b6a85dSStefano Zampini v_wgt = NULL; 617327b6a85dSStefano Zampini if (!color) { 6174e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 6175e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 6176e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6177c8587f34SStefano Zampini } else { 617852e5ac9dSStefano Zampini Mat subdomain_adj; 617952e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 618052e5ac9dSStefano Zampini MatPartitioning partitioner; 618127b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 618252e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 618357de7509SStefano Zampini PetscMPIInt size; 6184b0c7d250SStefano Zampini PetscBool aggregate; 6185b0c7d250SStefano Zampini 618627b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 618727b6a85dSStefano Zampini if (void_procs) { 618827b6a85dSStefano Zampini PetscInt prank = rank; 6189785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 619027b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 6191e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 6192e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 6193c8587f34SStefano Zampini } 6194e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 619527b6a85dSStefano Zampini } else { 619627b6a85dSStefano Zampini oldranks = NULL; 619727b6a85dSStefano Zampini } 6198b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 619927b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 6200b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 6201b0c7d250SStefano Zampini PetscMPIInt nrank; 6202b0c7d250SStefano Zampini PetscScalar *vals; 6203b0c7d250SStefano Zampini 620427b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 6205b0c7d250SStefano Zampini lrows = 0; 6206b0c7d250SStefano Zampini if (nrank<redprocs) { 6207b0c7d250SStefano Zampini lrows = size/redprocs; 6208b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 6209b0c7d250SStefano Zampini } 621027b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 6211b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 6212b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6213b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6214b0c7d250SStefano Zampini row = nrank; 6215b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 6216b0c7d250SStefano Zampini cols = adjncy; 6217b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 6218b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 6219b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 6220b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6221b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 622252e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 622352e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 622452e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6225b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 622627b6a85dSStefano Zampini if (use_vwgt) { 622727b6a85dSStefano Zampini Vec v; 622827b6a85dSStefano Zampini const PetscScalar *array; 622927b6a85dSStefano Zampini PetscInt nl; 623027b6a85dSStefano Zampini 623127b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 623227b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 623327b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 623427b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 623527b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 623627b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 623727b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 623822db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 623927b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 624027b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 624127b6a85dSStefano Zampini } 6242b0c7d250SStefano Zampini } else { 624327b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 624427b6a85dSStefano Zampini if (use_vwgt) { 624527b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 624627b6a85dSStefano Zampini v_wgt[0] = local_size; 624727b6a85dSStefano Zampini } 6248b0c7d250SStefano Zampini } 624922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 6250e7931f94SStefano Zampini 6251e7931f94SStefano Zampini /* Partition */ 625227b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 6253e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 625427b6a85dSStefano Zampini if (v_wgt) { 6255e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 6256c8587f34SStefano Zampini } 625757de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 625857de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 6259e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 6260e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 626122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 6262e7931f94SStefano Zampini 626352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 62646583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 626552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 626652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 626757de7509SStefano Zampini if (!aggregate) { 626857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 626927b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 627027b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 627127b6a85dSStefano Zampini #endif 627257de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 627327b6a85dSStefano Zampini } else if (oldranks) { 6274b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 627527b6a85dSStefano Zampini } else { 627627b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 627757de7509SStefano Zampini } 627828143c3dSStefano Zampini } else { 6279b0c7d250SStefano Zampini PetscInt idxs[1]; 6280b0c7d250SStefano Zampini PetscMPIInt tag; 6281b0c7d250SStefano Zampini MPI_Request *reqs; 6282b0c7d250SStefano Zampini 6283b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 6284b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 6285b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 628627b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 628728143c3dSStefano Zampini } 628827b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 6289b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6290b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 629157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 629227b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 629327b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 629427b6a85dSStefano Zampini #endif 629557de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 629627b6a85dSStefano Zampini } else if (oldranks) { 6297b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 629827b6a85dSStefano Zampini } else { 629927b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 6300e7931f94SStefano Zampini } 630157de7509SStefano Zampini } 630252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6303e7931f94SStefano Zampini /* clean up */ 6304e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 630552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 6306e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 6307e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 6308e7931f94SStefano Zampini } 630927b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 631057de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 6311e7931f94SStefano Zampini 6312e7931f94SStefano Zampini /* assemble parallel IS for sends */ 6313e7931f94SStefano Zampini i = 1; 631427b6a85dSStefano Zampini if (!color) i=0; 631557de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 6316e7931f94SStefano Zampini PetscFunctionReturn(0); 6317e7931f94SStefano Zampini } 6318e7931f94SStefano Zampini 6319e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 6320e7931f94SStefano Zampini 6321e7931f94SStefano Zampini #undef __FUNCT__ 6322e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 63231ae86dd6SStefano 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[]) 6324e7931f94SStefano Zampini { 632570cf5478SStefano Zampini Mat local_mat; 6326e7931f94SStefano Zampini IS is_sends_internal; 63279d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 63281ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 63299d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 6330e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 6331e7931f94SStefano Zampini PetscInt* l2gmap_indices; 6332e7931f94SStefano Zampini const PetscInt* is_indices; 6333e7931f94SStefano Zampini MatType new_local_type; 6334e7931f94SStefano Zampini /* buffers */ 6335e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 633628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 63379d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 6338e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 63391ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 6340e7931f94SStefano Zampini /* MPI */ 634128143c3dSStefano Zampini MPI_Comm comm,comm_n; 634228143c3dSStefano Zampini PetscSubcomm subcomm; 6343e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 634428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 634528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 63461ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 63471ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 63481ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 6349e7931f94SStefano Zampini PetscErrorCode ierr; 6350e7931f94SStefano Zampini 6351e7931f94SStefano Zampini PetscFunctionBegin; 635257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 6353e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 635428143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 635557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 635657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 635757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 635857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 635957de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 63601ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 63611ae86dd6SStefano Zampini if (nvecs) { 63621ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 63631ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 63641ae86dd6SStefano Zampini } 636557de7509SStefano Zampini /* further checks */ 6366e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6367e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 6368e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 6369e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 6370e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 637157de7509SStefano Zampini if (reuse && *mat_n) { 637270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 637357de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 637470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 637528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 637670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 637770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 637870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 637970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 638070cf5478SStefano Zampini } 6381e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 6382e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 638357de7509SStefano Zampini 6384e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 6385e7931f94SStefano Zampini if (!is_sends) { 638628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 638757de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 6388c8587f34SStefano Zampini } else { 6389e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 6390e7931f94SStefano Zampini is_sends_internal = is_sends; 6391c8587f34SStefano Zampini } 6392e7931f94SStefano Zampini 6393e7931f94SStefano Zampini /* get comm */ 6394a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 6395e7931f94SStefano Zampini 6396e7931f94SStefano Zampini /* compute number of sends */ 6397e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 6398e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 6399e7931f94SStefano Zampini 6400e7931f94SStefano Zampini /* compute number of receives */ 6401e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 6402785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 6403e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 6404e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6405e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 6406e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 6407e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 6408e7931f94SStefano Zampini 640928143c3dSStefano Zampini /* restrict comm if requested */ 641028143c3dSStefano Zampini subcomm = 0; 641128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 641228143c3dSStefano Zampini if (restrict_comm) { 6413779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 6414779c1cceSStefano Zampini 641528143c3dSStefano Zampini color = 0; 641653a05cb3SStefano Zampini if (restrict_full) { 641753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 641853a05cb3SStefano Zampini } else { 641953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 642053a05cb3SStefano Zampini } 6421b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 642228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 642328143c3dSStefano Zampini /* check if reuse has been requested */ 642457de7509SStefano Zampini if (reuse) { 642528143c3dSStefano Zampini if (*mat_n) { 642628143c3dSStefano Zampini PetscMPIInt subcommsize2; 642728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 642828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 642928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 643028143c3dSStefano Zampini } else { 643128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 643228143c3dSStefano Zampini } 643328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 6434779c1cceSStefano Zampini PetscMPIInt rank; 6435779c1cceSStefano Zampini 6436779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 643728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 643828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 643928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 6440306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 644128143c3dSStefano Zampini } 644228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 644328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 644428143c3dSStefano Zampini } else { 644528143c3dSStefano Zampini comm_n = comm; 644628143c3dSStefano Zampini } 644728143c3dSStefano Zampini 6448e7931f94SStefano Zampini /* prepare send/receive buffers */ 6449785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 6450e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 6451785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 6452e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 645328143c3dSStefano Zampini if (nis) { 6454854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 645528143c3dSStefano Zampini } 6456e7931f94SStefano Zampini 645728143c3dSStefano Zampini /* Get data from local matrices */ 64586c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 6459e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 6460e7931f94SStefano Zampini /* 6461e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 6462e7931f94SStefano Zampini send_buffer_idxs should contain: 6463e7931f94SStefano Zampini - MatType_PRIVATE type 6464e7931f94SStefano Zampini - PetscInt size_of_l2gmap 6465e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 6466e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 6467e7931f94SStefano Zampini */ 64686c4ed002SBarry Smith else { 6469e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 64703bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 6471854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 6472e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 6473e7931f94SStefano Zampini send_buffer_idxs[1] = i; 64743bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6475e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 64763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6477e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 6478e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6479e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 6480e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 6481c8587f34SStefano Zampini } 6482c8587f34SStefano Zampini } 6483e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 648428143c3dSStefano Zampini /* additional is (if any) */ 648528143c3dSStefano Zampini if (nis) { 648628143c3dSStefano Zampini PetscMPIInt psum; 648728143c3dSStefano Zampini PetscInt j; 648828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 648928143c3dSStefano Zampini PetscInt plen; 649028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 649128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 649228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 649328143c3dSStefano Zampini } 6494854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 649528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 649628143c3dSStefano Zampini PetscInt plen; 649728143c3dSStefano Zampini const PetscInt *is_array_idxs; 649828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 649928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 650028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 650128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 650228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 650328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 650428143c3dSStefano Zampini } 650528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 650628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 650728143c3dSStefano Zampini } 650828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 650928143c3dSStefano Zampini } 651028143c3dSStefano Zampini 6511e7931f94SStefano Zampini buf_size_idxs = 0; 6512e7931f94SStefano Zampini buf_size_vals = 0; 651328143c3dSStefano Zampini buf_size_idxs_is = 0; 65141ae86dd6SStefano Zampini buf_size_vecs = 0; 6515e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6516e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 6517e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 651828143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 65191ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 6520e7931f94SStefano Zampini } 6521785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 6522785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 652395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 65241ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 6525e7931f94SStefano Zampini 6526e7931f94SStefano Zampini /* get new tags for clean communications */ 6527e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 6528e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 652928143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 65301ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 6531e7931f94SStefano Zampini 6532e7931f94SStefano Zampini /* allocate for requests */ 6533785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 6534785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 653595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 65361ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 6537785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 6538785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 653995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 65401ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 6541e7931f94SStefano Zampini 6542e7931f94SStefano Zampini /* communications */ 6543e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6544e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 654528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 65461ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 6547e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6548e7931f94SStefano Zampini source_dest = onodes[i]; 6549e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 6550e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 6551e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6552e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 655328143c3dSStefano Zampini if (nis) { 655457de7509SStefano Zampini source_dest = onodes_is[i]; 655528143c3dSStefano 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); 655628143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 655728143c3dSStefano Zampini } 65581ae86dd6SStefano Zampini if (nvecs) { 65591ae86dd6SStefano Zampini source_dest = onodes[i]; 65601ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 65611ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 65621ae86dd6SStefano Zampini } 6563e7931f94SStefano Zampini } 6564e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6565e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 6566e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 6567e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 656828143c3dSStefano Zampini if (nis) { 656928143c3dSStefano 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); 657028143c3dSStefano Zampini } 65711ae86dd6SStefano Zampini if (nvecs) { 65721ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 65731ae86dd6SStefano 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); 65741ae86dd6SStefano Zampini } 6575e7931f94SStefano Zampini } 6576e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6577e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 6578e7931f94SStefano Zampini 6579e7931f94SStefano Zampini /* assemble new l2g map */ 6580e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6581e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 65829d30be91SStefano Zampini new_local_rows = 0; 6583e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 65849d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6585e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6586e7931f94SStefano Zampini } 65879d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 6588e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 65899d30be91SStefano Zampini new_local_rows = 0; 6590e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 65919d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 65929d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6593e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6594e7931f94SStefano Zampini } 65959d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 65969d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 6597e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 6598e7931f94SStefano Zampini 6599e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 6600e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 6601e7931f94SStefano 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) */ 6602e7931f94SStefano Zampini if (n_recvs) { 660328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 6604e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6605e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6606e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 6607e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 6608e7931f94SStefano Zampini break; 6609e7931f94SStefano Zampini } 6610e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6611e7931f94SStefano Zampini } 6612e7931f94SStefano Zampini switch (new_local_type_private) { 661328143c3dSStefano Zampini case MATDENSE_PRIVATE: 661428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 6615e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 6616e7931f94SStefano Zampini bs = 1; 661728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 661828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 661928143c3dSStefano Zampini bs = 1; 662028143c3dSStefano Zampini } 6621e7931f94SStefano Zampini break; 6622e7931f94SStefano Zampini case MATAIJ_PRIVATE: 6623e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 6624e7931f94SStefano Zampini bs = 1; 6625e7931f94SStefano Zampini break; 6626e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 6627e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 6628e7931f94SStefano Zampini break; 6629e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 6630e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 6631e7931f94SStefano Zampini break; 6632e7931f94SStefano Zampini default: 66339d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 6634e7931f94SStefano Zampini break; 6635e7931f94SStefano Zampini } 663628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 663728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 663828143c3dSStefano Zampini bs = 1; 6639e7931f94SStefano Zampini } 6640e7931f94SStefano Zampini 664170cf5478SStefano Zampini /* create MATIS object if needed */ 664257de7509SStefano Zampini if (!reuse) { 6643e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 6644e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 664570cf5478SStefano Zampini } else { 664670cf5478SStefano Zampini /* it also destroys the local matrices */ 664757de7509SStefano Zampini if (*mat_n) { 664870cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 664957de7509SStefano Zampini } else { /* this is a fake object */ 665057de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 665157de7509SStefano Zampini } 665270cf5478SStefano Zampini } 665370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 6654e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 66559d30be91SStefano Zampini 66569d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 66579d30be91SStefano Zampini 66589d30be91SStefano Zampini /* Global to local map of received indices */ 66599d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 66609d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 66619d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 66629d30be91SStefano Zampini 66639d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 66649d30be91SStefano Zampini buf_size_idxs = 0; 66659d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 66669d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 66679d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 66689d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 66699d30be91SStefano Zampini } 66709d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 66719d30be91SStefano Zampini 66729d30be91SStefano Zampini /* set preallocation */ 66739d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 66749d30be91SStefano Zampini if (!newisdense) { 66759d30be91SStefano Zampini PetscInt *new_local_nnz=0; 66769d30be91SStefano Zampini 66779d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 66789d30be91SStefano Zampini if (n_recvs) { 66799d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 66809d30be91SStefano Zampini } 66819d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 66829d30be91SStefano Zampini PetscInt j; 66839d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 66849d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 66859d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 66869d30be91SStefano Zampini } 66879d30be91SStefano Zampini } else { 66889d30be91SStefano Zampini /* TODO */ 66899d30be91SStefano Zampini } 66909d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 66919d30be91SStefano Zampini } 66929d30be91SStefano Zampini if (new_local_nnz) { 66939d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 66949d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 66959d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 66969d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 66979d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 66989d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 66999d30be91SStefano Zampini } else { 67009d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 67019d30be91SStefano Zampini } 67029d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 67039d30be91SStefano Zampini } else { 67049d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 67059d30be91SStefano Zampini } 6706e7931f94SStefano Zampini 6707e7931f94SStefano Zampini /* set values */ 6708e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 67099d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 6710e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6711e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 6712e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 67139d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 6714e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 6715e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 6716e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 671728143c3dSStefano Zampini } else { 671828143c3dSStefano Zampini /* TODO */ 6719e7931f94SStefano Zampini } 6720e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6721e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 6722e7931f94SStefano Zampini } 6723e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6724e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 672570cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 672670cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 67279d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 6728e7931f94SStefano Zampini 6729dfd14d43SStefano Zampini #if 0 673028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 6731e7931f94SStefano Zampini Vec lvec,rvec; 6732e7931f94SStefano Zampini PetscReal infty_error; 6733e7931f94SStefano Zampini 67342a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 6735e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 6736e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 6737e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 673870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 6739e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 6740e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 6741e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 6742e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 6743e7931f94SStefano Zampini } 674428143c3dSStefano Zampini #endif 6745e7931f94SStefano Zampini 674628143c3dSStefano Zampini /* assemble new additional is (if any) */ 674728143c3dSStefano Zampini if (nis) { 674828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 674928143c3dSStefano Zampini 675028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6751854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 675228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 675328143c3dSStefano Zampini psum = 0; 675428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 675528143c3dSStefano Zampini for (j=0;j<nis;j++) { 675628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 675728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 675828143c3dSStefano Zampini psum += plen; 675928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 676028143c3dSStefano Zampini } 676128143c3dSStefano Zampini } 6762854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 6763854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 676428143c3dSStefano Zampini for (i=1;i<nis;i++) { 676528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 676628143c3dSStefano Zampini } 676728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 676828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 676928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 677028143c3dSStefano Zampini for (j=0;j<nis;j++) { 677128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 677228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 677328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 677428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 677528143c3dSStefano Zampini } 677628143c3dSStefano Zampini } 677728143c3dSStefano Zampini for (i=0;i<nis;i++) { 677828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 677928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 678028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 678128143c3dSStefano Zampini } 678228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 678328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 678428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 678528143c3dSStefano Zampini } 6786e7931f94SStefano Zampini /* free workspace */ 678728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 6788e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6789e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 6790e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6791e7931f94SStefano Zampini if (isdense) { 6792e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6793e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 6794e7931f94SStefano Zampini } else { 6795e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 6796e7931f94SStefano Zampini } 679728143c3dSStefano Zampini if (nis) { 679828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 679928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 680028143c3dSStefano Zampini } 68011ae86dd6SStefano Zampini 68021ae86dd6SStefano Zampini if (nvecs) { 68031ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 68041ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 68051ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 68061ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 68071ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 68081ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 68091ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 68101ae86dd6SStefano Zampini /* set values */ 68111ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 68121ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 68131ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 68141ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 68151ae86dd6SStefano Zampini PetscInt j; 68161ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 68171ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 68181ae86dd6SStefano Zampini } 68191ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 68201ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 68211ae86dd6SStefano Zampini } 68221ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 68231ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 68241ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 68251ae86dd6SStefano Zampini } 68261ae86dd6SStefano Zampini 68271ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 68281ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 6829e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 6830e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 68311ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 683228143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 6833e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 6834e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 68351ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 683628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 6837e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 6838e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 6839e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 6840e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 6841e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 684228143c3dSStefano Zampini if (nis) { 684328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 684428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 684528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 684628143c3dSStefano Zampini } 684728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 684828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 684928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 685028143c3dSStefano Zampini for (i=0;i<nis;i++) { 685128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 685228143c3dSStefano Zampini } 68531ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 68541ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 68551ae86dd6SStefano Zampini } 685653a05cb3SStefano Zampini *mat_n = NULL; 685728143c3dSStefano Zampini } 6858e7931f94SStefano Zampini PetscFunctionReturn(0); 6859e7931f94SStefano Zampini } 6860a57a6d2fSStefano Zampini 686112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 6862af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 686312edc857SStefano Zampini 6864c8587f34SStefano Zampini #undef __FUNCT__ 6865c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 6866c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 6867c8587f34SStefano Zampini { 6868c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 6869c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 687020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 68711ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 68729881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 687320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 68746e683305SStefano Zampini IS coarse_is,*isarray; 68756e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 687630368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 6877f9eb5b7dSStefano Zampini PC pc_temp; 6878c8587f34SStefano Zampini PCType coarse_pc_type; 6879c8587f34SStefano Zampini KSPType coarse_ksp_type; 6880f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 68814f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 68826e683305SStefano Zampini Mat t_coarse_mat_is; 688357de7509SStefano Zampini PetscInt ncoarse; 688468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 688522bc73bbSStefano Zampini PetscScalar *array; 688657de7509SStefano Zampini MatReuse coarse_mat_reuse; 688757de7509SStefano Zampini PetscBool restr, full_restr, have_void; 68889881197aSStefano Zampini PetscErrorCode ierr; 6889fdc09c96SStefano Zampini 6890c8587f34SStefano Zampini PetscFunctionBegin; 6891c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 689268457ee5SStefano 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 */ 6893fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 68945a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 6895fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 6896f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 6897f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 6898f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 6899fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 690051bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 690151bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 6902dc4bcba2SStefano Zampini PC pc; 6903dc4bcba2SStefano Zampini PetscBool isbddc; 6904dc4bcba2SStefano Zampini 6905dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 6906dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 6907dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 6908dc4bcba2SStefano Zampini if (isbddc) { 690963c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 691063c961adSStefano Zampini } else { 6911727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 691263c961adSStefano Zampini } 6913fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 6914fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 6915fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 6916f4ddd8eeSStefano Zampini } 6917fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 6918fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 6919f4ddd8eeSStefano Zampini } 692070cf5478SStefano Zampini /* reset any subassembling information */ 692157de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 692270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 692357de7509SStefano Zampini } 69246e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 6925fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 6926f4ddd8eeSStefano Zampini } 692757de7509SStefano Zampini /* assemble coarse matrix */ 692857de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 692957de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 693057de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 693157de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 693218a45a71SStefano Zampini } else { 693357de7509SStefano Zampini coarse_mat = NULL; 693457de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 69356e683305SStefano Zampini } 6936e7931f94SStefano Zampini 6937abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 6938abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 6939abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 6940abbbba34SStefano Zampini 6941abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 694222bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 694322bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 694422bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 694522bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 6946e176bc59SStefano 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); 69476e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 69486e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 69496e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6950abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 6951abbbba34SStefano Zampini 695257de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 695357de7509SStefano Zampini im_active = !!(pcis->n); 695457de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 695557de7509SStefano Zampini 695614f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 695757de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 695857de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 695957de7509SStefano Zampini coarse_mat_is = NULL; 696057de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 696157de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 69621ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 696357de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 696457de7509SStefano Zampini if (multilevel_requested) { 696557de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 696657de7509SStefano Zampini restr = PETSC_FALSE; 696757de7509SStefano Zampini full_restr = PETSC_FALSE; 696857de7509SStefano Zampini } else { 696957de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 697057de7509SStefano Zampini restr = PETSC_TRUE; 697157de7509SStefano Zampini full_restr = PETSC_TRUE; 697257de7509SStefano Zampini } 69734b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 697457de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 697557de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 6976a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 697737eada30SStefano Zampini ierr = MatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 6978a198735bSStefano Zampini } else { 6979a198735bSStefano Zampini PetscMPIInt size,rank; 6980a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 6981a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 6982a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 6983a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 6984a198735bSStefano Zampini } 698557de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 698657de7509SStefano Zampini PetscInt psum; 698757de7509SStefano Zampini PetscMPIInt size; 698857de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 698957de7509SStefano Zampini else psum = 0; 699057de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 699157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 699257de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 699357de7509SStefano Zampini } 699457de7509SStefano Zampini /* determine if we can go multilevel */ 699557de7509SStefano Zampini if (multilevel_requested) { 699657de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 699757de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 699857de7509SStefano Zampini } 699957de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 700057de7509SStefano Zampini 7001e4d548c7SStefano Zampini /* dump subassembling pattern */ 7002e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 7003e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 7004e4d548c7SStefano Zampini } 7005e4d548c7SStefano Zampini 70066e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 7007c703fcc7SStefano Zampini if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 70086e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 70096e683305SStefano Zampini const PetscInt *idxs; 70106e683305SStefano Zampini ISLocalToGlobalMapping tmap; 70116e683305SStefano Zampini 70126e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 70130be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 70146e683305SStefano Zampini /* allocate space for temporary storage */ 7015854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 7016854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 70176e683305SStefano Zampini /* allocate for IS array */ 70186e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 70196e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 702027b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 702130368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 7022854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 70236e683305SStefano Zampini /* dofs splitting */ 70246e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 70256e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 70266e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 70276e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 70286e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 70296e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 70306e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 703130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 70326e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 70336e683305SStefano Zampini } 70346e683305SStefano Zampini /* neumann boundaries */ 70356e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 70366e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 70376e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 70386e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 70396e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 70406e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 70416e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 704230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 70436e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 70446e683305SStefano Zampini } 70456e683305SStefano Zampini /* free memory */ 70466e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 70476e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 70486e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 70496e683305SStefano Zampini } else { 70506e683305SStefano Zampini nis = 0; 70516e683305SStefano Zampini nisdofs = 0; 70526e683305SStefano Zampini nisneu = 0; 705330368db7SStefano Zampini nisvert = 0; 70546e683305SStefano Zampini isarray = NULL; 70556e683305SStefano Zampini } 70566e683305SStefano Zampini /* destroy no longer needed map */ 70576e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 70586e683305SStefano Zampini 705957de7509SStefano Zampini /* subassemble */ 706057de7509SStefano Zampini if (multilevel_allowed) { 70611ae86dd6SStefano Zampini Vec vp[1]; 70621ae86dd6SStefano Zampini PetscInt nvecs = 0; 706357de7509SStefano Zampini PetscBool reuse,reuser; 70641ae86dd6SStefano Zampini 706557de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 706657de7509SStefano Zampini else reuse = PETSC_FALSE; 706757de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 70681ae86dd6SStefano Zampini vp[0] = NULL; 70691ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 70701ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 70711ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 70721ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 70731ae86dd6SStefano Zampini nvecs = 1; 70741ae86dd6SStefano Zampini 70751ae86dd6SStefano Zampini if (pcbddc->divudotp) { 7076a198735bSStefano Zampini Mat B,loc_divudotp; 70771ae86dd6SStefano Zampini Vec v,p; 70781ae86dd6SStefano Zampini IS dummy; 70791ae86dd6SStefano Zampini PetscInt np; 70801ae86dd6SStefano Zampini 7081a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 7082a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 70831ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 7084a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 70851ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 70861ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 70871ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 70881ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 70891ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 70901ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 70911ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 70921ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 70931ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 70941ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 70951ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 70961ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 709774e2c79eSStefano Zampini } 70981ae86dd6SStefano Zampini } 70991ae86dd6SStefano Zampini if (reuser) { 71001ae86dd6SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 710174e2c79eSStefano Zampini } else { 71021ae86dd6SStefano 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); 71031ae86dd6SStefano Zampini } 71041ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 71051ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 71061ae86dd6SStefano Zampini PetscInt nl; 71071ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 71081ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 71091ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 71101ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 71111ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 71121ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 71131ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 71141ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 7115a198735bSStefano Zampini } else { 7116a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 71171ae86dd6SStefano Zampini } 71181ae86dd6SStefano Zampini } else { 71191ae86dd6SStefano 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); 71206e683305SStefano Zampini } 712157de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 712257de7509SStefano Zampini PetscMPIInt size; 7123f913dca9SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size);CHKERRQ(ierr); 712457de7509SStefano Zampini if (!multilevel_allowed) { 712557de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 71266e683305SStefano Zampini } else { 712757de7509SStefano Zampini Mat A; 7128779c1cceSStefano Zampini 712957de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 713057de7509SStefano Zampini if (coarse_mat_is) { 713157de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 713257de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 713357de7509SStefano Zampini coarse_mat = coarse_mat_is; 713457de7509SStefano Zampini } 713557de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 713657de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 713757de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 7138779c1cceSStefano Zampini } 7139779c1cceSStefano Zampini } 714057de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 714157de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 71426e683305SStefano Zampini 71436e683305SStefano Zampini /* create local to global scatters for coarse problem */ 714468457ee5SStefano Zampini if (compute_vecs) { 71456e683305SStefano Zampini PetscInt lrows; 71466e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 714757de7509SStefano Zampini if (coarse_mat) { 714857de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 71496e683305SStefano Zampini } else { 71506e683305SStefano Zampini lrows = 0; 71516e683305SStefano Zampini } 71526e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 71536e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 71546e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 71556e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 71566e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 71576e683305SStefano Zampini } 71586e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 7159c8587f34SStefano Zampini 7160f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 7161f9eb5b7dSStefano Zampini if (multilevel_allowed) { 7162f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 7163f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 7164f9eb5b7dSStefano Zampini } else { 7165f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 7166f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 7167c8587f34SStefano Zampini } 7168c8587f34SStefano Zampini 71696e683305SStefano Zampini /* print some info if requested */ 71706e683305SStefano Zampini if (pcbddc->dbg_flag) { 71716e683305SStefano Zampini if (!multilevel_allowed) { 71726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 71736e683305SStefano Zampini if (multilevel_requested) { 71746e683305SStefano 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); 71756e683305SStefano Zampini } else if (pcbddc->max_levels) { 71766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 71776e683305SStefano Zampini } 71786e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 71796e683305SStefano Zampini } 71806e683305SStefano Zampini } 71816e683305SStefano Zampini 7182f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 718357de7509SStefano Zampini if (coarse_mat) { 71846a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 71856e683305SStefano Zampini if (pcbddc->dbg_flag) { 718657de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 71876e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 71886e683305SStefano Zampini } 7189f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 7190312be037SStefano Zampini char prefix[256],str_level[16]; 7191e604994aSStefano Zampini size_t len; 719257de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 7193422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 7194c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 7195f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 719657de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7197c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 71986e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 7199c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 7200c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7201e604994aSStefano Zampini /* prefix */ 7202e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 7203e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 7204e604994aSStefano Zampini if (!pcbddc->current_level) { 7205e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 7206e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 7207c8587f34SStefano Zampini } else { 7208e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 7209312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 7210312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 721134d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 7212312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 7213e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 7214e604994aSStefano Zampini } 7215e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 72163e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 72173e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 72183e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 72193e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 7220f9eb5b7dSStefano Zampini /* allow user customization */ 7221f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 72223e3c6dadSStefano Zampini } 72233e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 722451bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 72253e3c6dadSStefano Zampini if (nisdofs) { 72263e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 72273e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 72283e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 72293e3c6dadSStefano Zampini } 72303e3c6dadSStefano Zampini } 72313e3c6dadSStefano Zampini if (nisneu) { 72323e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 72333e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 7234312be037SStefano Zampini } 723530368db7SStefano Zampini if (nisvert) { 723630368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 723730368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 723830368db7SStefano Zampini } 7239f9eb5b7dSStefano Zampini 7240f9eb5b7dSStefano Zampini /* get some info after set from options */ 7241f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 7242f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 72434f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 72446e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 7245f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7246f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 7247f9eb5b7dSStefano Zampini } 724839f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 72494f3a063dSStefano Zampini if (isredundant) { 72504f3a063dSStefano Zampini KSP inner_ksp; 72514f3a063dSStefano Zampini PC inner_pc; 72524f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 72534f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 72544f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 72554f3a063dSStefano Zampini } 7256f9eb5b7dSStefano Zampini 725757de7509SStefano Zampini /* parameters which miss an API */ 725857de7509SStefano Zampini if (isbddc) { 7259720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 7260720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 726157de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 726227b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 726327b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 7264a198735bSStefano Zampini Mat coarsedivudotp_is; 7265a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 7266a198735bSStefano Zampini IS row,col; 7267a198735bSStefano Zampini const PetscInt *gidxs; 7268a198735bSStefano Zampini PetscInt n,st,M,N; 7269a198735bSStefano Zampini 7270a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 7271a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 7272a198735bSStefano Zampini st = st-n; 7273a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 7274a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 7275a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 7276a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7277a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 7278a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7279a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 7280a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 7281a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 7282a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 7283a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 7284a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 7285a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 7286a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 7287a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 7288a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 7289a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 7290a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 7291a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 7292a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 72938ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 7294a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 7295720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 729659e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 7297720d30f9SStefano Zampini } 7298d4d8cf7bSStefano Zampini } 72999881197aSStefano Zampini 73003301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 73015a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 73023301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 73033301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 73043301b35fSStefano Zampini } 73053301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 73063301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 73073301b35fSStefano Zampini } 73083301b35fSStefano Zampini if (pc->pmat->spd_set) { 73093301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 73103301b35fSStefano Zampini } 731127b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 731227b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 731327b6a85dSStefano Zampini } 73146e683305SStefano Zampini /* set operators */ 73155f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 73166e683305SStefano Zampini if (pcbddc->dbg_flag) { 73176e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 73186e683305SStefano Zampini } 73196e683305SStefano Zampini } 73206e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 7321b1ecc7b1SStefano Zampini #if 0 7322b9b85e73SStefano Zampini { 7323b9b85e73SStefano Zampini PetscViewer viewer; 7324b9b85e73SStefano Zampini char filename[256]; 7325b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 7326b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 73276a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 7328b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 7329f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 7330b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 7331b9b85e73SStefano Zampini } 7332b9b85e73SStefano Zampini #endif 7333f9eb5b7dSStefano Zampini 733498a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 733598a51de6SStefano Zampini Vec crhs,csol; 733604708bb6SStefano Zampini 7337f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 7338f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 7339f347579bSStefano Zampini if (!csol) { 73402a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 7341f9eb5b7dSStefano Zampini } 7342f347579bSStefano Zampini if (!crhs) { 73432a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 7344f347579bSStefano Zampini } 7345b0f5fe93SStefano Zampini } 73461ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 7347b0f5fe93SStefano Zampini 7348b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 7349b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 7350b0f5fe93SStefano Zampini 7351b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 73524f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 73534f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 73544f1b2e48SStefano Zampini } 7355b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 7356b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 7357b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7358b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7359b0f5fe93SStefano Zampini if (coarse_mat) { 7360b0f5fe93SStefano Zampini Vec nullv; 7361b0f5fe93SStefano Zampini PetscScalar *array,*array2; 7362b0f5fe93SStefano Zampini PetscInt nl; 7363b0f5fe93SStefano Zampini 7364b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 7365b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 7366b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7367b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 7368b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 7369b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 7370b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7371b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 7372b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 7373b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 7374b0f5fe93SStefano Zampini } 7375b0f5fe93SStefano Zampini } 7376b0f5fe93SStefano Zampini 7377b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 7378b0f5fe93SStefano Zampini PetscBool ispreonly; 7379b0f5fe93SStefano Zampini 7380b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7381b0f5fe93SStefano Zampini PetscBool isnull; 7382b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 7383bef83e63SStefano Zampini if (isnull) { 7384b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 7385b0f5fe93SStefano Zampini } 7386bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 7387b0f5fe93SStefano Zampini } 7388b0f5fe93SStefano Zampini /* setup coarse ksp */ 7389b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 7390cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 7391cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 73926e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 7393c8587f34SStefano Zampini KSP check_ksp; 73942b510759SStefano Zampini KSPType check_ksp_type; 7395c8587f34SStefano Zampini PC check_pc; 73966e683305SStefano Zampini Vec check_vec,coarse_vec; 73976a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 73982b510759SStefano Zampini PetscInt its; 73996e683305SStefano Zampini PetscBool compute_eigs; 74006e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 74016e683305SStefano Zampini PetscInt neigs; 74028e185a42SStefano Zampini const char *prefix; 7403c8587f34SStefano Zampini 74042b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 74056e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 7406422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 740723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7408f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 7409e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 7410e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 7411e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 74122b510759SStefano Zampini if (ispreonly) { 74132b510759SStefano Zampini check_ksp_type = KSPPREONLY; 74146e683305SStefano Zampini compute_eigs = PETSC_FALSE; 74152b510759SStefano Zampini } else { 7416cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 74176e683305SStefano Zampini compute_eigs = PETSC_TRUE; 7418c8587f34SStefano Zampini } 7419c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 74206e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 74216e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 74226e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 7423a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 7424a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 7425a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 7426a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 7427c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 7428c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 7429c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 7430c8587f34SStefano Zampini /* create random vec */ 74312701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 7432c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 74336e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 7434c8587f34SStefano Zampini /* solve coarse problem */ 74356e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 7436cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 74376e683305SStefano Zampini if (compute_eigs) { 7438854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 7439854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 74406e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 74411ae86dd6SStefano Zampini if (neigs) { 74426e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 74436e683305SStefano Zampini lambda_min = eigs_r[0]; 74446e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 74452701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 74462701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 7447cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 7448cbcc2c2aSStefano Zampini } 7449c8587f34SStefano Zampini } 7450c8587f34SStefano Zampini } 74511ae86dd6SStefano Zampini } 7452cbcc2c2aSStefano Zampini 7453c8587f34SStefano Zampini /* check coarse problem residual error */ 74546e683305SStefano Zampini if (pcbddc->dbg_flag) { 74556e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 74566e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 74576e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 7458c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 74596e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 74606e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 7461779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 74626e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 74636e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 74646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 74656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 7466b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7467b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 7468b0f5fe93SStefano Zampini } 74696e683305SStefano Zampini if (compute_eigs) { 74706e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 7471b03ebc13SStefano Zampini KSPConvergedReason reason; 7472deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 7473c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 7474b03ebc13SStefano Zampini ierr = KSPGetConvergedReason(check_ksp,&reason);CHKERRQ(ierr); 74756e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 7476b03ebc13SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem eigenvalues (estimated with %d iterations of %s, conv reason %d): %1.6e %1.6e (%1.6e %1.6e)\n",its,check_ksp_type,reason,lambda_min,lambda_max,lambda_min_s,lambda_max_s);CHKERRQ(ierr); 74776e683305SStefano Zampini for (i=0;i<neigs;i++) { 74786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 7479c8587f34SStefano Zampini } 74806e683305SStefano Zampini } 74816e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 74826e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 74836e683305SStefano Zampini } 7484e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 74852701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 7486c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 74876e683305SStefano Zampini if (compute_eigs) { 74886e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 74896e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 7490c8587f34SStefano Zampini } 74916e683305SStefano Zampini } 74926e683305SStefano Zampini } 7493bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 7494cbcc2c2aSStefano Zampini /* print additional info */ 7495cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 74966e683305SStefano Zampini /* waits until all processes reaches this point */ 74976e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 7498cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 7499cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7500cbcc2c2aSStefano Zampini } 7501cbcc2c2aSStefano Zampini 75022b510759SStefano Zampini /* free memory */ 7503fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 7504c8587f34SStefano Zampini PetscFunctionReturn(0); 7505c8587f34SStefano Zampini } 7506674ae819SStefano Zampini 7507f34684f1SStefano Zampini #undef __FUNCT__ 7508f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 7509f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 7510f34684f1SStefano Zampini { 7511f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7512f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 7513f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 7514dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 7515dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 751673be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 7517dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 7518f34684f1SStefano Zampini PetscErrorCode ierr; 7519f34684f1SStefano Zampini 7520f34684f1SStefano Zampini PetscFunctionBegin; 7521f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 75226c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 7523dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 75243bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 7525dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7526dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 75276583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 7528dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 7529dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 7530dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 75316c4ed002SBarry 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); 7532dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 7533dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7534dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 7535dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7536dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7537f34684f1SStefano Zampini 7538f34684f1SStefano Zampini /* check numbering */ 7539f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 7540019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 7541dc456d91SStefano Zampini PetscInt i; 7542b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 7543f34684f1SStefano Zampini 7544f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7545f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 7546f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 75471575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7548019a44ceSStefano Zampini /* counter */ 7549019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7550019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 7551019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7552019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7553019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7554019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7555f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 7556f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 7557727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 7558f34684f1SStefano Zampini } 7559f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7560f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7561f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7562e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7563e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7564e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7565e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7566f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7567019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 7568f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 7569019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 75702c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 757175c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 7572b9b85e73SStefano Zampini set_error = PETSC_TRUE; 75732c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 75742c66d082SStefano 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); 7575f34684f1SStefano Zampini } 7576f34684f1SStefano Zampini } 7577019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 7578b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 7579f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7580f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 7581f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 7582f34684f1SStefano Zampini } 7583f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7584f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7585e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7586e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7587f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 7588f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 7589b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 7590ca8b9ea9SStefano Zampini PetscInt *gidxs; 7591ca8b9ea9SStefano Zampini 7592ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 75933bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 7594f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 7595f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7596f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 7597f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 75984bc2dc4bSStefano 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); 7599f34684f1SStefano Zampini } 7600f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7601ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 7602f34684f1SStefano Zampini } 7603f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 76041575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7605302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 7606f34684f1SStefano Zampini } 76078bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 7608f34684f1SStefano Zampini /* get back data */ 7609f34684f1SStefano Zampini *coarse_size_n = coarse_size; 7610f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 7611674ae819SStefano Zampini PetscFunctionReturn(0); 7612674ae819SStefano Zampini } 7613674ae819SStefano Zampini 7614e456f2a8SStefano Zampini #undef __FUNCT__ 7615e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 7616a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 7617e456f2a8SStefano Zampini { 7618e456f2a8SStefano Zampini IS localis_t; 7619a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 7620e456f2a8SStefano Zampini PetscScalar *vals; 7621e456f2a8SStefano Zampini PetscErrorCode ierr; 7622e456f2a8SStefano Zampini 7623e456f2a8SStefano Zampini PetscFunctionBegin; 7624a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 7625e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 7626854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 7627e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 7628e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 7629a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 7630a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 76311035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 7632a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 76331035eff8SStefano Zampini } 7634a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 7635e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 7636e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 7637a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 7638a7dc3881SStefano Zampini /* now compute set in local ordering */ 7639a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7640a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7641a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 7642a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 7643a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 7644ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 7645e456f2a8SStefano Zampini lsize++; 7646e456f2a8SStefano Zampini } 7647e456f2a8SStefano Zampini } 7648854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 7649a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 7650ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 7651e456f2a8SStefano Zampini idxs[lsize++] = i; 7652e456f2a8SStefano Zampini } 7653e456f2a8SStefano Zampini } 7654a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 7655a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 7656e456f2a8SStefano Zampini *localis = localis_t; 7657e456f2a8SStefano Zampini PetscFunctionReturn(0); 7658e456f2a8SStefano Zampini } 7659906d46d4SStefano Zampini 7660b96c3477SStefano Zampini #undef __FUNCT__ 7661b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 766208122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 7663b96c3477SStefano Zampini { 7664a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 7665b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 7666b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 7667a64f4aa4SStefano Zampini Mat S_j; 7668b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 7669b96c3477SStefano Zampini PetscBool free_used_adj; 7670b96c3477SStefano Zampini PetscErrorCode ierr; 7671b96c3477SStefano Zampini 7672b96c3477SStefano Zampini PetscFunctionBegin; 7673b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 7674b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 767508122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 7676b96c3477SStefano Zampini used_xadj = NULL; 7677b96c3477SStefano Zampini used_adjncy = NULL; 7678b96c3477SStefano Zampini } else { 767908122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 768008122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 768108122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 768208122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 7683b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 7684b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 7685b96c3477SStefano Zampini } else { 76862fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 7687b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 7688b96c3477SStefano Zampini PetscInt nvtxs; 7689b96c3477SStefano Zampini 76902fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 76912fffb893SStefano Zampini if (flg_row) { 7692b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 7693b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 7694b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 7695b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 76962fffb893SStefano Zampini } else { 76972fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 76982fffb893SStefano Zampini used_xadj = NULL; 76992fffb893SStefano Zampini used_adjncy = NULL; 77002fffb893SStefano Zampini } 77012fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 7702b96c3477SStefano Zampini } 7703b96c3477SStefano Zampini } 7704d5574798SStefano Zampini 7705d5574798SStefano Zampini /* setup sub_schurs data */ 7706a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 7707df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 7708df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 7709a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 771091af6908SStefano 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); 7711a64f4aa4SStefano Zampini } else { 77126816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 77134d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 7714a3df083aSStefano Zampini PetscInt benign_n; 771572b8c272SStefano Zampini Mat change = NULL; 77169d54b7f4SStefano Zampini Vec scaling = NULL; 771772b8c272SStefano Zampini IS change_primal = NULL; 7718a3df083aSStefano Zampini 77195feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 77205feab87aSStefano Zampini PetscInt n_vertices; 77215feab87aSStefano Zampini 77225feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 77232034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 77245feab87aSStefano Zampini } 772504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 772604708bb6SStefano Zampini if (!isseqaij) { 772704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 772804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 772904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 773004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 773104708bb6SStefano Zampini } else { 7732511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 773304708bb6SStefano Zampini } 773404708bb6SStefano Zampini } 7735a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 7736a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 7737ca92afb2SStefano Zampini } else { 7738a3df083aSStefano Zampini benign_n = 0; 7739ca92afb2SStefano Zampini } 7740b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 7741b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 7742b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 774372b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 774422db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 7745b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 774622db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 7747b7ab4a40SStefano Zampini } 7748b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 7749b7ab4a40SStefano 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 */ 7750b7ab4a40SStefano Zampini if (need_change) { 775188c03ad3SStefano Zampini PC_IS *pcisf; 775288c03ad3SStefano Zampini PC_BDDC *pcbddcf; 775388c03ad3SStefano Zampini PC pcf; 775488c03ad3SStefano Zampini 7755e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 775688c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 775788c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 775888c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 775988c03ad3SStefano Zampini /* hacks */ 776088c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 776172b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 776272b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 776372b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 776472b8c272SStefano Zampini pcisf->n = pcis->n; 776572b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 776688c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 776788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 776888c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 776988c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 777088c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 777188c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 777272b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 777388c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 777488c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 777572b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 777672b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 777772b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 777872b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 777972b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 778088c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 778172b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 778288c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 778388c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 778488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 778588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 778688c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 778788c03ad3SStefano Zampini pcf->ops->destroy = NULL; 778888c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 778988c03ad3SStefano Zampini } 77909d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 779191af6908SStefano 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); 779272b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 779372b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 7794ca92afb2SStefano Zampini } 7795d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 7796b96c3477SStefano Zampini 7797b96c3477SStefano Zampini /* free adjacency */ 7798b96c3477SStefano Zampini if (free_used_adj) { 7799b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 7800b96c3477SStefano Zampini } 7801b96c3477SStefano Zampini PetscFunctionReturn(0); 7802b96c3477SStefano Zampini } 7803b96c3477SStefano Zampini 7804b96c3477SStefano Zampini #undef __FUNCT__ 7805b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 780608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 7807b96c3477SStefano Zampini { 7808b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 7809b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 7810b96c3477SStefano Zampini PCBDDCGraph graph; 7811b96c3477SStefano Zampini PetscErrorCode ierr; 7812b96c3477SStefano Zampini 7813b96c3477SStefano Zampini PetscFunctionBegin; 7814b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 781508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 78163301b35fSStefano Zampini IS verticesIS,verticescomm; 78173301b35fSStefano Zampini PetscInt vsize,*idxs; 7818b96c3477SStefano Zampini 7819b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 78203301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 78213301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 78223301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 78233301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 7824c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 7825b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 7826be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 7827441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 78283301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 7829b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 7830b96c3477SStefano Zampini } else { 7831b96c3477SStefano Zampini graph = pcbddc->mat_graph; 7832b96c3477SStefano Zampini } 7833e4d548c7SStefano Zampini /* print some info */ 78345c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 7835e4d548c7SStefano Zampini IS vertices; 7836e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 7837c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 7838e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 7839e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 7840e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7841e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 7842e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 7843e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 7844e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 7845e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7846e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7847c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 7848e4d548c7SStefano Zampini } 7849b96c3477SStefano Zampini 7850b96c3477SStefano Zampini /* sub_schurs init */ 7851b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 7852b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 7853b334f244SStefano Zampini } 78548b6046baSStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 7855a64f4aa4SStefano Zampini 7856b96c3477SStefano Zampini /* free graph struct */ 785708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 7858b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 7859b96c3477SStefano Zampini } 7860b96c3477SStefano Zampini PetscFunctionReturn(0); 7861b96c3477SStefano Zampini } 7862fa34dd3eSStefano Zampini 7863fa34dd3eSStefano Zampini #undef __FUNCT__ 7864fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 7865fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 7866fa34dd3eSStefano Zampini { 7867fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 7868fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 7869fa34dd3eSStefano Zampini PetscErrorCode ierr; 7870fa34dd3eSStefano Zampini 7871fa34dd3eSStefano Zampini PetscFunctionBegin; 7872fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 7873fa34dd3eSStefano Zampini IS zerodiag = NULL; 78744f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 7875fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 78764f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 787775c01103SStefano Zampini PetscReal norm; 7878fa34dd3eSStefano Zampini PetscInt i; 7879fa34dd3eSStefano Zampini 7880fa34dd3eSStefano Zampini /* B0 and B0_B */ 7881fa34dd3eSStefano Zampini if (zerodiag) { 7882fa34dd3eSStefano Zampini IS dummy; 7883fa34dd3eSStefano Zampini 78844f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 78854f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 7886fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 7887fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 7888fa34dd3eSStefano Zampini } 7889fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 7890fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 7891fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 7892fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7893fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7894fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7895fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7896fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 7897fa34dd3eSStefano Zampini /* S_j */ 7898fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 7899fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 7900fa34dd3eSStefano Zampini 7901fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 7902fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 7903fa34dd3eSStefano Zampini /* continuous in primal space */ 7904fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 7905fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7906fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7907fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 79084f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 79094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 7910fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 7911fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 7912fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7913fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7914fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7915fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7916fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 7917fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 7918fa34dd3eSStefano Zampini 7919fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 7920fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 7921fa34dd3eSStefano Zampini /* local with Schur */ 7922fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 7923fa34dd3eSStefano Zampini if (zerodiag) { 7924fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 79254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 7926fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 7927fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 7928fa34dd3eSStefano Zampini } 7929fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 7930fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7931fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7932fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7933fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 7934fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 7935fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 7936fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7937fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 7938fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7939fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7940fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7941fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7942fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 7943fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 7944fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 7945fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 7946fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 7947fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7948fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7949fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7950fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7951fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 7952fa34dd3eSStefano Zampini if (zerodiag) { 7953fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 7954fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 79554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 7956fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 7957fa34dd3eSStefano Zampini } 7958fa34dd3eSStefano Zampini /* BDDC */ 7959fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 7960fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 7961fa34dd3eSStefano Zampini 7962fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 7963fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 7964fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 7965fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 79664f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 79674f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 7968fa34dd3eSStefano Zampini } 79694f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 7970fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 7971fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 7972fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 7973fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 7974fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 7975fa34dd3eSStefano Zampini } 7976fa34dd3eSStefano Zampini PetscFunctionReturn(0); 7977fa34dd3eSStefano Zampini } 7978