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> 4837cedc9SStefano Zampini #include <../src/mat/impls/dense/seq/dense.h> 5c80a6c00SStefano Zampini #include <petscdmplex.h> 6674ae819SStefano Zampini #include <petscblaslapack.h> 7daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 8c80a6c00SStefano Zampini #include <petsc/private/dmpleximpl.h> 97620a527SStefano Zampini #include <petscdmda.h> 10674ae819SStefano Zampini 111e0482f5SStefano Zampini static PetscErrorCode MatMPIAIJRestrict(Mat,MPI_Comm,Mat*); 121e0482f5SStefano Zampini 13f498cd09SStefano Zampini /* if range is true, it returns B s.t. span{B} = range(A) 14f498cd09SStefano Zampini if range is false, it returns B s.t. range(B) _|_ range(A) */ 15f498cd09SStefano Zampini PetscErrorCode MatDenseOrthogonalRangeOrComplement(Mat A, PetscBool range, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B) 16a13144ffSStefano Zampini { 17a13144ffSStefano Zampini PetscScalar *uwork,*data,*U, ds = 0.; 18a13144ffSStefano Zampini PetscReal *sing; 19a13144ffSStefano Zampini PetscBLASInt bM,bN,lwork,lierr,di = 1; 20a13144ffSStefano Zampini PetscInt ulw,i,nr,nc,n; 21a13144ffSStefano Zampini PetscErrorCode ierr; 22abee2b68SSebastian Grimberg #if defined(PETSC_USE_COMPLEX) 23abee2b68SSebastian Grimberg PetscReal *rwork2; 24abee2b68SSebastian Grimberg #endif 25a13144ffSStefano Zampini 26a13144ffSStefano Zampini PetscFunctionBegin; 27a13144ffSStefano Zampini ierr = MatGetSize(A,&nr,&nc);CHKERRQ(ierr); 28a13144ffSStefano Zampini if (!nr || !nc) PetscFunctionReturn(0); 29a13144ffSStefano Zampini 30a13144ffSStefano Zampini /* workspace */ 31a13144ffSStefano Zampini if (!work) { 32a13144ffSStefano Zampini ulw = PetscMax(PetscMax(1,5*PetscMin(nr,nc)),3*PetscMin(nr,nc)+PetscMax(nr,nc)); 33f913dca9SStefano Zampini ierr = PetscMalloc1(ulw,&uwork);CHKERRQ(ierr); 34a13144ffSStefano Zampini } else { 35a13144ffSStefano Zampini ulw = lw; 36a13144ffSStefano Zampini uwork = work; 37a13144ffSStefano Zampini } 38a13144ffSStefano Zampini n = PetscMin(nr,nc); 39a13144ffSStefano Zampini if (!rwork) { 40a13144ffSStefano Zampini ierr = PetscMalloc1(n,&sing);CHKERRQ(ierr); 41a13144ffSStefano Zampini } else { 42a13144ffSStefano Zampini sing = rwork; 43a13144ffSStefano Zampini } 44a13144ffSStefano Zampini 45a13144ffSStefano Zampini /* SVD */ 46a13144ffSStefano Zampini ierr = PetscMalloc1(nr*nr,&U);CHKERRQ(ierr); 47a13144ffSStefano Zampini ierr = PetscBLASIntCast(nr,&bM);CHKERRQ(ierr); 48a13144ffSStefano Zampini ierr = PetscBLASIntCast(nc,&bN);CHKERRQ(ierr); 49a13144ffSStefano Zampini ierr = PetscBLASIntCast(ulw,&lwork);CHKERRQ(ierr); 50a13144ffSStefano Zampini ierr = MatDenseGetArray(A,&data);CHKERRQ(ierr); 51a13144ffSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 52abee2b68SSebastian Grimberg #if !defined(PETSC_USE_COMPLEX) 53a13144ffSStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,&lierr)); 54abee2b68SSebastian Grimberg #else 55abee2b68SSebastian Grimberg ierr = PetscMalloc1(5*n,&rwork2);CHKERRQ(ierr); 56abee2b68SSebastian Grimberg PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,rwork2,&lierr)); 57abee2b68SSebastian Grimberg ierr = PetscFree(rwork2);CHKERRQ(ierr); 58abee2b68SSebastian Grimberg #endif 59a13144ffSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 60a13144ffSStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 61a13144ffSStefano Zampini ierr = MatDenseRestoreArray(A,&data);CHKERRQ(ierr); 62a13144ffSStefano Zampini for (i=0;i<n;i++) if (sing[i] < PETSC_SMALL) break; 63a13144ffSStefano Zampini if (!rwork) { 64a13144ffSStefano Zampini ierr = PetscFree(sing);CHKERRQ(ierr); 65a13144ffSStefano Zampini } 66a13144ffSStefano Zampini if (!work) { 67a13144ffSStefano Zampini ierr = PetscFree(uwork);CHKERRQ(ierr); 68a13144ffSStefano Zampini } 69a13144ffSStefano Zampini /* create B */ 70f498cd09SStefano Zampini if (!range) { 71a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,nr-i,NULL,B);CHKERRQ(ierr); 72a13144ffSStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 73580bdb30SBarry Smith ierr = PetscArraycpy(data,U+nr*i,(nr-i)*nr);CHKERRQ(ierr); 74f498cd09SStefano Zampini } else { 75f498cd09SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,i,NULL,B);CHKERRQ(ierr); 76f498cd09SStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 77580bdb30SBarry Smith ierr = PetscArraycpy(data,U,i*nr);CHKERRQ(ierr); 78f498cd09SStefano Zampini } 79a13144ffSStefano Zampini ierr = MatDenseRestoreArray(*B,&data);CHKERRQ(ierr); 80a13144ffSStefano Zampini ierr = PetscFree(U);CHKERRQ(ierr); 81a13144ffSStefano Zampini PetscFunctionReturn(0); 82a13144ffSStefano Zampini } 83a13144ffSStefano Zampini 841e0482f5SStefano Zampini /* TODO REMOVE */ 851e0482f5SStefano Zampini #if defined(PRINT_GDET) 861e0482f5SStefano Zampini static int inc = 0; 871e0482f5SStefano Zampini static int lev = 0; 881e0482f5SStefano Zampini #endif 891e0482f5SStefano Zampini 901e0482f5SStefano Zampini PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, IS corners, Mat* Gins, Mat* GKins, PetscScalar cvals[2], PetscScalar *work, PetscReal *rwork) 91a13144ffSStefano Zampini { 92a13144ffSStefano Zampini PetscErrorCode ierr; 93a13144ffSStefano Zampini Mat GE,GEd; 94a13144ffSStefano Zampini PetscInt rsize,csize,esize; 95a13144ffSStefano Zampini PetscScalar *ptr; 96a13144ffSStefano Zampini 97a13144ffSStefano Zampini PetscFunctionBegin; 98a13144ffSStefano Zampini ierr = ISGetSize(edge,&esize);CHKERRQ(ierr); 99c3c0e390SStefano Zampini if (!esize) PetscFunctionReturn(0); 100a13144ffSStefano Zampini ierr = ISGetSize(extrow,&rsize);CHKERRQ(ierr); 101a13144ffSStefano Zampini ierr = ISGetSize(extcol,&csize);CHKERRQ(ierr); 102a13144ffSStefano Zampini 103a13144ffSStefano Zampini /* gradients */ 104a13144ffSStefano Zampini ptr = work + 5*esize; 1057dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,extrow,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 106a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,rsize,csize,ptr,Gins);CHKERRQ(ierr); 107a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,Gins);CHKERRQ(ierr); 108a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 109a13144ffSStefano Zampini 110a13144ffSStefano Zampini /* constants */ 111a13144ffSStefano Zampini ptr += rsize*csize; 112a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,esize,csize,ptr,&GEd);CHKERRQ(ierr); 1137dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,edge,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 114a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,&GEd);CHKERRQ(ierr); 115a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 116f498cd09SStefano Zampini ierr = MatDenseOrthogonalRangeOrComplement(GEd,PETSC_FALSE,5*esize,work,rwork,GKins);CHKERRQ(ierr); 117a13144ffSStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1181e0482f5SStefano Zampini 1191e0482f5SStefano Zampini if (corners) { 1201e0482f5SStefano Zampini Mat GEc; 1211683a169SBarry Smith const PetscScalar *vals; 1221683a169SBarry Smith PetscScalar v; 1231e0482f5SStefano Zampini 1247dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,edge,corners,MAT_INITIAL_MATRIX,&GEc);CHKERRQ(ierr); 1251e0482f5SStefano Zampini ierr = MatTransposeMatMult(GEc,*GKins,MAT_INITIAL_MATRIX,1.0,&GEd);CHKERRQ(ierr); 1261683a169SBarry Smith ierr = MatDenseGetArrayRead(GEd,&vals);CHKERRQ(ierr); 127637e8532SStefano Zampini /* v = PetscAbsScalar(vals[0]) */; 128637e8532SStefano Zampini v = 1.; 1291e0482f5SStefano Zampini cvals[0] = vals[0]/v; 1301e0482f5SStefano Zampini cvals[1] = vals[1]/v; 1311683a169SBarry Smith ierr = MatDenseRestoreArrayRead(GEd,&vals);CHKERRQ(ierr); 1321e0482f5SStefano Zampini ierr = MatScale(*GKins,1./v);CHKERRQ(ierr); 1331e0482f5SStefano Zampini #if defined(PRINT_GDET) 1341e0482f5SStefano Zampini { 1351e0482f5SStefano Zampini PetscViewer viewer; 1361e0482f5SStefano Zampini char filename[256]; 1371e0482f5SStefano Zampini sprintf(filename,"Gdet_l%d_r%d_cc%d.m",lev,PetscGlobalRank,inc++); 1381e0482f5SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1391e0482f5SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1401e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEc,"GEc");CHKERRQ(ierr); 1411e0482f5SStefano Zampini ierr = MatView(GEc,viewer);CHKERRQ(ierr); 1421e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)(*GKins),"GK");CHKERRQ(ierr); 1431e0482f5SStefano Zampini ierr = MatView(*GKins,viewer);CHKERRQ(ierr); 1441e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEd,"Gproj");CHKERRQ(ierr); 1451e0482f5SStefano Zampini ierr = MatView(GEd,viewer);CHKERRQ(ierr); 1461e0482f5SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1471e0482f5SStefano Zampini } 1481e0482f5SStefano Zampini #endif 1491e0482f5SStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1501e0482f5SStefano Zampini ierr = MatDestroy(&GEc);CHKERRQ(ierr); 1511e0482f5SStefano Zampini } 1521e0482f5SStefano Zampini 153a13144ffSStefano Zampini PetscFunctionReturn(0); 154a13144ffSStefano Zampini } 155a13144ffSStefano Zampini 156a13144ffSStefano Zampini PetscErrorCode PCBDDCNedelecSupport(PC pc) 157a13144ffSStefano Zampini { 158a13144ffSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 159a13144ffSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1600569b399SStefano Zampini Mat G,T,conn,lG,lGt,lGis,lGall,lGe,lGinit; 161eee23b56SStefano Zampini Vec tvec; 162a13144ffSStefano Zampini PetscSF sfv; 1631e0482f5SStefano Zampini ISLocalToGlobalMapping el2g,vl2g,fl2g,al2g; 164a13144ffSStefano Zampini MPI_Comm comm; 165c2151214SStefano Zampini IS lned,primals,allprimals,nedfieldlocal; 166c2151214SStefano Zampini IS *eedges,*extrows,*extcols,*alleedges; 1677d871cd7SStefano Zampini PetscBT btv,bte,btvc,btb,btbd,btvcand,btvi,btee,bter; 168a13144ffSStefano Zampini PetscScalar *vals,*work; 169a13144ffSStefano Zampini PetscReal *rwork; 170a13144ffSStefano Zampini const PetscInt *idxs,*ii,*jj,*iit,*jjt; 1711e0482f5SStefano Zampini PetscInt ne,nv,Lv,order,n,field; 172a13144ffSStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 173eee23b56SStefano Zampini PetscInt i,j,extmem,cum,maxsize,nee; 174b03ebc13SStefano Zampini PetscInt *extrow,*extrowcum,*marks,*vmarks,*gidxs; 175a13144ffSStefano Zampini PetscInt *sfvleaves,*sfvroots; 176b03ebc13SStefano Zampini PetscInt *corners,*cedges; 177637e8532SStefano Zampini PetscInt *ecount,**eneighs,*vcount,**vneighs; 178b03ebc13SStefano Zampini PetscInt *emarks; 179213b8bfaSStefano Zampini PetscBool print,eerr,done,lrc[2],conforming,global,singular,setprimal; 180a13144ffSStefano Zampini PetscErrorCode ierr; 181a13144ffSStefano Zampini 182a13144ffSStefano Zampini PetscFunctionBegin; 183213b8bfaSStefano Zampini /* If the discrete gradient is defined for a subset of dofs and global is true, 184213b8bfaSStefano Zampini it assumes G is given in global ordering for all the dofs. 185213b8bfaSStefano Zampini Otherwise, the ordering is global for the Nedelec field */ 186213b8bfaSStefano Zampini order = pcbddc->nedorder; 187213b8bfaSStefano Zampini conforming = pcbddc->conforming; 188213b8bfaSStefano Zampini field = pcbddc->nedfield; 189213b8bfaSStefano Zampini global = pcbddc->nedglobal; 190213b8bfaSStefano Zampini setprimal = PETSC_FALSE; 191a13144ffSStefano Zampini print = PETSC_FALSE; 192213b8bfaSStefano Zampini singular = PETSC_FALSE; 193a13144ffSStefano Zampini 194213b8bfaSStefano Zampini /* Command line customization */ 195213b8bfaSStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC Nedelec options","PC");CHKERRQ(ierr); 196213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_field_primal","All edge dofs set as primals: Toselli's algorithm C",NULL,setprimal,&setprimal,NULL);CHKERRQ(ierr); 197213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_singular","Infer nullspace from discrete gradient",NULL,singular,&singular,NULL);CHKERRQ(ierr); 198213b8bfaSStefano Zampini ierr = PetscOptionsInt("-pc_bddc_nedelec_order","Test variable order code (to be removed)",NULL,order,&order,NULL);CHKERRQ(ierr); 199213b8bfaSStefano Zampini /* print debug info TODO: to be removed */ 200213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_print","Print debug info",NULL,print,&print,NULL);CHKERRQ(ierr); 201213b8bfaSStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 202213b8bfaSStefano Zampini 203213b8bfaSStefano Zampini /* Return if there are no edges in the decomposition and the problem is not singular */ 2041e0482f5SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&al2g,NULL);CHKERRQ(ierr); 2051e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(al2g,&n);CHKERRQ(ierr); 206213b8bfaSStefano Zampini ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 207213b8bfaSStefano Zampini if (!singular) { 208a13144ffSStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 209a13144ffSStefano Zampini lrc[0] = PETSC_FALSE; 210c2151214SStefano Zampini for (i=0;i<n;i++) { 211a13144ffSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 212a13144ffSStefano Zampini lrc[0] = PETSC_TRUE; 213a13144ffSStefano Zampini break; 214a13144ffSStefano Zampini } 215a13144ffSStefano Zampini } 216a13144ffSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 217820f2d46SBarry Smith ierr = MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);CHKERRMPI(ierr); 218a13144ffSStefano Zampini if (!lrc[1]) PetscFunctionReturn(0); 219213b8bfaSStefano Zampini } 220a13144ffSStefano Zampini 221213b8bfaSStefano Zampini /* Get Nedelec field */ 2226080607fSStefano 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); 223213b8bfaSStefano Zampini if (pcbddc->n_ISForDofsLocal && field >= 0) { 224c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);CHKERRQ(ierr); 225c2151214SStefano Zampini nedfieldlocal = pcbddc->ISForDofsLocal[field]; 226c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 227213b8bfaSStefano Zampini } else if (!pcbddc->n_ISForDofsLocal && field != PETSC_DECIDE) { 228213b8bfaSStefano Zampini ne = n; 229213b8bfaSStefano Zampini nedfieldlocal = NULL; 230213b8bfaSStefano Zampini global = PETSC_TRUE; 231213b8bfaSStefano Zampini } else if (field == PETSC_DECIDE) { 232213b8bfaSStefano Zampini PetscInt rst,ren,*idx; 233213b8bfaSStefano Zampini 234580bdb30SBarry Smith ierr = PetscArrayzero(matis->sf_leafdata,n);CHKERRQ(ierr); 235580bdb30SBarry Smith ierr = PetscArrayzero(matis->sf_rootdata,pc->pmat->rmap->n);CHKERRQ(ierr); 236213b8bfaSStefano Zampini ierr = MatGetOwnershipRange(pcbddc->discretegradient,&rst,&ren);CHKERRQ(ierr); 237213b8bfaSStefano Zampini for (i=rst;i<ren;i++) { 238213b8bfaSStefano Zampini PetscInt nc; 239213b8bfaSStefano Zampini 240213b8bfaSStefano Zampini ierr = MatGetRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 241213b8bfaSStefano Zampini if (nc > 1) matis->sf_rootdata[i-rst] = 1; 242213b8bfaSStefano Zampini ierr = MatRestoreRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 243213b8bfaSStefano Zampini } 244ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata,MPI_REPLACE);CHKERRQ(ierr); 245ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata,MPI_REPLACE);CHKERRQ(ierr); 246213b8bfaSStefano Zampini ierr = PetscMalloc1(n,&idx);CHKERRQ(ierr); 247213b8bfaSStefano Zampini for (i=0,ne=0;i<n;i++) if (matis->sf_leafdata[i]) idx[ne++] = i; 248213b8bfaSStefano Zampini ierr = ISCreateGeneral(comm,ne,idx,PETSC_OWN_POINTER,&nedfieldlocal);CHKERRQ(ierr); 249213b8bfaSStefano Zampini } else { 250213b8bfaSStefano Zampini SETERRQ(comm,PETSC_ERR_USER,"When multiple fields are present, the Nedelec field has to be specified"); 251213b8bfaSStefano Zampini } 252213b8bfaSStefano Zampini 253213b8bfaSStefano Zampini /* Sanity checks */ 254213b8bfaSStefano Zampini if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time"); 255213b8bfaSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis"); 2566080607fSStefano 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); 257213b8bfaSStefano Zampini 258213b8bfaSStefano Zampini /* Just set primal dofs and return */ 2591e0482f5SStefano Zampini if (setprimal) { 260eee23b56SStefano Zampini IS enedfieldlocal; 261eee23b56SStefano Zampini PetscInt *eidxs; 262eee23b56SStefano Zampini 263eee23b56SStefano Zampini ierr = PetscMalloc1(ne,&eidxs);CHKERRQ(ierr); 264eee23b56SStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 265213b8bfaSStefano Zampini if (nedfieldlocal) { 266213b8bfaSStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 267eee23b56SStefano Zampini for (i=0,cum=0;i<ne;i++) { 268eee23b56SStefano Zampini if (PetscRealPart(vals[idxs[i]]) > 2.) { 269eee23b56SStefano Zampini eidxs[cum++] = idxs[i]; 270eee23b56SStefano Zampini } 271eee23b56SStefano Zampini } 272eee23b56SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 273213b8bfaSStefano Zampini } else { 274213b8bfaSStefano Zampini for (i=0,cum=0;i<ne;i++) { 275213b8bfaSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 276213b8bfaSStefano Zampini eidxs[cum++] = i; 277213b8bfaSStefano Zampini } 278213b8bfaSStefano Zampini } 279213b8bfaSStefano Zampini } 280213b8bfaSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 281eee23b56SStefano Zampini ierr = ISCreateGeneral(comm,cum,eidxs,PETSC_COPY_VALUES,&enedfieldlocal);CHKERRQ(ierr); 282eee23b56SStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,enedfieldlocal);CHKERRQ(ierr); 283eee23b56SStefano Zampini ierr = PetscFree(eidxs);CHKERRQ(ierr); 2841e0482f5SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 285eee23b56SStefano Zampini ierr = ISDestroy(&enedfieldlocal);CHKERRQ(ierr); 2861e0482f5SStefano Zampini PetscFunctionReturn(0); 2871e0482f5SStefano Zampini } 288a13144ffSStefano Zampini 289213b8bfaSStefano Zampini /* Compute some l2g maps */ 290213b8bfaSStefano Zampini if (nedfieldlocal) { 291c2151214SStefano Zampini IS is; 292c2151214SStefano Zampini 293c2151214SStefano Zampini /* need to map from the local Nedelec field to local numbering */ 294c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);CHKERRQ(ierr); 2951e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering for the whole dofs*/ 2961e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(al2g,nedfieldlocal,&is);CHKERRQ(ierr); 2971e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is,&al2g);CHKERRQ(ierr); 2981e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering (for Nedelec only) */ 2991e0482f5SStefano Zampini if (global) { 3001e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3011e0482f5SStefano Zampini el2g = al2g; 3021e0482f5SStefano Zampini } else { 3031e0482f5SStefano Zampini IS gis; 3041e0482f5SStefano Zampini 3051e0482f5SStefano Zampini ierr = ISRenumber(is,NULL,NULL,&gis);CHKERRQ(ierr); 3061e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(gis,&el2g);CHKERRQ(ierr); 3071e0482f5SStefano Zampini ierr = ISDestroy(&gis);CHKERRQ(ierr); 3081e0482f5SStefano Zampini } 309c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 310c2151214SStefano Zampini } else { 3111e0482f5SStefano Zampini /* restore default */ 3121e0482f5SStefano Zampini pcbddc->nedfield = -1; 3131e0482f5SStefano Zampini /* one ref for the destruction of al2g, one for el2g */ 3141e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3151e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3161e0482f5SStefano Zampini el2g = al2g; 317c2151214SStefano Zampini fl2g = NULL; 318c2151214SStefano Zampini } 319a13144ffSStefano Zampini 320213b8bfaSStefano Zampini /* Start communication to drop connections for interior edges (for cc analysis only) */ 321580bdb30SBarry Smith ierr = PetscArrayzero(matis->sf_leafdata,n);CHKERRQ(ierr); 322580bdb30SBarry Smith ierr = PetscArrayzero(matis->sf_rootdata,pc->pmat->rmap->n);CHKERRQ(ierr); 323c2151214SStefano Zampini if (nedfieldlocal) { 324c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 325c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1; 326c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 327c2151214SStefano Zampini } else { 328c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1; 329c2151214SStefano Zampini } 330c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 331c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 332213b8bfaSStefano Zampini 333213b8bfaSStefano Zampini if (!singular) { /* drop connections with interior edges to avoid unneeded communications and memory movements */ 334213b8bfaSStefano Zampini ierr = MatDuplicate(pcbddc->discretegradient,MAT_COPY_VALUES,&G);CHKERRQ(ierr); 335213b8bfaSStefano Zampini ierr = MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 3361e0482f5SStefano Zampini if (global) { 3371e0482f5SStefano Zampini PetscInt rst; 3381e0482f5SStefano Zampini 339c2151214SStefano Zampini ierr = MatGetOwnershipRange(G,&rst,NULL);CHKERRQ(ierr); 340c2151214SStefano Zampini for (i=0,cum=0;i<pc->pmat->rmap->n;i++) { 341c2151214SStefano Zampini if (matis->sf_rootdata[i] < 2) { 342c2151214SStefano Zampini matis->sf_rootdata[cum++] = i + rst; 343c2151214SStefano Zampini } 344c2151214SStefano Zampini } 345a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);CHKERRQ(ierr); 346c2151214SStefano Zampini ierr = MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);CHKERRQ(ierr); 3471e0482f5SStefano Zampini } else { 3481e0482f5SStefano Zampini PetscInt *tbz; 3491e0482f5SStefano Zampini 3501e0482f5SStefano Zampini ierr = PetscMalloc1(ne,&tbz);CHKERRQ(ierr); 351ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata,MPI_REPLACE);CHKERRQ(ierr); 352ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata,MPI_REPLACE);CHKERRQ(ierr); 3531e0482f5SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3541e0482f5SStefano Zampini for (i=0,cum=0;i<ne;i++) 3551e0482f5SStefano Zampini if (matis->sf_leafdata[idxs[i]] == 1) 3561e0482f5SStefano Zampini tbz[cum++] = i; 3571e0482f5SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3581e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApply(el2g,cum,tbz,tbz);CHKERRQ(ierr); 3591e0482f5SStefano Zampini ierr = MatZeroRows(G,cum,tbz,0.,NULL,NULL);CHKERRQ(ierr); 3601e0482f5SStefano Zampini ierr = PetscFree(tbz);CHKERRQ(ierr); 3611e0482f5SStefano Zampini } 362213b8bfaSStefano Zampini } else { /* we need the entire G to infer the nullspace */ 363213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->discretegradient);CHKERRQ(ierr); 364213b8bfaSStefano Zampini G = pcbddc->discretegradient; 365213b8bfaSStefano Zampini } 366a13144ffSStefano Zampini 367a13144ffSStefano Zampini /* Extract subdomain relevant rows of G */ 368a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(el2g,&idxs);CHKERRQ(ierr); 369a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);CHKERRQ(ierr); 3707dae84e0SHong Zhang ierr = MatCreateSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);CHKERRQ(ierr); 371a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);CHKERRQ(ierr); 372a13144ffSStefano Zampini ierr = ISDestroy(&lned);CHKERRQ(ierr); 373a13144ffSStefano Zampini ierr = MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);CHKERRQ(ierr); 374a13144ffSStefano Zampini ierr = MatDestroy(&lGall);CHKERRQ(ierr); 375a13144ffSStefano Zampini ierr = MatISGetLocalMat(lGis,&lG);CHKERRQ(ierr); 376a13144ffSStefano Zampini 377213b8bfaSStefano Zampini /* SF for nodal dofs communications */ 378c2151214SStefano Zampini ierr = MatGetLocalSize(G,NULL,&Lv);CHKERRQ(ierr); 379a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);CHKERRQ(ierr); 380a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)vl2g);CHKERRQ(ierr); 381a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(vl2g,&nv);CHKERRQ(ierr); 382a13144ffSStefano Zampini ierr = PetscSFCreate(comm,&sfv);CHKERRQ(ierr); 383a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(vl2g,&idxs);CHKERRQ(ierr); 384a13144ffSStefano Zampini ierr = PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);CHKERRQ(ierr); 385a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);CHKERRQ(ierr); 386213b8bfaSStefano Zampini i = singular ? 2 : 1; 387213b8bfaSStefano Zampini ierr = PetscMalloc2(i*nv,&sfvleaves,i*Lv,&sfvroots);CHKERRQ(ierr); 388a13144ffSStefano Zampini 3891e0482f5SStefano Zampini /* Destroy temporary G created in MATIS format and modified G */ 390213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)lG);CHKERRQ(ierr); 391a13144ffSStefano Zampini ierr = MatDestroy(&lGis);CHKERRQ(ierr); 392213b8bfaSStefano Zampini ierr = MatDestroy(&G);CHKERRQ(ierr); 393a13144ffSStefano Zampini 394213b8bfaSStefano Zampini if (print) { 395213b8bfaSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"initial_lG");CHKERRQ(ierr); 396213b8bfaSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 397213b8bfaSStefano Zampini } 398213b8bfaSStefano Zampini 399213b8bfaSStefano Zampini /* Save lG for values insertion in change of basis */ 4000569b399SStefano Zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);CHKERRQ(ierr); 4010569b399SStefano Zampini 402a13144ffSStefano Zampini /* Analyze the edge-nodes connections (duplicate lG) */ 4034e64d54eSstefano_zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGe);CHKERRQ(ierr); 4044e64d54eSstefano_zampini ierr = MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 405a13144ffSStefano Zampini ierr = PetscBTCreate(nv,&btv);CHKERRQ(ierr); 406a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&bte);CHKERRQ(ierr); 407a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&btb);CHKERRQ(ierr); 4087d871cd7SStefano Zampini ierr = PetscBTCreate(ne,&btbd);CHKERRQ(ierr); 409c2151214SStefano Zampini ierr = PetscBTCreate(nv,&btvcand);CHKERRQ(ierr); 410a13144ffSStefano Zampini /* need to import the boundary specification to ensure the 411a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 412a13144ffSStefano Zampini if (pcbddc->DirichletBoundariesLocal) { 413c2151214SStefano Zampini IS is; 414c2151214SStefano Zampini 415c2151214SStefano Zampini if (fl2g) { 416c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);CHKERRQ(ierr); 417c2151214SStefano Zampini } else { 418c2151214SStefano Zampini is = pcbddc->DirichletBoundariesLocal; 419c2151214SStefano Zampini } 420c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 421c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 422a13144ffSStefano Zampini for (i=0;i<cum;i++) { 423a13144ffSStefano Zampini if (idxs[i] >= 0) { 424a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 4257d871cd7SStefano Zampini ierr = PetscBTSet(btbd,idxs[i]);CHKERRQ(ierr); 426a13144ffSStefano Zampini } 427a13144ffSStefano Zampini } 428c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 429c2151214SStefano Zampini if (fl2g) { 430c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 431c2151214SStefano Zampini } 432a13144ffSStefano Zampini } 433a13144ffSStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 434c2151214SStefano Zampini IS is; 435c2151214SStefano Zampini 436c2151214SStefano Zampini if (fl2g) { 437c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);CHKERRQ(ierr); 438c2151214SStefano Zampini } else { 439c2151214SStefano Zampini is = pcbddc->NeumannBoundariesLocal; 440c2151214SStefano Zampini } 441c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 442c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 443a13144ffSStefano Zampini for (i=0;i<cum;i++) { 444a13144ffSStefano Zampini if (idxs[i] >= 0) { 445a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 446a13144ffSStefano Zampini } 447a13144ffSStefano Zampini } 448c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 449c2151214SStefano Zampini if (fl2g) { 450c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 451a13144ffSStefano Zampini } 452c2151214SStefano Zampini } 453c2151214SStefano Zampini 454213b8bfaSStefano Zampini /* Count neighs per dof */ 455b63b1311SStefano Zampini ierr = ISLocalToGlobalMappingGetNodeInfo(el2g,NULL,&ecount,&eneighs);CHKERRQ(ierr); 456b63b1311SStefano Zampini ierr = ISLocalToGlobalMappingGetNodeInfo(vl2g,NULL,&vcount,&vneighs);CHKERRQ(ierr); 457637e8532SStefano Zampini 4587d871cd7SStefano Zampini /* need to remove coarse faces' dofs and coarse edges' dirichlet dofs 4597d871cd7SStefano Zampini for proper detection of coarse edges' endpoints */ 46062b0c6f7SStefano Zampini ierr = PetscBTCreate(ne,&btee);CHKERRQ(ierr); 46162b0c6f7SStefano Zampini for (i=0;i<ne;i++) { 462b63b1311SStefano Zampini if ((ecount[i] > 2 && !PetscBTLookup(btbd,i)) || (ecount[i] == 2 && PetscBTLookup(btb,i))) { 46362b0c6f7SStefano Zampini ierr = PetscBTSet(btee,i);CHKERRQ(ierr); 46462b0c6f7SStefano Zampini } 46562b0c6f7SStefano Zampini } 466637e8532SStefano Zampini ierr = PetscMalloc1(ne,&marks);CHKERRQ(ierr); 46762b0c6f7SStefano Zampini if (!conforming) { 46862b0c6f7SStefano Zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 46962b0c6f7SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 47062b0c6f7SStefano Zampini } 4714e64d54eSstefano_zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 472dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGe,&vals);CHKERRQ(ierr); 47362b0c6f7SStefano Zampini cum = 0; 474a13144ffSStefano Zampini for (i=0;i<ne;i++) { 475dec27d64SStefano Zampini /* eliminate rows corresponding to edge dofs belonging to coarse faces */ 47662b0c6f7SStefano Zampini if (!PetscBTLookup(btee,i)) { 477a13144ffSStefano Zampini marks[cum++] = i; 478dec27d64SStefano Zampini continue; 479dec27d64SStefano Zampini } 480dec27d64SStefano Zampini /* set badly connected edge dofs as primal */ 48162b0c6f7SStefano Zampini if (!conforming) { 48262b0c6f7SStefano Zampini if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */ 483a13144ffSStefano Zampini marks[cum++] = i; 484a13144ffSStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 485a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 486a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 487a13144ffSStefano Zampini } 48862b0c6f7SStefano Zampini } else { 48962b0c6f7SStefano Zampini /* every edge dofs should be connected trough a certain number of nodal dofs 49062b0c6f7SStefano Zampini to other edge dofs belonging to coarse edges 49162b0c6f7SStefano Zampini - at most 2 endpoints 49262b0c6f7SStefano Zampini - order-1 interior nodal dofs 49362b0c6f7SStefano Zampini - no undefined nodal dofs (nconn < order) 49462b0c6f7SStefano Zampini */ 49562b0c6f7SStefano Zampini PetscInt ends = 0,ints = 0, undef = 0; 49662b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 49762b0c6f7SStefano Zampini PetscInt v = jj[j],k; 49862b0c6f7SStefano Zampini PetscInt nconn = iit[v+1]-iit[v]; 49962b0c6f7SStefano Zampini for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--; 50062b0c6f7SStefano Zampini if (nconn > order) ends++; 50162b0c6f7SStefano Zampini else if (nconn == order) ints++; 50262b0c6f7SStefano Zampini else undef++; 50362b0c6f7SStefano Zampini } 50462b0c6f7SStefano Zampini if (undef || ends > 2 || ints != order -1) { 50562b0c6f7SStefano Zampini marks[cum++] = i; 50662b0c6f7SStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 50762b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 50862b0c6f7SStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 50962b0c6f7SStefano Zampini } 51062b0c6f7SStefano Zampini } 51162b0c6f7SStefano Zampini } 512a13144ffSStefano Zampini } 513dec27d64SStefano Zampini /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */ 514dec27d64SStefano Zampini if (!order && ii[i+1] != ii[i]) { 515dec27d64SStefano Zampini PetscScalar val = 1./(ii[i+1]-ii[i]-1); 516dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vals[j] = val; 517a13144ffSStefano Zampini } 518dec27d64SStefano Zampini } 51962b0c6f7SStefano Zampini ierr = PetscBTDestroy(&btee);CHKERRQ(ierr); 520dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGe,&vals);CHKERRQ(ierr); 5214e64d54eSstefano_zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 52262b0c6f7SStefano Zampini if (!conforming) { 52362b0c6f7SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 52462b0c6f7SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 52562b0c6f7SStefano Zampini } 5264e64d54eSstefano_zampini ierr = MatZeroRows(lGe,cum,marks,0.,NULL,NULL);CHKERRQ(ierr); 527637e8532SStefano Zampini 528b03ebc13SStefano Zampini /* identify splitpoints and corner candidates */ 5294e64d54eSstefano_zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 530a13144ffSStefano Zampini if (print) { 5314e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");CHKERRQ(ierr); 5324e64d54eSstefano_zampini ierr = MatView(lGe,NULL);CHKERRQ(ierr); 5334e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");CHKERRQ(ierr); 534a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 535a13144ffSStefano Zampini } 536a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 537dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGt,&vals);CHKERRQ(ierr); 538a13144ffSStefano Zampini for (i=0;i<nv;i++) { 539637e8532SStefano Zampini PetscInt ord = order, test = ii[i+1]-ii[i], vc = vcount[i]; 5407d871cd7SStefano Zampini PetscBool sneighs = PETSC_TRUE, bdir = PETSC_FALSE; 541b03ebc13SStefano Zampini if (!order) { /* variable order */ 542dec27d64SStefano Zampini PetscReal vorder = 0.; 543dec27d64SStefano Zampini 544dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]); 545dec27d64SStefano Zampini test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON); 5466080607fSStefano Zampini if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%D)",vorder,test); 547dec27d64SStefano Zampini ord = 1; 548dec27d64SStefano Zampini } 549cf9c20a2SJed Brown if (PetscUnlikelyDebug(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); 550637e8532SStefano Zampini for (j=ii[i];j<ii[i+1] && sneighs;j++) { 5517d871cd7SStefano Zampini if (PetscBTLookup(btbd,jj[j])) { 5527d871cd7SStefano Zampini bdir = PETSC_TRUE; 5537d871cd7SStefano Zampini break; 5547d871cd7SStefano Zampini } 555637e8532SStefano Zampini if (vc != ecount[jj[j]]) { 556637e8532SStefano Zampini sneighs = PETSC_FALSE; 557637e8532SStefano Zampini } else { 558637e8532SStefano Zampini PetscInt k,*vn = vneighs[i], *en = eneighs[jj[j]]; 559637e8532SStefano Zampini for (k=0;k<vc;k++) { 560637e8532SStefano Zampini if (vn[k] != en[k]) { 561637e8532SStefano Zampini sneighs = PETSC_FALSE; 562637e8532SStefano Zampini break; 563637e8532SStefano Zampini } 564637e8532SStefano Zampini } 565637e8532SStefano Zampini } 566637e8532SStefano Zampini } 5677d871cd7SStefano Zampini if (!sneighs || test >= 3*ord || bdir) { /* splitpoints */ 5686080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %D (%D %D %D)\n",i,!sneighs,test >= 3*ord,bdir); 569a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 570dec27d64SStefano Zampini } else if (test == ord) { 571b03ebc13SStefano Zampini if (order == 1 || (!order && ii[i+1]-ii[i] == 1)) { 5726080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %D\n",i); 573a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 574a13144ffSStefano Zampini } else { 5756080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %D\n",i); 576a13144ffSStefano Zampini ierr = PetscBTSet(btvcand,i);CHKERRQ(ierr); 577a13144ffSStefano Zampini } 578a13144ffSStefano Zampini } 579a13144ffSStefano Zampini } 580b63b1311SStefano Zampini ierr = ISLocalToGlobalMappingRestoreNodeInfo(el2g,NULL,&ecount,&eneighs);CHKERRQ(ierr); 581b63b1311SStefano Zampini ierr = ISLocalToGlobalMappingRestoreNodeInfo(vl2g,NULL,&vcount,&vneighs);CHKERRQ(ierr); 5827d871cd7SStefano Zampini ierr = PetscBTDestroy(&btbd);CHKERRQ(ierr); 583b03ebc13SStefano Zampini 584b03ebc13SStefano Zampini /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */ 585b03ebc13SStefano Zampini if (order != 1) { 586b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"INSPECTING CANDIDATES\n"); 587b03ebc13SStefano Zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 588b03ebc13SStefano Zampini for (i=0;i<nv;i++) { 589b03ebc13SStefano Zampini if (PetscBTLookup(btvcand,i)) { 590b03ebc13SStefano Zampini PetscBool found = PETSC_FALSE; 591b03ebc13SStefano Zampini for (j=ii[i];j<ii[i+1] && !found;j++) { 592b03ebc13SStefano Zampini PetscInt k,e = jj[j]; 593b03ebc13SStefano Zampini if (PetscBTLookup(bte,e)) continue; 594b03ebc13SStefano Zampini for (k=iit[e];k<iit[e+1];k++) { 595b03ebc13SStefano Zampini PetscInt v = jjt[k]; 596b03ebc13SStefano Zampini if (v != i && PetscBTLookup(btvcand,v)) { 597b03ebc13SStefano Zampini found = PETSC_TRUE; 598b03ebc13SStefano Zampini break; 599b03ebc13SStefano Zampini } 600b03ebc13SStefano Zampini } 601b03ebc13SStefano Zampini } 602b03ebc13SStefano Zampini if (!found) { 6036080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %D CLEARED\n",i); 604b03ebc13SStefano Zampini ierr = PetscBTClear(btvcand,i);CHKERRQ(ierr); 605b03ebc13SStefano Zampini } else { 6066080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %D ACCEPTED\n",i); 607b03ebc13SStefano Zampini } 608b03ebc13SStefano Zampini } 609b03ebc13SStefano Zampini } 610b03ebc13SStefano Zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 611b03ebc13SStefano Zampini } 612dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGt,&vals);CHKERRQ(ierr); 613a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 614b03ebc13SStefano Zampini ierr = MatDestroy(&lGe);CHKERRQ(ierr); 615a13144ffSStefano Zampini 616a13144ffSStefano Zampini /* Get the local G^T explicitly */ 6170569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 618a13144ffSStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 6194e64d54eSstefano_zampini ierr = MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 620a13144ffSStefano Zampini 6214e64d54eSstefano_zampini /* Mark interior nodal dofs */ 622a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 6234e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvi);CHKERRQ(ierr); 624a13144ffSStefano Zampini for (i=1;i<n_neigh;i++) { 625a13144ffSStefano Zampini for (j=0;j<n_shared[i];j++) { 6264e64d54eSstefano_zampini ierr = PetscBTSet(btvi,shared[i][j]);CHKERRQ(ierr); 627a13144ffSStefano Zampini } 628a13144ffSStefano Zampini } 629a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 630a13144ffSStefano Zampini 631a13144ffSStefano Zampini /* communicate corners and splitpoints */ 632a13144ffSStefano Zampini ierr = PetscMalloc1(nv,&vmarks);CHKERRQ(ierr); 633580bdb30SBarry Smith ierr = PetscArrayzero(sfvleaves,nv);CHKERRQ(ierr); 634580bdb30SBarry Smith ierr = PetscArrayzero(sfvroots,Lv);CHKERRQ(ierr); 635a13144ffSStefano Zampini for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1; 636a13144ffSStefano Zampini 637a13144ffSStefano Zampini if (print) { 638a13144ffSStefano Zampini IS tbz; 639a13144ffSStefano Zampini 640a13144ffSStefano Zampini cum = 0; 641a13144ffSStefano Zampini for (i=0;i<nv;i++) 642a13144ffSStefano Zampini if (sfvleaves[i]) 643a13144ffSStefano Zampini vmarks[cum++] = i; 644a13144ffSStefano Zampini 645a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 646a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");CHKERRQ(ierr); 647a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 648a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 649a13144ffSStefano Zampini } 650a13144ffSStefano Zampini 651a13144ffSStefano Zampini ierr = PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 652a13144ffSStefano Zampini ierr = PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 653ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves,MPI_REPLACE);CHKERRQ(ierr); 654ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves,MPI_REPLACE);CHKERRQ(ierr); 655a13144ffSStefano Zampini 6564e64d54eSstefano_zampini /* Zero rows of lGt corresponding to identified corners 6574e64d54eSstefano_zampini and interior nodal dofs */ 658a13144ffSStefano Zampini cum = 0; 659a13144ffSStefano Zampini for (i=0;i<nv;i++) { 660a13144ffSStefano Zampini if (sfvleaves[i]) { 661a13144ffSStefano Zampini vmarks[cum++] = i; 662a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 663a13144ffSStefano Zampini } 6644e64d54eSstefano_zampini if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i; 665a13144ffSStefano Zampini } 6664e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvi);CHKERRQ(ierr); 667a13144ffSStefano Zampini if (print) { 668a13144ffSStefano Zampini IS tbz; 669a13144ffSStefano Zampini 670a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 6714e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");CHKERRQ(ierr); 672a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 673a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 674a13144ffSStefano Zampini } 675a13144ffSStefano Zampini ierr = MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);CHKERRQ(ierr); 676a13144ffSStefano Zampini ierr = PetscFree(vmarks);CHKERRQ(ierr); 677a13144ffSStefano Zampini ierr = PetscSFDestroy(&sfv);CHKERRQ(ierr); 678a13144ffSStefano Zampini ierr = PetscFree2(sfvleaves,sfvroots);CHKERRQ(ierr); 679a13144ffSStefano Zampini 680a13144ffSStefano Zampini /* Recompute G */ 681a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 682a13144ffSStefano Zampini ierr = MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);CHKERRQ(ierr); 683a13144ffSStefano Zampini if (print) { 684a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"used_lG");CHKERRQ(ierr); 685a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 686a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lGt,"used_lGt");CHKERRQ(ierr); 687a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 688a13144ffSStefano Zampini } 689a13144ffSStefano Zampini 690a13144ffSStefano Zampini /* Get primal dofs (if any) */ 691a13144ffSStefano Zampini cum = 0; 692a13144ffSStefano Zampini for (i=0;i<ne;i++) { 693a13144ffSStefano Zampini if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i; 694a13144ffSStefano Zampini } 695c2151214SStefano Zampini if (fl2g) { 696c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);CHKERRQ(ierr); 697c2151214SStefano Zampini } 698a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 699a13144ffSStefano Zampini if (print) { 700a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");CHKERRQ(ierr); 701a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 702a13144ffSStefano Zampini } 703a13144ffSStefano Zampini ierr = PetscBTDestroy(&bte);CHKERRQ(ierr); 704c2151214SStefano Zampini /* TODO: what if the user passed in some of them ? */ 705a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 706a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 707a13144ffSStefano Zampini 708a13144ffSStefano Zampini /* Compute edge connectivity */ 709a13144ffSStefano Zampini ierr = PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");CHKERRQ(ierr); 7104222ddf1SHong Zhang 7114222ddf1SHong Zhang /* Symbolic conn = lG*lGt */ 7124222ddf1SHong Zhang ierr = MatProductCreate(lG,lGt,NULL,&conn);CHKERRQ(ierr); 7134222ddf1SHong Zhang ierr = MatProductSetType(conn,MATPRODUCT_AB);CHKERRQ(ierr); 7144222ddf1SHong Zhang ierr = MatProductSetAlgorithm(conn,"default");CHKERRQ(ierr); 7154222ddf1SHong Zhang ierr = MatProductSetFill(conn,PETSC_DEFAULT);CHKERRQ(ierr); 7164222ddf1SHong Zhang ierr = PetscObjectSetOptionsPrefix((PetscObject)conn,"econn_");CHKERRQ(ierr); 7174222ddf1SHong Zhang ierr = MatProductSetFromOptions(conn);CHKERRQ(ierr); 7184222ddf1SHong Zhang ierr = MatProductSymbolic(conn);CHKERRQ(ierr); 7194222ddf1SHong Zhang 720a13144ffSStefano Zampini ierr = MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 721c2151214SStefano Zampini if (fl2g) { 722c2151214SStefano Zampini PetscBT btf; 723c2151214SStefano Zampini PetscInt *iia,*jja,*iiu,*jju; 724c2151214SStefano Zampini PetscBool rest = PETSC_FALSE,free = PETSC_FALSE; 725c2151214SStefano Zampini 726c2151214SStefano Zampini /* create CSR for all local dofs */ 727c2151214SStefano Zampini ierr = PetscMalloc1(n+1,&iia);CHKERRQ(ierr); 728c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */ 7296080607fSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid size of CSR graph %D. Should be %D",pcbddc->mat_graph->nvtxs_csr,n); 730c2151214SStefano Zampini iiu = pcbddc->mat_graph->xadj; 731c2151214SStefano Zampini jju = pcbddc->mat_graph->adjncy; 732c2151214SStefano Zampini } else if (pcbddc->use_local_adj) { 733c2151214SStefano Zampini rest = PETSC_TRUE; 734c2151214SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 735c2151214SStefano Zampini } else { 736c2151214SStefano Zampini free = PETSC_TRUE; 737c2151214SStefano Zampini ierr = PetscMalloc2(n+1,&iiu,n,&jju);CHKERRQ(ierr); 738c2151214SStefano Zampini iiu[0] = 0; 739c2151214SStefano Zampini for (i=0;i<n;i++) { 740c2151214SStefano Zampini iiu[i+1] = i+1; 741c2151214SStefano Zampini jju[i] = -1; 742d904f53bSStefano Zampini } 743c2151214SStefano Zampini } 744c2151214SStefano Zampini 745c2151214SStefano Zampini /* import sizes of CSR */ 746c2151214SStefano Zampini iia[0] = 0; 747c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i]; 748c2151214SStefano Zampini 749c2151214SStefano Zampini /* overwrite entries corresponding to the Nedelec field */ 750c2151214SStefano Zampini ierr = PetscBTCreate(n,&btf);CHKERRQ(ierr); 751c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 752c2151214SStefano Zampini for (i=0;i<ne;i++) { 753c2151214SStefano Zampini ierr = PetscBTSet(btf,idxs[i]);CHKERRQ(ierr); 754c2151214SStefano Zampini iia[idxs[i]+1] = ii[i+1]-ii[i]; 755c2151214SStefano Zampini } 756c2151214SStefano Zampini 757c2151214SStefano Zampini /* iia in CSR */ 758c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] += iia[i]; 759c2151214SStefano Zampini 760c2151214SStefano Zampini /* jja in CSR */ 761c2151214SStefano Zampini ierr = PetscMalloc1(iia[n],&jja);CHKERRQ(ierr); 762c2151214SStefano Zampini for (i=0;i<n;i++) 763c2151214SStefano Zampini if (!PetscBTLookup(btf,i)) 764c2151214SStefano Zampini for (j=0;j<iiu[i+1]-iiu[i];j++) 765c2151214SStefano Zampini jja[iia[i]+j] = jju[iiu[i]+j]; 766c2151214SStefano Zampini 767c2151214SStefano Zampini /* map edge dofs connectivity */ 7681e0482f5SStefano Zampini if (jj) { 769c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);CHKERRQ(ierr); 770c2151214SStefano Zampini for (i=0;i<ne;i++) { 771c2151214SStefano Zampini PetscInt e = idxs[i]; 772c2151214SStefano Zampini for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j]; 773c2151214SStefano Zampini } 7741e0482f5SStefano Zampini } 775c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 776c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);CHKERRQ(ierr); 777c2151214SStefano Zampini if (rest) { 778c2151214SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 779c2151214SStefano Zampini } 780c2151214SStefano Zampini if (free) { 781c2151214SStefano Zampini ierr = PetscFree2(iiu,jju);CHKERRQ(ierr); 782c2151214SStefano Zampini } 783c2151214SStefano Zampini ierr = PetscBTDestroy(&btf);CHKERRQ(ierr); 784c2151214SStefano Zampini } else { 785c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);CHKERRQ(ierr); 786c2151214SStefano Zampini } 787c2151214SStefano Zampini 788a13144ffSStefano Zampini /* Analyze interface for edge dofs */ 789a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 790213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 791a13144ffSStefano Zampini 792a13144ffSStefano Zampini /* Get coarse edges in the edge space */ 793c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 794a13144ffSStefano Zampini ierr = MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 795a13144ffSStefano Zampini 796c2151214SStefano Zampini if (fl2g) { 797c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 798c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 799c2151214SStefano Zampini for (i=0;i<nee;i++) { 800c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 801c2151214SStefano Zampini } 802c2151214SStefano Zampini } else { 803c2151214SStefano Zampini eedges = alleedges; 804c2151214SStefano Zampini primals = allprimals; 805c2151214SStefano Zampini } 806c2151214SStefano Zampini 807a13144ffSStefano Zampini /* Mark fine edge dofs with their coarse edge id */ 808580bdb30SBarry Smith ierr = PetscArrayzero(marks,ne);CHKERRQ(ierr); 809c2151214SStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 810c2151214SStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 811c2151214SStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = nee+1; 812c2151214SStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 813c2151214SStefano Zampini if (print) { 814c2151214SStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");CHKERRQ(ierr); 815c2151214SStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 816c2151214SStefano Zampini } 817c2151214SStefano Zampini 818c2151214SStefano Zampini maxsize = 0; 819a13144ffSStefano Zampini for (i=0;i<nee;i++) { 820a13144ffSStefano Zampini PetscInt size,mark = i+1; 821a13144ffSStefano Zampini 822a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 823a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 824a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 825a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 826a13144ffSStefano Zampini maxsize = PetscMax(maxsize,size); 827a13144ffSStefano Zampini } 828a13144ffSStefano Zampini 829a13144ffSStefano Zampini /* Find coarse edge endpoints */ 830a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 831a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 832a13144ffSStefano Zampini for (i=0;i<nee;i++) { 833a13144ffSStefano Zampini PetscInt mark = i+1,size; 834a13144ffSStefano Zampini 835a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 8361e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 8376080607fSStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i); 838a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 839a13144ffSStefano Zampini if (print) { 8406080607fSStefano Zampini ierr = PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %D\n",i);CHKERRQ(ierr); 8416080607fSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 842a13144ffSStefano Zampini } 843a13144ffSStefano Zampini for (j=0;j<size;j++) { 844a13144ffSStefano Zampini PetscInt k, ee = idxs[j]; 8456080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," idx %D\n",ee); 846a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 8476080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," inspect %D\n",jj[k]); 848a13144ffSStefano Zampini if (PetscBTLookup(btv,jj[k])) { 8496080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %D\n",jj[k]); 850a13144ffSStefano Zampini } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */ 851a13144ffSStefano Zampini PetscInt k2; 852a13144ffSStefano Zampini PetscBool corner = PETSC_FALSE; 853a13144ffSStefano Zampini for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) { 8546080607fSStefano 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])); 855c2151214SStefano Zampini /* it's a corner if either is connected with an edge dof belonging to a different cc or 856c2151214SStefano Zampini if the edge dof lie on the natural part of the boundary */ 857c2151214SStefano Zampini if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) { 858a13144ffSStefano Zampini corner = PETSC_TRUE; 859a13144ffSStefano Zampini break; 860a13144ffSStefano Zampini } 861a13144ffSStefano Zampini } 862a13144ffSStefano Zampini if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */ 8636080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found %D\n",jj[k]); 864a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 865a13144ffSStefano Zampini } else { 866a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n"); 867a13144ffSStefano Zampini } 868a13144ffSStefano Zampini } 869a13144ffSStefano Zampini } 870a13144ffSStefano Zampini } 871a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 872a13144ffSStefano Zampini } 873a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 874a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 875c2151214SStefano Zampini ierr = PetscBTDestroy(&btb);CHKERRQ(ierr); 876a13144ffSStefano Zampini 877a13144ffSStefano Zampini /* Reset marked primal dofs */ 878a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 879a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 880a13144ffSStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = 0; 881a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 882a13144ffSStefano Zampini 8830569b399SStefano Zampini /* Now use the initial lG */ 8840569b399SStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 8850569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 8860569b399SStefano Zampini lG = lGinit; 8870569b399SStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 8880569b399SStefano Zampini 889a13144ffSStefano Zampini /* Compute extended cols indices */ 890b03ebc13SStefano Zampini ierr = PetscBTCreate(nv,&btvc);CHKERRQ(ierr); 891b03ebc13SStefano Zampini ierr = PetscBTCreate(nee,&bter);CHKERRQ(ierr); 892a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 893a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lG,&i);CHKERRQ(ierr); 894a13144ffSStefano Zampini i *= maxsize; 895b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 896a13144ffSStefano Zampini ierr = PetscMalloc2(i,&extrow,i,&gidxs);CHKERRQ(ierr); 897a13144ffSStefano Zampini eerr = PETSC_FALSE; 898a13144ffSStefano Zampini for (i=0;i<nee;i++) { 899b03ebc13SStefano Zampini PetscInt size,found = 0; 900a13144ffSStefano Zampini 901a13144ffSStefano Zampini cum = 0; 902a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 9031e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 9046080607fSStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i); 905a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 906b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 907a13144ffSStefano Zampini for (j=0;j<size;j++) { 908a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 909b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 910b03ebc13SStefano Zampini PetscInt vv = jj[k]; 911b03ebc13SStefano Zampini if (!PetscBTLookup(btv,vv)) extrow[cum++] = vv; 912b03ebc13SStefano Zampini else if (!PetscBTLookupSet(btvc,vv)) found++; 913b03ebc13SStefano Zampini } 914a13144ffSStefano Zampini } 915a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 916a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 917a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 918a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 919a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 920a13144ffSStefano Zampini /* it may happen that endpoints are not defined at this point 921a13144ffSStefano Zampini if it is the case, mark this edge for a second pass */ 922b03ebc13SStefano Zampini if (cum != size -1 || found != 2) { 923b03ebc13SStefano Zampini ierr = PetscBTSet(bter,i);CHKERRQ(ierr); 924a13144ffSStefano Zampini if (print) { 925a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge");CHKERRQ(ierr); 926a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 927a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol");CHKERRQ(ierr); 928a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 929a13144ffSStefano Zampini } 930a13144ffSStefano Zampini eerr = PETSC_TRUE; 931a13144ffSStefano Zampini } 932a13144ffSStefano Zampini } 9334e64d54eSstefano_zampini /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */ 934820f2d46SBarry Smith ierr = MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);CHKERRMPI(ierr); 935a13144ffSStefano Zampini if (done) { 936a13144ffSStefano Zampini PetscInt *newprimals; 937a13144ffSStefano Zampini 938a13144ffSStefano Zampini ierr = PetscMalloc1(ne,&newprimals);CHKERRQ(ierr); 939a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 940a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 941580bdb30SBarry Smith ierr = PetscArraycpy(newprimals,idxs,cum);CHKERRQ(ierr); 942a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 9430569b399SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 9446080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"DOING SECOND PASS (eerr %D)\n",eerr); 945a13144ffSStefano Zampini for (i=0;i<nee;i++) { 946b03ebc13SStefano Zampini PetscBool has_candidates = PETSC_FALSE; 947b03ebc13SStefano Zampini if (PetscBTLookup(bter,i)) { 948a13144ffSStefano Zampini PetscInt size,mark = i+1; 949a13144ffSStefano Zampini 950a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 951a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 952c2151214SStefano Zampini /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */ 953a13144ffSStefano Zampini for (j=0;j<size;j++) { 954a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 9556080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"Inspecting edge dof %D [%D %D)\n",ee,ii[ee],ii[ee+1]); 956a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 957a13144ffSStefano Zampini /* set all candidates located on the edge as corners */ 958a13144ffSStefano Zampini if (PetscBTLookup(btvcand,jj[k])) { 959a13144ffSStefano Zampini PetscInt k2,vv = jj[k]; 960b03ebc13SStefano Zampini has_candidates = PETSC_TRUE; 9616080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Candidate set to vertex %D\n",vv); 962a13144ffSStefano Zampini ierr = PetscBTSet(btv,vv);CHKERRQ(ierr); 963a13144ffSStefano Zampini /* set all edge dofs connected to candidate as primals */ 964a13144ffSStefano Zampini for (k2=iit[vv];k2<iit[vv+1];k2++) { 965a13144ffSStefano Zampini if (marks[jjt[k2]] == mark) { 966a13144ffSStefano Zampini PetscInt k3,ee2 = jjt[k2]; 9676080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected edge dof set to primal %D\n",ee2); 968a13144ffSStefano Zampini newprimals[cum++] = ee2; 969a13144ffSStefano Zampini /* finally set the new corners */ 970a13144ffSStefano Zampini for (k3=ii[ee2];k3<ii[ee2+1];k3++) { 9716080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected nodal dof set to vertex %D\n",jj[k3]); 972a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k3]);CHKERRQ(ierr); 973a13144ffSStefano Zampini } 974a13144ffSStefano Zampini } 975a13144ffSStefano Zampini } 976b03ebc13SStefano Zampini } else { 9776080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Not a candidate vertex %D\n",jj[k]); 978a13144ffSStefano Zampini } 979a13144ffSStefano Zampini } 980a13144ffSStefano Zampini } 981b03ebc13SStefano Zampini if (!has_candidates) { /* circular edge */ 982b03ebc13SStefano Zampini PetscInt k, ee = idxs[0],*tmarks; 983b03ebc13SStefano Zampini 984b03ebc13SStefano Zampini ierr = PetscCalloc1(ne,&tmarks);CHKERRQ(ierr); 9856080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Circular edge %D\n",i); 986b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 987b03ebc13SStefano Zampini PetscInt k2; 9886080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Set to corner %D\n",jj[k]); 989b03ebc13SStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 990b03ebc13SStefano Zampini for (k2=iit[jj[k]];k2<iit[jj[k]+1];k2++) tmarks[jjt[k2]]++; 991b03ebc13SStefano Zampini } 992b03ebc13SStefano Zampini for (j=0;j<size;j++) { 993b03ebc13SStefano Zampini if (tmarks[idxs[j]] > 1) { 9946080607fSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Edge dof set to primal %D\n",idxs[j]); 995b03ebc13SStefano Zampini newprimals[cum++] = idxs[j]; 996b03ebc13SStefano Zampini } 997b03ebc13SStefano Zampini } 998b03ebc13SStefano Zampini ierr = PetscFree(tmarks);CHKERRQ(ierr); 999b03ebc13SStefano Zampini } 1000a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1001a13144ffSStefano Zampini } 1002a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 1003a13144ffSStefano Zampini } 1004b03ebc13SStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 10050569b399SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 1006a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,newprimals);CHKERRQ(ierr); 1007c2151214SStefano Zampini if (fl2g) { 1008c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);CHKERRQ(ierr); 1009c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1010c2151214SStefano Zampini for (i=0;i<nee;i++) { 1011c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1012c2151214SStefano Zampini } 1013c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1014c2151214SStefano Zampini } 1015c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1016a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 1017a13144ffSStefano Zampini ierr = PetscFree(newprimals);CHKERRQ(ierr); 1018a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 1019a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1020a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 1021213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 1022c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1023c2151214SStefano Zampini if (fl2g) { 1024c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 1025c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 1026c2151214SStefano Zampini for (i=0;i<nee;i++) { 1027c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 1028c2151214SStefano Zampini } 1029c2151214SStefano Zampini } else { 1030c2151214SStefano Zampini eedges = alleedges; 1031c2151214SStefano Zampini primals = allprimals; 1032c2151214SStefano Zampini } 1033b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 1034a13144ffSStefano Zampini 1035a13144ffSStefano Zampini /* Mark again */ 1036580bdb30SBarry Smith ierr = PetscArrayzero(marks,ne);CHKERRQ(ierr); 1037a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1038a13144ffSStefano Zampini PetscInt size,mark = i+1; 1039a13144ffSStefano Zampini 1040a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 1041a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1042a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 1043a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1044a13144ffSStefano Zampini } 1045a13144ffSStefano Zampini if (print) { 1046a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");CHKERRQ(ierr); 1047a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 1048a13144ffSStefano Zampini } 1049a13144ffSStefano Zampini 1050a13144ffSStefano Zampini /* Recompute extended cols */ 1051a13144ffSStefano Zampini eerr = PETSC_FALSE; 1052a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1053a13144ffSStefano Zampini PetscInt size; 1054a13144ffSStefano Zampini 1055a13144ffSStefano Zampini cum = 0; 1056a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 10571e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 10586080607fSStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i); 1059a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1060a13144ffSStefano Zampini for (j=0;j<size;j++) { 1061a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 10621e0482f5SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) if (!PetscBTLookup(btv,jj[k])) extrow[cum++] = jj[k]; 1063a13144ffSStefano Zampini } 1064a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1065a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 1066a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 1067a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 1068a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 1069a13144ffSStefano Zampini if (cum != size -1) { 1070a13144ffSStefano Zampini if (print) { 1071a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");CHKERRQ(ierr); 1072a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 1073a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");CHKERRQ(ierr); 1074a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 1075a13144ffSStefano Zampini } 1076a13144ffSStefano Zampini eerr = PETSC_TRUE; 1077a13144ffSStefano Zampini } 1078a13144ffSStefano Zampini } 1079a13144ffSStefano Zampini } 1080a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1081a13144ffSStefano Zampini ierr = PetscFree2(extrow,gidxs);CHKERRQ(ierr); 1082b03ebc13SStefano Zampini ierr = PetscBTDestroy(&bter);CHKERRQ(ierr); 10837d871cd7SStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 1084a13144ffSStefano Zampini /* an error should not occur at this point */ 1085a13144ffSStefano Zampini if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS"); 1086a13144ffSStefano Zampini 10874e64d54eSstefano_zampini /* Check the number of endpoints */ 10880569b399SStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1089b03ebc13SStefano Zampini ierr = PetscMalloc1(2*nee,&corners);CHKERRQ(ierr); 1090b03ebc13SStefano Zampini ierr = PetscMalloc1(nee,&cedges);CHKERRQ(ierr); 10914e64d54eSstefano_zampini for (i=0;i<nee;i++) { 1092b03ebc13SStefano Zampini PetscInt size, found = 0, gc[2]; 10934e64d54eSstefano_zampini 1094b03ebc13SStefano Zampini /* init with defaults */ 1095b03ebc13SStefano Zampini cedges[i] = corners[i*2] = corners[i*2+1] = -1; 10964e64d54eSstefano_zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 10971e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 10986080607fSStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i); 10994e64d54eSstefano_zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1100b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 11014e64d54eSstefano_zampini for (j=0;j<size;j++) { 11024e64d54eSstefano_zampini PetscInt k,ee = idxs[j]; 11034e64d54eSstefano_zampini for (k=ii[ee];k<ii[ee+1];k++) { 11044e64d54eSstefano_zampini PetscInt vv = jj[k]; 11054e64d54eSstefano_zampini if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) { 11066080607fSStefano Zampini if (found == 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found more then two corners for edge %D",i); 1107b03ebc13SStefano Zampini corners[i*2+found++] = vv; 11084e64d54eSstefano_zampini } 11094e64d54eSstefano_zampini } 11104e64d54eSstefano_zampini } 1111b03ebc13SStefano Zampini if (found != 2) { 1112b03ebc13SStefano Zampini PetscInt e; 1113b03ebc13SStefano Zampini if (fl2g) { 1114b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,1,idxs,&e);CHKERRQ(ierr); 1115b03ebc13SStefano Zampini } else { 1116b03ebc13SStefano Zampini e = idxs[0]; 1117b03ebc13SStefano Zampini } 11186080607fSStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %D corners for edge %D (astart %D, estart %D)",found,i,e,idxs[0]); 1119b03ebc13SStefano Zampini } 1120eee23b56SStefano Zampini 1121eee23b56SStefano Zampini /* get primal dof index on this coarse edge */ 1122b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,2,corners+2*i,gc);CHKERRQ(ierr); 1123b03ebc13SStefano Zampini if (gc[0] > gc[1]) { 1124b03ebc13SStefano Zampini PetscInt swap = corners[2*i]; 1125b03ebc13SStefano Zampini corners[2*i] = corners[2*i+1]; 1126b03ebc13SStefano Zampini corners[2*i+1] = swap; 1127b03ebc13SStefano Zampini } 1128eee23b56SStefano Zampini cedges[i] = idxs[size-1]; 11294e64d54eSstefano_zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 11306080607fSStefano 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]); 11314e64d54eSstefano_zampini } 11320569b399SStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 11334e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvc);CHKERRQ(ierr); 11344e64d54eSstefano_zampini 113576bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { 1136a13144ffSStefano Zampini /* Inspects columns of lG (rows of lGt) and make sure the change of basis will 1137a13144ffSStefano Zampini not interfere with neighbouring coarse edges */ 1138a13144ffSStefano Zampini ierr = PetscMalloc1(nee+1,&emarks);CHKERRQ(ierr); 1139a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1140a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1141a13144ffSStefano Zampini PetscInt emax = 0,eemax = 0; 1142a13144ffSStefano Zampini 1143a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1144580bdb30SBarry Smith ierr = PetscArrayzero(emarks,nee+1);CHKERRQ(ierr); 1145a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++; 1146a13144ffSStefano Zampini for (j=1;j<nee+1;j++) { 1147a13144ffSStefano Zampini if (emax < emarks[j]) { 1148a13144ffSStefano Zampini emax = emarks[j]; 1149a13144ffSStefano Zampini eemax = j; 1150a13144ffSStefano Zampini } 1151a13144ffSStefano Zampini } 1152a13144ffSStefano Zampini /* not relevant for edges */ 1153a13144ffSStefano Zampini if (!eemax) continue; 1154a13144ffSStefano Zampini 1155a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 1156a13144ffSStefano Zampini if (marks[jj[j]] && marks[jj[j]] != eemax) { 11576080607fSStefano 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",marks[jj[j]]-1,eemax,i,jj[j]); 1158a13144ffSStefano Zampini } 1159a13144ffSStefano Zampini } 1160a13144ffSStefano Zampini } 1161a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 1162a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 116376bd3646SJed Brown } 1164a13144ffSStefano Zampini 1165a13144ffSStefano Zampini /* Compute extended rows indices for edge blocks of the change of basis */ 1166a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1167a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);CHKERRQ(ierr); 1168a13144ffSStefano Zampini extmem *= maxsize; 1169a13144ffSStefano Zampini ierr = PetscMalloc1(extmem*nee,&extrow);CHKERRQ(ierr); 1170a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extrows);CHKERRQ(ierr); 1171a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&extrowcum);CHKERRQ(ierr); 1172a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1173a13144ffSStefano Zampini PetscInt mark = 0,size,start; 1174213b8bfaSStefano Zampini 1175a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1176a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) 1177a13144ffSStefano Zampini if (marks[jj[j]] && !mark) 1178a13144ffSStefano Zampini mark = marks[jj[j]]; 1179a13144ffSStefano Zampini 1180a13144ffSStefano Zampini /* not relevant */ 1181a13144ffSStefano Zampini if (!mark) continue; 1182a13144ffSStefano Zampini 1183a13144ffSStefano Zampini /* import extended row */ 1184a13144ffSStefano Zampini mark--; 1185a13144ffSStefano Zampini start = mark*extmem+extrowcum[mark]; 1186a13144ffSStefano Zampini size = ii[i+1]-ii[i]; 11876080607fSStefano Zampini if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %D > %D",extrowcum[mark] + size,extmem); 1188580bdb30SBarry Smith ierr = PetscArraycpy(extrow+start,jj+ii[i],size);CHKERRQ(ierr); 1189a13144ffSStefano Zampini extrowcum[mark] += size; 1190a13144ffSStefano Zampini } 1191a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1192213b8bfaSStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 1193213b8bfaSStefano Zampini ierr = PetscFree(marks);CHKERRQ(ierr); 1194213b8bfaSStefano Zampini 1195213b8bfaSStefano Zampini /* Compress extrows */ 1196a13144ffSStefano Zampini cum = 0; 1197a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1198a13144ffSStefano Zampini PetscInt size = extrowcum[i],*start = extrow + i*extmem; 1199a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&size,start);CHKERRQ(ierr); 1200a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);CHKERRQ(ierr); 1201a13144ffSStefano Zampini cum = PetscMax(cum,size); 1202a13144ffSStefano Zampini } 1203a13144ffSStefano Zampini ierr = PetscFree(extrowcum);CHKERRQ(ierr); 1204a13144ffSStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 1205a13144ffSStefano Zampini ierr = PetscBTDestroy(&btvcand);CHKERRQ(ierr); 1206a13144ffSStefano Zampini 1207a13144ffSStefano Zampini /* Workspace for lapack inner calls and VecSetValues */ 1208a13144ffSStefano Zampini ierr = PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);CHKERRQ(ierr); 1209a13144ffSStefano Zampini 1210a13144ffSStefano Zampini /* Create change of basis matrix (preallocation can be improved) */ 1211a13144ffSStefano Zampini ierr = MatCreate(comm,&T);CHKERRQ(ierr); 1212c2151214SStefano Zampini ierr = MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n, 1213c2151214SStefano Zampini pc->pmat->rmap->N,pc->pmat->rmap->N);CHKERRQ(ierr); 1214a13144ffSStefano Zampini ierr = MatSetType(T,MATAIJ);CHKERRQ(ierr); 1215a13144ffSStefano Zampini ierr = MatSeqAIJSetPreallocation(T,10,NULL);CHKERRQ(ierr); 1216a13144ffSStefano Zampini ierr = MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);CHKERRQ(ierr); 12171e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(T,al2g,al2g);CHKERRQ(ierr); 1218a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1219a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 1220213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&al2g);CHKERRQ(ierr); 1221a13144ffSStefano Zampini 1222a13144ffSStefano Zampini /* Defaults to identity */ 1223c2151214SStefano Zampini ierr = MatCreateVecs(pc->pmat,&tvec,NULL);CHKERRQ(ierr); 1224a13144ffSStefano Zampini ierr = VecSet(tvec,1.0);CHKERRQ(ierr); 1225a13144ffSStefano Zampini ierr = MatDiagonalSet(T,tvec,INSERT_VALUES);CHKERRQ(ierr); 1226a13144ffSStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 1227a13144ffSStefano Zampini 12281e0482f5SStefano Zampini /* Create discrete gradient for the coarser level if needed */ 12291e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 12301e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 12311e0482f5SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 12321e0482f5SStefano Zampini ISLocalToGlobalMapping cel2g,cvl2g; 12331e0482f5SStefano Zampini IS wis,gwis; 12341e0482f5SStefano Zampini PetscInt cnv,cne; 12351e0482f5SStefano Zampini 12361e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,nee,cedges,PETSC_COPY_VALUES,&wis);CHKERRQ(ierr); 12371e0482f5SStefano Zampini if (fl2g) { 12381e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(fl2g,wis,&pcbddc->nedclocal);CHKERRQ(ierr); 12391e0482f5SStefano Zampini } else { 12401e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)wis);CHKERRQ(ierr); 12411e0482f5SStefano Zampini pcbddc->nedclocal = wis; 12421e0482f5SStefano Zampini } 12431e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(el2g,wis,&gwis);CHKERRQ(ierr); 12441e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12451e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cne,&wis);CHKERRQ(ierr); 12461e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cel2g);CHKERRQ(ierr); 12471e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12481e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 12491e0482f5SStefano Zampini 12501e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,2*nee,corners,PETSC_USE_POINTER,&wis);CHKERRQ(ierr); 12511e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(vl2g,wis,&gwis);CHKERRQ(ierr); 12521e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12531e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cnv,&wis);CHKERRQ(ierr); 12541e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cvl2g);CHKERRQ(ierr); 12551e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12561e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 12571e0482f5SStefano Zampini 12581e0482f5SStefano Zampini ierr = MatCreate(comm,&pcbddc->nedcG);CHKERRQ(ierr); 12591e0482f5SStefano Zampini ierr = MatSetSizes(pcbddc->nedcG,PETSC_DECIDE,PETSC_DECIDE,cne,cnv);CHKERRQ(ierr); 12601e0482f5SStefano Zampini ierr = MatSetType(pcbddc->nedcG,MATAIJ);CHKERRQ(ierr); 12611e0482f5SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->nedcG,2,NULL);CHKERRQ(ierr); 12621e0482f5SStefano Zampini ierr = MatMPIAIJSetPreallocation(pcbddc->nedcG,2,NULL,2,NULL);CHKERRQ(ierr); 12631e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->nedcG,cel2g,cvl2g);CHKERRQ(ierr); 12641e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cel2g);CHKERRQ(ierr); 12651e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cvl2g);CHKERRQ(ierr); 12661e0482f5SStefano Zampini } 1267213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&vl2g);CHKERRQ(ierr); 12681e0482f5SStefano Zampini 12691e0482f5SStefano Zampini #if defined(PRINT_GDET) 12701e0482f5SStefano Zampini inc = 0; 12711e0482f5SStefano Zampini lev = pcbddc->current_level; 12721e0482f5SStefano Zampini #endif 1273213b8bfaSStefano Zampini 1274213b8bfaSStefano Zampini /* Insert values in the change of basis matrix */ 1275a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1276a13144ffSStefano Zampini Mat Gins = NULL, GKins = NULL; 12771e0482f5SStefano Zampini IS cornersis = NULL; 12781e0482f5SStefano Zampini PetscScalar cvals[2]; 1279a13144ffSStefano Zampini 12801e0482f5SStefano Zampini if (pcbddc->nedcG) { 12811e0482f5SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,2,corners+2*i,PETSC_USE_POINTER,&cornersis);CHKERRQ(ierr); 12821e0482f5SStefano Zampini } 12831e0482f5SStefano Zampini ierr = PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],cornersis,&Gins,&GKins,cvals,work,rwork);CHKERRQ(ierr); 1284a13144ffSStefano Zampini if (Gins && GKins) { 12851683a169SBarry Smith const PetscScalar *data; 1286a13144ffSStefano Zampini const PetscInt *rows,*cols; 1287a13144ffSStefano Zampini PetscInt nrh,nch,nrc,ncc; 1288a13144ffSStefano Zampini 1289a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&cols);CHKERRQ(ierr); 1290a13144ffSStefano Zampini /* H1 */ 1291a13144ffSStefano Zampini ierr = ISGetIndices(extrows[i],&rows);CHKERRQ(ierr); 1292a13144ffSStefano Zampini ierr = MatGetSize(Gins,&nrh,&nch);CHKERRQ(ierr); 12931683a169SBarry Smith ierr = MatDenseGetArrayRead(Gins,&data);CHKERRQ(ierr); 1294a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);CHKERRQ(ierr); 12951683a169SBarry Smith ierr = MatDenseRestoreArrayRead(Gins,&data);CHKERRQ(ierr); 1296a13144ffSStefano Zampini ierr = ISRestoreIndices(extrows[i],&rows);CHKERRQ(ierr); 1297a13144ffSStefano Zampini /* complement */ 1298a13144ffSStefano Zampini ierr = MatGetSize(GKins,&nrc,&ncc);CHKERRQ(ierr); 12996080607fSStefano Zampini if (!ncc) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Constant function has not been generated for coarse edge %D",i); 13006080607fSStefano Zampini if (ncc + nch != nrc) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"The sum of the number of columns of GKins %D and Gins %D does not match %D for coarse edge %D",ncc,nch,nrc,i); 13016080607fSStefano Zampini if (ncc != 1 && pcbddc->nedcG) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot generate the coarse discrete gradient for coarse edge %D with ncc %D",i,ncc); 13021683a169SBarry Smith ierr = MatDenseGetArrayRead(GKins,&data);CHKERRQ(ierr); 1303a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);CHKERRQ(ierr); 13041683a169SBarry Smith ierr = MatDenseRestoreArrayRead(GKins,&data);CHKERRQ(ierr); 13051e0482f5SStefano Zampini 13061e0482f5SStefano Zampini /* coarse discrete gradient */ 13071e0482f5SStefano Zampini if (pcbddc->nedcG) { 13081e0482f5SStefano Zampini PetscInt cols[2]; 13091e0482f5SStefano Zampini 13101e0482f5SStefano Zampini cols[0] = 2*i; 13111e0482f5SStefano Zampini cols[1] = 2*i+1; 13121e0482f5SStefano Zampini ierr = MatSetValuesLocal(pcbddc->nedcG,1,&i,2,cols,cvals,INSERT_VALUES);CHKERRQ(ierr); 13131e0482f5SStefano Zampini } 1314a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&cols);CHKERRQ(ierr); 1315a13144ffSStefano Zampini } 1316a13144ffSStefano Zampini ierr = ISDestroy(&extrows[i]);CHKERRQ(ierr); 1317a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 13181e0482f5SStefano Zampini ierr = ISDestroy(&cornersis);CHKERRQ(ierr); 1319a13144ffSStefano Zampini ierr = MatDestroy(&Gins);CHKERRQ(ierr); 1320a13144ffSStefano Zampini ierr = MatDestroy(&GKins);CHKERRQ(ierr); 1321a13144ffSStefano Zampini } 1322213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&el2g);CHKERRQ(ierr); 1323a13144ffSStefano Zampini 1324a13144ffSStefano Zampini /* Start assembling */ 1325a13144ffSStefano Zampini ierr = MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13261e0482f5SStefano Zampini if (pcbddc->nedcG) { 13271e0482f5SStefano Zampini ierr = MatAssemblyBegin(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13281e0482f5SStefano Zampini } 1329a13144ffSStefano Zampini 1330a13144ffSStefano Zampini /* Free */ 1331c2151214SStefano Zampini if (fl2g) { 1332c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1333c2151214SStefano Zampini for (i=0;i<nee;i++) { 1334c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1335c2151214SStefano Zampini } 1336c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1337c2151214SStefano Zampini } 1338eee23b56SStefano Zampini 1339eee23b56SStefano Zampini /* hack mat_graph with primal dofs on the coarse edges */ 1340eee23b56SStefano Zampini { 1341eee23b56SStefano Zampini PCBDDCGraph graph = pcbddc->mat_graph; 1342eee23b56SStefano Zampini PetscInt *oqueue = graph->queue; 1343eee23b56SStefano Zampini PetscInt *ocptr = graph->cptr; 1344eee23b56SStefano Zampini PetscInt ncc,*idxs; 1345eee23b56SStefano Zampini 1346eee23b56SStefano Zampini /* find first primal edge */ 1347eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1348eee23b56SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1349eee23b56SStefano Zampini } else { 1350eee23b56SStefano Zampini if (fl2g) { 1351eee23b56SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,nee,cedges,cedges);CHKERRQ(ierr); 1352eee23b56SStefano Zampini } 1353eee23b56SStefano Zampini idxs = cedges; 1354eee23b56SStefano Zampini } 1355eee23b56SStefano Zampini cum = 0; 1356eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1357eee23b56SStefano Zampini 1358eee23b56SStefano Zampini /* adapt connected components */ 1359eee23b56SStefano Zampini ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,ocptr[graph->ncc],&graph->queue);CHKERRQ(ierr); 1360eee23b56SStefano Zampini graph->cptr[0] = 0; 1361eee23b56SStefano Zampini for (i=0,ncc=0;i<graph->ncc;i++) { 1362eee23b56SStefano Zampini PetscInt lc = ocptr[i+1]-ocptr[i]; 1363eee23b56SStefano Zampini if (cum != nee && oqueue[ocptr[i+1]-1] == cedges[cum]) { /* this cc has a primal dof */ 1364eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc]+1; 1365eee23b56SStefano Zampini graph->queue[graph->cptr[ncc]] = cedges[cum]; 1366eee23b56SStefano Zampini ncc++; 1367eee23b56SStefano Zampini lc--; 1368eee23b56SStefano Zampini cum++; 1369eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1370eee23b56SStefano Zampini } 1371eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc] + lc; 1372eee23b56SStefano Zampini for (j=0;j<lc;j++) graph->queue[graph->cptr[ncc]+j] = oqueue[ocptr[i]+j]; 1373eee23b56SStefano Zampini ncc++; 1374eee23b56SStefano Zampini } 1375eee23b56SStefano Zampini graph->ncc = ncc; 1376eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1377eee23b56SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1378eee23b56SStefano Zampini } 1379eee23b56SStefano Zampini ierr = PetscFree2(ocptr,oqueue);CHKERRQ(ierr); 1380eee23b56SStefano Zampini } 1381213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&fl2g);CHKERRQ(ierr); 1382c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1383c2151214SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1384213b8bfaSStefano Zampini ierr = MatDestroy(&conn);CHKERRQ(ierr); 1385eee23b56SStefano Zampini 1386c2151214SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 1387a13144ffSStefano Zampini ierr = PetscFree(extrow);CHKERRQ(ierr); 1388a13144ffSStefano Zampini ierr = PetscFree2(work,rwork);CHKERRQ(ierr); 1389b03ebc13SStefano Zampini ierr = PetscFree(corners);CHKERRQ(ierr); 1390b03ebc13SStefano Zampini ierr = PetscFree(cedges);CHKERRQ(ierr); 1391a13144ffSStefano Zampini ierr = PetscFree(extrows);CHKERRQ(ierr); 1392a13144ffSStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 1393a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 1394a13144ffSStefano Zampini 1395a13144ffSStefano Zampini /* Complete assembling */ 1396a13144ffSStefano Zampini ierr = MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13971e0482f5SStefano Zampini if (pcbddc->nedcG) { 13981e0482f5SStefano Zampini ierr = MatAssemblyEnd(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13991e0482f5SStefano Zampini #if 0 14001e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->nedcG,"coarse_G");CHKERRQ(ierr); 14011e0482f5SStefano Zampini ierr = MatView(pcbddc->nedcG,NULL);CHKERRQ(ierr); 14021e0482f5SStefano Zampini #endif 14031e0482f5SStefano Zampini } 1404a13144ffSStefano Zampini 1405a13144ffSStefano Zampini /* set change of basis */ 1406213b8bfaSStefano Zampini ierr = PCBDDCSetChangeOfBasisMat(pc,T,singular);CHKERRQ(ierr); 1407a13144ffSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1408a13144ffSStefano Zampini 1409a13144ffSStefano Zampini PetscFunctionReturn(0); 1410a13144ffSStefano Zampini } 1411a13144ffSStefano Zampini 1412d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 1413d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 1414d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 1415d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 1416d8203eabSStefano Zampini { 1417d8203eabSStefano Zampini PetscErrorCode ierr; 1418d8203eabSStefano Zampini PetscInt i; 1419d8203eabSStefano Zampini 1420d8203eabSStefano Zampini PetscFunctionBegin; 1421d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 1422d8203eabSStefano Zampini PetscInt first,last; 1423d8203eabSStefano Zampini 1424d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 142586fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 1426d8203eabSStefano Zampini if (i>=first && i < last) { 1427d8203eabSStefano Zampini PetscScalar *data; 1428d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1429d8203eabSStefano Zampini if (!has_const) { 1430d8203eabSStefano Zampini data[i-first] = 1.; 1431d8203eabSStefano Zampini } else { 143286fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 143386fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 1434d8203eabSStefano Zampini } 1435d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1436d8203eabSStefano Zampini } 1437d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1438d8203eabSStefano Zampini } 1439d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 1440d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 1441d8203eabSStefano Zampini PetscInt first,last; 14428860a134SJunchao Zhang ierr = VecLockReadPop(quad_vecs[i]);CHKERRQ(ierr); 1443d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 1444d8203eabSStefano Zampini if (i>=first && i < last) { 1445d8203eabSStefano Zampini PetscScalar *data; 1446d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1447d8203eabSStefano Zampini if (!has_const) { 1448d8203eabSStefano Zampini data[i-first] = 0.; 1449d8203eabSStefano Zampini } else { 145086fa73c5SStefano Zampini data[2*i-first] = 0.; 145186fa73c5SStefano Zampini data[2*i-first+1] = 0.; 1452d8203eabSStefano Zampini } 1453d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1454d8203eabSStefano Zampini } 1455d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 14568860a134SJunchao Zhang ierr = VecLockReadPush(quad_vecs[i]);CHKERRQ(ierr); 1457d8203eabSStefano Zampini } 1458d8203eabSStefano Zampini PetscFunctionReturn(0); 1459d8203eabSStefano Zampini } 1460d8203eabSStefano Zampini 14618ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 1462669cc0f4SStefano Zampini { 1463a198735bSStefano Zampini Mat loc_divudotp; 1464fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 14658ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 1466669cc0f4SStefano Zampini PetscScalar *vals; 1467669cc0f4SStefano Zampini const PetscScalar *array; 14680f04eeffSStefano Zampini PetscInt i,maxneighs = 0,maxsize,*gidxs; 1469a040e873SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 14701ae86dd6SStefano Zampini PetscMPIInt rank; 1471a198735bSStefano Zampini PetscErrorCode ierr; 1472669cc0f4SStefano Zampini 1473669cc0f4SStefano Zampini PetscFunctionBegin; 1474a040e873SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 14750f04eeffSStefano Zampini for (i=0;i<n_neigh;i++) maxneighs = PetscMax(graph->count[shared[i][0]]+1,maxneighs); 1476820f2d46SBarry Smith ierr = MPIU_Allreduce(MPI_IN_PLACE,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRMPI(ierr); 14778037d520SStefano Zampini if (!maxneighs) { 14788037d520SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 14798037d520SStefano Zampini *nnsp = NULL; 14808037d520SStefano Zampini PetscFunctionReturn(0); 1481669cc0f4SStefano Zampini } 1482669cc0f4SStefano Zampini maxsize = 0; 1483a040e873SStefano Zampini for (i=0;i<n_neigh;i++) maxsize = PetscMax(n_shared[i],maxsize); 14840f27d399SStefano Zampini ierr = PetscMalloc2(maxsize,&gidxs,maxsize,&vals);CHKERRQ(ierr); 1485669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 1486669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 14878ae0ca82SStefano Zampini if (!transpose) { 14888ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 14898ae0ca82SStefano Zampini } else { 14908ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 14918ae0ca82SStefano Zampini } 1492669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 14931ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 1494d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 1495669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 14968860a134SJunchao Zhang ierr = VecLockReadPop(quad_vecs[i]);CHKERRQ(ierr); 1497669cc0f4SStefano Zampini } 1498d8203eabSStefano Zampini 1499669cc0f4SStefano Zampini /* compute local quad vec */ 1500a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 15018ae0ca82SStefano Zampini if (!transpose) { 1502a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 15038ae0ca82SStefano Zampini } else { 15048ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 15058ae0ca82SStefano Zampini } 1506669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 15078ae0ca82SStefano Zampini if (!transpose) { 1508a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 15098ae0ca82SStefano Zampini } else { 15108ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 15118ae0ca82SStefano Zampini } 1512fa23a32eSStefano Zampini if (vl2l) { 1513187c917aSStefano Zampini Mat lA; 1514187c917aSStefano Zampini VecScatter sc; 1515187c917aSStefano Zampini 1516187c917aSStefano Zampini ierr = MatISGetLocalMat(A,&lA);CHKERRQ(ierr); 1517187c917aSStefano Zampini ierr = MatCreateVecs(lA,&vins,NULL);CHKERRQ(ierr); 15189448b7f1SJunchao Zhang ierr = VecScatterCreate(v,NULL,vins,vl2l,&sc);CHKERRQ(ierr); 1519187c917aSStefano Zampini ierr = VecScatterBegin(sc,v,vins,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1520187c917aSStefano Zampini ierr = VecScatterEnd(sc,v,vins,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1521187c917aSStefano Zampini ierr = VecScatterDestroy(&sc);CHKERRQ(ierr); 1522fa23a32eSStefano Zampini } else { 1523fa23a32eSStefano Zampini vins = v; 1524fa23a32eSStefano Zampini } 1525fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 1526669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 15279a962809SStefano Zampini 15281ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 1529ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRMPI(ierr); 15300f04eeffSStefano Zampini for (i=1;i<n_neigh;i++) { 1531669cc0f4SStefano Zampini const PetscInt *idxs; 1532669cc0f4SStefano Zampini PetscInt idx,nn,j; 1533669cc0f4SStefano Zampini 1534a040e873SStefano Zampini idxs = shared[i]; 1535a040e873SStefano Zampini nn = n_shared[i]; 1536669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 15371ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1538669cc0f4SStefano Zampini idx = -(idx+1); 15390f04eeffSStefano Zampini if (idx < 0 || idx >= maxneighs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid index %D not in [0,%D)",idx,maxneighs); 15400f27d399SStefano Zampini ierr = ISLocalToGlobalMappingApply(map,nn,idxs,gidxs);CHKERRQ(ierr); 15410f27d399SStefano Zampini ierr = VecSetValues(quad_vecs[idx],nn,gidxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1542669cc0f4SStefano Zampini } 1543a040e873SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 1544fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 1545fa23a32eSStefano Zampini if (vl2l) { 1546187c917aSStefano Zampini ierr = VecDestroy(&vins);CHKERRQ(ierr); 1547fa23a32eSStefano Zampini } 1548669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 15490f27d399SStefano Zampini ierr = PetscFree2(gidxs,vals);CHKERRQ(ierr); 1550669cc0f4SStefano Zampini 1551669cc0f4SStefano Zampini /* assemble near null space */ 1552669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1553669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 1554669cc0f4SStefano Zampini } 1555669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1556669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 15573272d46bSStefano Zampini ierr = VecViewFromOptions(quad_vecs[i],NULL,"-pc_bddc_quad_vecs_view");CHKERRQ(ierr); 15588860a134SJunchao Zhang ierr = VecLockReadPush(quad_vecs[i]);CHKERRQ(ierr); 1559669cc0f4SStefano Zampini } 1560669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 1561669cc0f4SStefano Zampini PetscFunctionReturn(0); 1562669cc0f4SStefano Zampini } 1563669cc0f4SStefano Zampini 15647620a527SStefano Zampini PetscErrorCode PCBDDCAddPrimalVerticesLocalIS(PC pc, IS primalv) 15657620a527SStefano Zampini { 15667620a527SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 15677620a527SStefano Zampini PetscErrorCode ierr; 15687620a527SStefano Zampini 15697620a527SStefano Zampini PetscFunctionBegin; 15707620a527SStefano Zampini if (primalv) { 15717620a527SStefano Zampini if (pcbddc->user_primal_vertices_local) { 15727620a527SStefano Zampini IS list[2], newp; 15737620a527SStefano Zampini 15747620a527SStefano Zampini list[0] = primalv; 15757620a527SStefano Zampini list[1] = pcbddc->user_primal_vertices_local; 15767620a527SStefano Zampini ierr = ISConcatenate(PetscObjectComm((PetscObject)pc),2,list,&newp);CHKERRQ(ierr); 15777620a527SStefano Zampini ierr = ISSortRemoveDups(newp);CHKERRQ(ierr); 15787620a527SStefano Zampini ierr = ISDestroy(&list[1]);CHKERRQ(ierr); 15797620a527SStefano Zampini pcbddc->user_primal_vertices_local = newp; 15807620a527SStefano Zampini } else { 15817620a527SStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primalv);CHKERRQ(ierr); 15827620a527SStefano Zampini } 15837620a527SStefano Zampini } 15847620a527SStefano Zampini PetscFunctionReturn(0); 15857620a527SStefano Zampini } 1586669cc0f4SStefano Zampini 15871c7a958bSStefano Zampini static PetscErrorCode func_coords_private(PetscInt dim, PetscReal t, const PetscReal X[], PetscInt Nf, PetscScalar *out, void *ctx) 15881c7a958bSStefano Zampini { 15891c7a958bSStefano Zampini PetscInt f, *comp = (PetscInt *)ctx; 15901c7a958bSStefano Zampini 15911c7a958bSStefano Zampini PetscFunctionBegin; 15921c7a958bSStefano Zampini for (f=0;f<Nf;f++) out[f] = X[*comp]; 15931c7a958bSStefano Zampini PetscFunctionReturn(0); 15941c7a958bSStefano Zampini } 1595674ae819SStefano Zampini 15961f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 15971f4df5f7SStefano Zampini { 15981f4df5f7SStefano Zampini PetscErrorCode ierr; 15991f4df5f7SStefano Zampini Vec local,global; 16001f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16011f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 16025c5e10d6SStefano Zampini PetscBool monolithic = PETSC_FALSE; 16031f4df5f7SStefano Zampini 16041f4df5f7SStefano Zampini PetscFunctionBegin; 16055c5e10d6SStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC topology options","PC");CHKERRQ(ierr); 16065c5e10d6SStefano Zampini ierr = PetscOptionsBool("-pc_bddc_monolithic","Discard any information on dofs splitting",NULL,monolithic,&monolithic,NULL);CHKERRQ(ierr); 16075c5e10d6SStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 16081f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 160921ef3d20SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 16101f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 1611b470e4b4SRichard Tran Mills ierr = VecBindToCPU(global,PETSC_TRUE);CHKERRQ(ierr); 1612b470e4b4SRichard Tran Mills ierr = VecBindToCPU(local,PETSC_TRUE);CHKERRQ(ierr); 16136a8fc67bSStefano Zampini if (monolithic) { /* just get block size to properly compute vertices */ 16146a8fc67bSStefano Zampini if (pcbddc->vertex_size == 1) { 16156a8fc67bSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->vertex_size);CHKERRQ(ierr); 16166a8fc67bSStefano Zampini } 16176a8fc67bSStefano Zampini goto boundary; 16186a8fc67bSStefano Zampini } 16195c5e10d6SStefano Zampini 16201f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 16211f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 16221f4df5f7SStefano Zampini PetscInt i; 16230c85b387SStefano Zampini 16241f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16251f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 16260c85b387SStefano Zampini PetscInt bs; 16270c85b387SStefano Zampini 16281f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16290c85b387SStefano Zampini ierr = ISGetBlockSize(pcbddc->ISForDofs[i],&bs);CHKERRQ(ierr); 16300c85b387SStefano Zampini ierr = ISSetBlockSize(pcbddc->ISForDofsLocal[i],bs);CHKERRQ(ierr); 16311f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 16321f4df5f7SStefano Zampini } 16331f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 16341f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 16351f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 16361f4df5f7SStefano Zampini } 16371f4df5f7SStefano Zampini } else { 163821ef3d20SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present */ 163921ef3d20SStefano Zampini DM dm; 164021ef3d20SStefano Zampini 164121ef3d20SStefano Zampini ierr = MatGetDM(pc->pmat, &dm);CHKERRQ(ierr); 16424f819b78SStefano Zampini if (!dm) { 16434f819b78SStefano Zampini ierr = PCGetDM(pc, &dm);CHKERRQ(ierr); 164421ef3d20SStefano Zampini } 164521ef3d20SStefano Zampini if (dm) { 164621ef3d20SStefano Zampini IS *fields; 164721ef3d20SStefano Zampini PetscInt nf,i; 16480c85b387SStefano Zampini 164921ef3d20SStefano Zampini ierr = DMCreateFieldDecomposition(dm,&nf,NULL,&fields,NULL);CHKERRQ(ierr); 165021ef3d20SStefano Zampini ierr = PetscMalloc1(nf,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 165121ef3d20SStefano Zampini for (i=0;i<nf;i++) { 16520c85b387SStefano Zampini PetscInt bs; 16530c85b387SStefano Zampini 165421ef3d20SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,fields[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16550c85b387SStefano Zampini ierr = ISGetBlockSize(fields[i],&bs);CHKERRQ(ierr); 16560c85b387SStefano Zampini ierr = ISSetBlockSize(pcbddc->ISForDofsLocal[i],bs);CHKERRQ(ierr); 165721ef3d20SStefano Zampini ierr = ISDestroy(&fields[i]);CHKERRQ(ierr); 165821ef3d20SStefano Zampini } 165921ef3d20SStefano Zampini ierr = PetscFree(fields);CHKERRQ(ierr); 166021ef3d20SStefano Zampini pcbddc->n_ISForDofsLocal = nf; 166121ef3d20SStefano Zampini } else { /* See if MATIS has fields attached by the conversion from MatNest */ 166221ef3d20SStefano Zampini PetscContainer c; 166321ef3d20SStefano Zampini 166421ef3d20SStefano Zampini ierr = PetscObjectQuery((PetscObject)pc->pmat,"_convert_nest_lfields",(PetscObject*)&c);CHKERRQ(ierr); 166521ef3d20SStefano Zampini if (c) { 166621ef3d20SStefano Zampini MatISLocalFields lf; 166721ef3d20SStefano Zampini ierr = PetscContainerGetPointer(c,(void**)&lf);CHKERRQ(ierr); 166821ef3d20SStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,lf->nr,lf->rf);CHKERRQ(ierr); 166921ef3d20SStefano Zampini } else { /* fallback, create the default fields if bs > 1 */ 16701f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1671986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 167221ef3d20SStefano Zampini if (i > 1) { 1673986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 16741f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16751f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16761f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16771f4df5f7SStefano Zampini } 16781f4df5f7SStefano Zampini } 167921ef3d20SStefano Zampini } 168021ef3d20SStefano Zampini } 16817a0e7b2cSstefano_zampini } else { 16827a0e7b2cSstefano_zampini PetscInt i; 16837a0e7b2cSstefano_zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16847a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16857a0e7b2cSstefano_zampini } 16861f4df5f7SStefano Zampini } 1687986cdee1SStefano Zampini } 16881f4df5f7SStefano Zampini 16895c5e10d6SStefano Zampini boundary: 16901f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 16911f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16927a0e7b2cSstefano_zampini } else if (pcbddc->DirichletBoundariesLocal) { 16937a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16941f4df5f7SStefano Zampini } 16951f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 16961f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16977a0e7b2cSstefano_zampini } else if (pcbddc->NeumannBoundariesLocal) { 16987a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16991f4df5f7SStefano Zampini } 17001f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 17011f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 17021f4df5f7SStefano Zampini } 17031f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 17041f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 17057620a527SStefano Zampini /* detect local disconnected subdomains if requested (use matis->A) */ 17067620a527SStefano Zampini if (pcbddc->detect_disconnected) { 17077620a527SStefano Zampini IS primalv = NULL; 17087620a527SStefano Zampini PetscInt i; 17098361f951SStefano Zampini PetscBool filter = pcbddc->detect_disconnected_filter; 17107a0e7b2cSstefano_zampini 17117620a527SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 17127620a527SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 17137620a527SStefano Zampini } 17147620a527SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 17158361f951SStefano Zampini ierr = PCBDDCDetectDisconnectedComponents(pc,filter,&pcbddc->n_local_subs,&pcbddc->local_subs,&primalv);CHKERRQ(ierr); 17167620a527SStefano Zampini ierr = PCBDDCAddPrimalVerticesLocalIS(pc,primalv);CHKERRQ(ierr); 17177620a527SStefano Zampini ierr = ISDestroy(&primalv);CHKERRQ(ierr); 17187620a527SStefano Zampini } 17197620a527SStefano Zampini /* early stage corner detection */ 17207620a527SStefano Zampini { 17217620a527SStefano Zampini DM dm; 17227620a527SStefano Zampini 17237620a527SStefano Zampini ierr = MatGetDM(pc->pmat,&dm);CHKERRQ(ierr); 17244f819b78SStefano Zampini if (!dm) { 17254f819b78SStefano Zampini ierr = PCGetDM(pc,&dm);CHKERRQ(ierr); 17264f819b78SStefano Zampini } 17277620a527SStefano Zampini if (dm) { 17287620a527SStefano Zampini PetscBool isda; 17297620a527SStefano Zampini 17307620a527SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)dm,DMDA,&isda);CHKERRQ(ierr); 17317620a527SStefano Zampini if (isda) { 17327620a527SStefano Zampini ISLocalToGlobalMapping l2l; 17337620a527SStefano Zampini IS corners; 17347620a527SStefano Zampini Mat lA; 17354f819b78SStefano Zampini PetscBool gl,lo; 17367620a527SStefano Zampini 17374f819b78SStefano Zampini { 17384f819b78SStefano Zampini Vec cvec; 17394f819b78SStefano Zampini const PetscScalar *coords; 17404f819b78SStefano Zampini PetscInt dof,n,cdim; 17414f819b78SStefano Zampini PetscBool memc = PETSC_TRUE; 17424f819b78SStefano Zampini 17434f819b78SStefano Zampini ierr = DMDAGetInfo(dm,NULL,NULL,NULL,NULL,NULL,NULL,NULL,&dof,NULL,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 17444f819b78SStefano Zampini ierr = DMGetCoordinates(dm,&cvec);CHKERRQ(ierr); 17454f819b78SStefano Zampini ierr = VecGetLocalSize(cvec,&n);CHKERRQ(ierr); 17464f819b78SStefano Zampini ierr = VecGetBlockSize(cvec,&cdim);CHKERRQ(ierr); 17474f819b78SStefano Zampini n /= cdim; 17484f819b78SStefano Zampini ierr = PetscFree(pcbddc->mat_graph->coords);CHKERRQ(ierr); 17494f819b78SStefano Zampini ierr = PetscMalloc1(dof*n*cdim,&pcbddc->mat_graph->coords);CHKERRQ(ierr); 17504f819b78SStefano Zampini ierr = VecGetArrayRead(cvec,&coords);CHKERRQ(ierr); 17514f819b78SStefano Zampini #if defined(PETSC_USE_COMPLEX) 17524f819b78SStefano Zampini memc = PETSC_FALSE; 17534f819b78SStefano Zampini #endif 17544f819b78SStefano Zampini if (dof != 1) memc = PETSC_FALSE; 17554f819b78SStefano Zampini if (memc) { 1756580bdb30SBarry Smith ierr = PetscArraycpy(pcbddc->mat_graph->coords,coords,cdim*n*dof);CHKERRQ(ierr); 17574f819b78SStefano Zampini } else { /* BDDC graph does not use any blocked information, we need to replicate the data */ 17584f819b78SStefano Zampini PetscReal *bcoords = pcbddc->mat_graph->coords; 17594f819b78SStefano Zampini PetscInt i, b, d; 17604f819b78SStefano Zampini 17614f819b78SStefano Zampini for (i=0;i<n;i++) { 17624f819b78SStefano Zampini for (b=0;b<dof;b++) { 17634f819b78SStefano Zampini for (d=0;d<cdim;d++) { 17644f819b78SStefano Zampini bcoords[i*dof*cdim + b*cdim + d] = PetscRealPart(coords[i*cdim+d]); 17654f819b78SStefano Zampini } 17664f819b78SStefano Zampini } 17674f819b78SStefano Zampini } 17684f819b78SStefano Zampini } 17694f819b78SStefano Zampini ierr = VecRestoreArrayRead(cvec,&coords);CHKERRQ(ierr); 17704f819b78SStefano Zampini pcbddc->mat_graph->cdim = cdim; 17714f819b78SStefano Zampini pcbddc->mat_graph->cnloc = dof*n; 17724f819b78SStefano Zampini pcbddc->mat_graph->cloc = PETSC_FALSE; 17734f819b78SStefano Zampini } 1774d4a6ed37SStefano Zampini ierr = DMDAGetSubdomainCornersIS(dm,&corners);CHKERRQ(ierr); 17757620a527SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&lA);CHKERRQ(ierr); 17767620a527SStefano Zampini ierr = MatGetLocalToGlobalMapping(lA,&l2l,NULL);CHKERRQ(ierr); 17777620a527SStefano Zampini ierr = MatISRestoreLocalMat(pc->pmat,&lA);CHKERRQ(ierr); 17784f819b78SStefano Zampini lo = (PetscBool)(l2l && corners); 1779820f2d46SBarry Smith ierr = MPIU_Allreduce(&lo,&gl,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 17804f819b78SStefano Zampini if (gl) { /* From PETSc's DMDA */ 17817620a527SStefano Zampini const PetscInt *idx; 178272ed36d8SStefano Zampini PetscInt dof,bs,*idxout,n; 17837620a527SStefano Zampini 178472ed36d8SStefano Zampini ierr = DMDAGetInfo(dm,NULL,NULL,NULL,NULL,NULL,NULL,NULL,&dof,NULL,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 17857620a527SStefano Zampini ierr = ISLocalToGlobalMappingGetBlockSize(l2l,&bs);CHKERRQ(ierr); 17867620a527SStefano Zampini ierr = ISGetLocalSize(corners,&n);CHKERRQ(ierr); 17877620a527SStefano Zampini ierr = ISGetIndices(corners,&idx);CHKERRQ(ierr); 178872ed36d8SStefano Zampini if (bs == dof) { 17897620a527SStefano Zampini ierr = PetscMalloc1(n,&idxout);CHKERRQ(ierr); 17907620a527SStefano Zampini ierr = ISLocalToGlobalMappingApplyBlock(l2l,n,idx,idxout);CHKERRQ(ierr); 179172ed36d8SStefano Zampini } else { /* the original DMDA local-to-local map have been modified */ 179272ed36d8SStefano Zampini PetscInt i,d; 179372ed36d8SStefano Zampini 179472ed36d8SStefano Zampini ierr = PetscMalloc1(dof*n,&idxout);CHKERRQ(ierr); 179572ed36d8SStefano Zampini for (i=0;i<n;i++) for (d=0;d<dof;d++) idxout[dof*i+d] = dof*idx[i]+d; 179672ed36d8SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2l,dof*n,idxout,idxout);CHKERRQ(ierr); 179772ed36d8SStefano Zampini 179872ed36d8SStefano Zampini bs = 1; 179972ed36d8SStefano Zampini n *= dof; 180072ed36d8SStefano Zampini } 18017620a527SStefano Zampini ierr = ISRestoreIndices(corners,&idx);CHKERRQ(ierr); 1802d4a6ed37SStefano Zampini ierr = DMDARestoreSubdomainCornersIS(dm,&corners);CHKERRQ(ierr); 18037620a527SStefano Zampini ierr = ISCreateBlock(PetscObjectComm((PetscObject)pc),bs,n,idxout,PETSC_OWN_POINTER,&corners);CHKERRQ(ierr); 18047620a527SStefano Zampini ierr = PCBDDCAddPrimalVerticesLocalIS(pc,corners);CHKERRQ(ierr); 18057620a527SStefano Zampini ierr = ISDestroy(&corners);CHKERRQ(ierr); 18061c7a958bSStefano Zampini pcbddc->corner_selected = PETSC_TRUE; 18074f819b78SStefano Zampini pcbddc->corner_selection = PETSC_TRUE; 18084f819b78SStefano Zampini } 18094f819b78SStefano Zampini if (corners) { 1810d4a6ed37SStefano Zampini ierr = DMDARestoreSubdomainCornersIS(dm,&corners);CHKERRQ(ierr); 18117620a527SStefano Zampini } 18127620a527SStefano Zampini } 18137620a527SStefano Zampini } 18147620a527SStefano Zampini } 18151c7a958bSStefano Zampini if (pcbddc->corner_selection && !pcbddc->mat_graph->cdim) { 18161c7a958bSStefano Zampini DM dm; 18171c7a958bSStefano Zampini 18181c7a958bSStefano Zampini ierr = MatGetDM(pc->pmat,&dm);CHKERRQ(ierr); 18194f819b78SStefano Zampini if (!dm) { 18204f819b78SStefano Zampini ierr = PCGetDM(pc,&dm);CHKERRQ(ierr); 18211c7a958bSStefano Zampini } 18224f819b78SStefano Zampini if (dm) { /* this can get very expensive, I need to find a faster alternative */ 18231c7a958bSStefano Zampini Vec vcoords; 18241c7a958bSStefano Zampini PetscSection section; 18251c7a958bSStefano Zampini PetscReal *coords; 18261c7a958bSStefano Zampini PetscInt d,cdim,nl,nf,**ctxs; 18271c7a958bSStefano Zampini PetscErrorCode (**funcs)(PetscInt, PetscReal, const PetscReal *, PetscInt, PetscScalar *, void *); 18281c7a958bSStefano Zampini 18291c7a958bSStefano Zampini ierr = DMGetCoordinateDim(dm,&cdim);CHKERRQ(ierr); 183092fd8e1eSJed Brown ierr = DMGetLocalSection(dm,§ion);CHKERRQ(ierr); 18311c7a958bSStefano Zampini ierr = PetscSectionGetNumFields(section,&nf);CHKERRQ(ierr); 18321c7a958bSStefano Zampini ierr = DMCreateGlobalVector(dm,&vcoords);CHKERRQ(ierr); 18331c7a958bSStefano Zampini ierr = VecGetLocalSize(vcoords,&nl);CHKERRQ(ierr); 18341c7a958bSStefano Zampini ierr = PetscMalloc1(nl*cdim,&coords);CHKERRQ(ierr); 18351c7a958bSStefano Zampini ierr = PetscMalloc2(nf,&funcs,nf,&ctxs);CHKERRQ(ierr); 18361c7a958bSStefano Zampini ierr = PetscMalloc1(nf,&ctxs[0]);CHKERRQ(ierr); 18371c7a958bSStefano Zampini for (d=0;d<nf;d++) funcs[d] = func_coords_private; 18381c7a958bSStefano Zampini for (d=1;d<nf;d++) ctxs[d] = ctxs[d-1] + 1; 18391c7a958bSStefano Zampini for (d=0;d<cdim;d++) { 18401c7a958bSStefano Zampini PetscInt i; 18411c7a958bSStefano Zampini const PetscScalar *v; 18421c7a958bSStefano Zampini 18431c7a958bSStefano Zampini for (i=0;i<nf;i++) ctxs[i][0] = d; 18441c7a958bSStefano Zampini ierr = DMProjectFunction(dm,0.0,funcs,(void**)ctxs,INSERT_VALUES,vcoords);CHKERRQ(ierr); 18451c7a958bSStefano Zampini ierr = VecGetArrayRead(vcoords,&v);CHKERRQ(ierr); 18461c7a958bSStefano Zampini for (i=0;i<nl;i++) coords[i*cdim+d] = PetscRealPart(v[i]); 18471c7a958bSStefano Zampini ierr = VecRestoreArrayRead(vcoords,&v);CHKERRQ(ierr); 18481c7a958bSStefano Zampini } 18491c7a958bSStefano Zampini ierr = VecDestroy(&vcoords);CHKERRQ(ierr); 18501c7a958bSStefano Zampini ierr = PCSetCoordinates(pc,cdim,nl,coords);CHKERRQ(ierr); 18511c7a958bSStefano Zampini ierr = PetscFree(coords);CHKERRQ(ierr); 18521c7a958bSStefano Zampini ierr = PetscFree(ctxs[0]);CHKERRQ(ierr); 18531c7a958bSStefano Zampini ierr = PetscFree2(funcs,ctxs);CHKERRQ(ierr); 18541c7a958bSStefano Zampini } 18551c7a958bSStefano Zampini } 18567a0e7b2cSstefano_zampini PetscFunctionReturn(0); 18577a0e7b2cSstefano_zampini } 18587a0e7b2cSstefano_zampini 18597a0e7b2cSstefano_zampini PetscErrorCode PCBDDCConsistencyCheckIS(PC pc, MPI_Op mop, IS *is) 18607a0e7b2cSstefano_zampini { 18617a0e7b2cSstefano_zampini Mat_IS *matis = (Mat_IS*)(pc->pmat->data); 18627a0e7b2cSstefano_zampini PetscErrorCode ierr; 18637a0e7b2cSstefano_zampini IS nis; 18647a0e7b2cSstefano_zampini const PetscInt *idxs; 18657a0e7b2cSstefano_zampini PetscInt i,nd,n = matis->A->rmap->n,*nidxs,nnd; 18667a0e7b2cSstefano_zampini PetscBool *ld; 18677a0e7b2cSstefano_zampini 18687a0e7b2cSstefano_zampini PetscFunctionBegin; 18697a0e7b2cSstefano_zampini if (mop != MPI_LAND && mop != MPI_LOR) SETERRQ(PetscObjectComm((PetscObject)(pc)),PETSC_ERR_SUP,"Supported are MPI_LAND and MPI_LOR"); 18707a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 18717a0e7b2cSstefano_zampini /* init rootdata with true */ 18727a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_rootdata; 18737a0e7b2cSstefano_zampini for (i=0;i<pc->pmat->rmap->n;i++) ld[i] = PETSC_TRUE; 18747a0e7b2cSstefano_zampini } else { 1875580bdb30SBarry Smith ierr = PetscArrayzero(matis->sf_rootdata,pc->pmat->rmap->n);CHKERRQ(ierr); 18767a0e7b2cSstefano_zampini } 1877580bdb30SBarry Smith ierr = PetscArrayzero(matis->sf_leafdata,n);CHKERRQ(ierr); 18787a0e7b2cSstefano_zampini ierr = ISGetLocalSize(*is,&nd);CHKERRQ(ierr); 18797a0e7b2cSstefano_zampini ierr = ISGetIndices(*is,&idxs);CHKERRQ(ierr); 18807a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_leafdata; 18817a0e7b2cSstefano_zampini for (i=0;i<nd;i++) 18827a0e7b2cSstefano_zampini if (-1 < idxs[i] && idxs[i] < n) 18837a0e7b2cSstefano_zampini ld[idxs[i]] = PETSC_TRUE; 18847a0e7b2cSstefano_zampini ierr = ISRestoreIndices(*is,&idxs);CHKERRQ(ierr); 18857a0e7b2cSstefano_zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 18867a0e7b2cSstefano_zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 1887ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata,MPI_REPLACE);CHKERRQ(ierr); 1888ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata,MPI_REPLACE);CHKERRQ(ierr); 18897a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 18907a0e7b2cSstefano_zampini ierr = PetscMalloc1(nd,&nidxs);CHKERRQ(ierr); 18917a0e7b2cSstefano_zampini } else { 18927a0e7b2cSstefano_zampini ierr = PetscMalloc1(n,&nidxs);CHKERRQ(ierr); 18937a0e7b2cSstefano_zampini } 18947a0e7b2cSstefano_zampini for (i=0,nnd=0;i<n;i++) 18957a0e7b2cSstefano_zampini if (ld[i]) 18967a0e7b2cSstefano_zampini nidxs[nnd++] = i; 18977a0e7b2cSstefano_zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(*is)),nnd,nidxs,PETSC_OWN_POINTER,&nis);CHKERRQ(ierr); 18987a0e7b2cSstefano_zampini ierr = ISDestroy(is);CHKERRQ(ierr); 18997a0e7b2cSstefano_zampini *is = nis; 19001f4df5f7SStefano Zampini PetscFunctionReturn(0); 19011f4df5f7SStefano Zampini } 19021f4df5f7SStefano Zampini 19033e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 19043e589ea0SStefano Zampini { 19053e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 19063e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 19073e589ea0SStefano Zampini PetscErrorCode ierr; 19083e589ea0SStefano Zampini 19093e589ea0SStefano Zampini PetscFunctionBegin; 19103e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 19113e589ea0SStefano Zampini PetscFunctionReturn(0); 19123e589ea0SStefano Zampini } 19133e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 19143e589ea0SStefano Zampini Vec swap; 19153e589ea0SStefano Zampini 19163e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 19173e589ea0SStefano Zampini swap = pcbddc->work_change; 19183e589ea0SStefano Zampini pcbddc->work_change = r; 19193e589ea0SStefano Zampini r = swap; 19203e589ea0SStefano Zampini } 19213e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 19223e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192355c176c0SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][0],pc,0,0,0);CHKERRQ(ierr); 19243e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 192555c176c0SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][0],pc,0,0,0);CHKERRQ(ierr); 1926c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_D,pc,pcis->vec2_D);CHKERRQ(ierr); 19273e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 19283e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 19293e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 19303e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1931f913dca9SStefano Zampini pcbddc->work_change = r; 19323e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 19333e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 19343e589ea0SStefano Zampini } 19353e589ea0SStefano Zampini PetscFunctionReturn(0); 19363e589ea0SStefano Zampini } 19373e589ea0SStefano Zampini 1938a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1939a3df083aSStefano Zampini { 1940a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1941a3df083aSStefano Zampini PetscErrorCode ierr; 1942a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1943a3df083aSStefano Zampini 1944a3df083aSStefano Zampini PetscFunctionBegin; 1945a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1946a3df083aSStefano Zampini if (transpose) { 1947a3df083aSStefano Zampini apply_right = ctx->apply_left; 1948a3df083aSStefano Zampini apply_left = ctx->apply_right; 1949a3df083aSStefano Zampini } else { 1950a3df083aSStefano Zampini apply_right = ctx->apply_right; 1951a3df083aSStefano Zampini apply_left = ctx->apply_left; 1952a3df083aSStefano Zampini } 1953a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1954a3df083aSStefano Zampini if (apply_right) { 1955a3df083aSStefano Zampini const PetscScalar *ax; 1956a3df083aSStefano Zampini PetscInt nl,i; 1957a3df083aSStefano Zampini 1958a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1959a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1960580bdb30SBarry Smith ierr = PetscArraycpy(ctx->work,ax,nl);CHKERRQ(ierr); 1961a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1962a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1963a3df083aSStefano Zampini PetscScalar sum,val; 1964a3df083aSStefano Zampini const PetscInt *idxs; 1965a3df083aSStefano Zampini PetscInt nz,j; 1966a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1967a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1968a3df083aSStefano Zampini sum = 0.; 1969a3df083aSStefano Zampini if (ctx->apply_p0) { 1970a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1971a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1972a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1973a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1974a3df083aSStefano Zampini } 1975a3df083aSStefano Zampini } else { 1976a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1977a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1978a3df083aSStefano Zampini } 1979a3df083aSStefano Zampini } 1980a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1981a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1982a3df083aSStefano Zampini } 1983a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1984a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1985a3df083aSStefano Zampini } 1986a3df083aSStefano Zampini if (transpose) { 1987a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1988a3df083aSStefano Zampini } else { 1989a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1990a3df083aSStefano Zampini } 1991a3df083aSStefano Zampini if (reset_x) { 1992a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1993a3df083aSStefano Zampini } 1994a3df083aSStefano Zampini if (apply_left) { 1995a3df083aSStefano Zampini PetscScalar *ay; 1996a3df083aSStefano Zampini PetscInt i; 1997a3df083aSStefano Zampini 1998a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1999a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 2000a3df083aSStefano Zampini PetscScalar sum,val; 2001a3df083aSStefano Zampini const PetscInt *idxs; 2002a3df083aSStefano Zampini PetscInt nz,j; 2003a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2004a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 2005a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 2006a3df083aSStefano Zampini if (ctx->apply_p0) { 2007a3df083aSStefano Zampini sum = 0.; 2008a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 2009a3df083aSStefano Zampini sum += ay[idxs[j]]; 2010a3df083aSStefano Zampini ay[idxs[j]] += val; 2011a3df083aSStefano Zampini } 2012a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 2013a3df083aSStefano Zampini } else { 2014a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 2015a3df083aSStefano Zampini ay[idxs[j]] += val; 2016a3df083aSStefano Zampini } 2017a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 2018a3df083aSStefano Zampini } 2019a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 2020a3df083aSStefano Zampini } 2021a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 2022a3df083aSStefano Zampini } 2023a3df083aSStefano Zampini PetscFunctionReturn(0); 2024a3df083aSStefano Zampini } 2025a3df083aSStefano Zampini 2026a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 2027a3df083aSStefano Zampini { 2028a3df083aSStefano Zampini PetscErrorCode ierr; 2029a3df083aSStefano Zampini 2030a3df083aSStefano Zampini PetscFunctionBegin; 2031a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 2032a3df083aSStefano Zampini PetscFunctionReturn(0); 2033a3df083aSStefano Zampini } 2034a3df083aSStefano Zampini 2035a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 2036a3df083aSStefano Zampini { 2037a3df083aSStefano Zampini PetscErrorCode ierr; 2038a3df083aSStefano Zampini 2039a3df083aSStefano Zampini PetscFunctionBegin; 2040a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 2041a3df083aSStefano Zampini PetscFunctionReturn(0); 2042a3df083aSStefano Zampini } 2043a3df083aSStefano Zampini 2044a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 2045a3df083aSStefano Zampini { 2046a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2047a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2048a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 2049a3df083aSStefano Zampini PetscErrorCode ierr; 2050a3df083aSStefano Zampini 2051a3df083aSStefano Zampini PetscFunctionBegin; 2052a3df083aSStefano Zampini if (!restore) { 20531dd7afcfSStefano Zampini Mat A_IB,A_BI; 2054a3df083aSStefano Zampini PetscScalar *work; 2055b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 2056a3df083aSStefano Zampini 20579a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 20589a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 2059a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 2060a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 2061a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2062a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 2063a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 2064a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 2065a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 2066a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 2067a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 2068a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 2069a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 2070a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 2071059032f7SStefano Zampini if (reuse) { 2072a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 20731dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 2074059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 2075059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 2076059032f7SStefano Zampini PetscInt i; 2077059032f7SStefano Zampini 2078059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 2079059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 2080059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2081059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 2082059032f7SStefano Zampini } 2083059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 20841dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 2085059032f7SStefano Zampini } 2086a3df083aSStefano Zampini ctx->A = pcis->A_IB; 2087a3df083aSStefano Zampini ctx->work = work; 2088a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 2089a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2090a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2091a3df083aSStefano Zampini pcis->A_IB = A_IB; 2092a3df083aSStefano Zampini 2093a3df083aSStefano Zampini /* A_BI as A_IB^T */ 2094a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 2095a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 2096a3df083aSStefano Zampini pcis->A_BI = A_BI; 2097a3df083aSStefano Zampini } else { 20981dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 20991dd7afcfSStefano Zampini PetscFunctionReturn(0); 21001dd7afcfSStefano Zampini } 2101a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 2102a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 2103a3df083aSStefano Zampini pcis->A_IB = ctx->A; 21041dd7afcfSStefano Zampini ctx->A = NULL; 21051dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 21061dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 21071dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 21081dd7afcfSStefano Zampini if (ctx->free) { 2109059032f7SStefano Zampini PetscInt i; 21101dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 2111059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 2112059032f7SStefano Zampini } 2113059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 2114059032f7SStefano Zampini } 2115a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 2116a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 2117a3df083aSStefano Zampini } 2118a3df083aSStefano Zampini PetscFunctionReturn(0); 2119a3df083aSStefano Zampini } 2120a3df083aSStefano Zampini 2121a3df083aSStefano Zampini /* used just in bddc debug mode */ 2122a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 2123a3df083aSStefano Zampini { 2124a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2125a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2126a3df083aSStefano Zampini Mat An; 2127a3df083aSStefano Zampini PetscErrorCode ierr; 2128a3df083aSStefano Zampini 2129a3df083aSStefano Zampini PetscFunctionBegin; 2130a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 2131a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 2132a3df083aSStefano Zampini if (is1) { 21337dae84e0SHong Zhang ierr = MatCreateSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 2134a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 2135a3df083aSStefano Zampini } else { 2136a3df083aSStefano Zampini *B = An; 2137a3df083aSStefano Zampini } 2138a3df083aSStefano Zampini PetscFunctionReturn(0); 2139a3df083aSStefano Zampini } 2140a3df083aSStefano Zampini 21411cf9b237SStefano Zampini /* TODO: add reuse flag */ 21421cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 21431cf9b237SStefano Zampini { 21441cf9b237SStefano Zampini Mat Bt; 21451cf9b237SStefano Zampini PetscScalar *a,*bdata; 21461cf9b237SStefano Zampini const PetscInt *ii,*ij; 21471cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 21481cf9b237SStefano Zampini PetscBool flg_row; 21491cf9b237SStefano Zampini PetscErrorCode ierr; 21501cf9b237SStefano Zampini 21511cf9b237SStefano Zampini PetscFunctionBegin; 21521cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 21531cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 21541cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 21551cf9b237SStefano Zampini nnz = n; 21561cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 21571cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 21581cf9b237SStefano Zampini } 21591cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 21601cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 21611cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 21621cf9b237SStefano Zampini nnz = 0; 21631cf9b237SStefano Zampini bii[0] = 0; 21641cf9b237SStefano Zampini for (i=0;i<n;i++) { 21651cf9b237SStefano Zampini PetscInt j; 21661cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 21671cf9b237SStefano Zampini PetscScalar entry = a[j]; 21683272d46bSStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || (n == m && ij[j] == i)) { 21691cf9b237SStefano Zampini bij[nnz] = ij[j]; 21701cf9b237SStefano Zampini bdata[nnz] = entry; 21711cf9b237SStefano Zampini nnz++; 21721cf9b237SStefano Zampini } 21731cf9b237SStefano Zampini } 21741cf9b237SStefano Zampini bii[i+1] = nnz; 21751cf9b237SStefano Zampini } 21761cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 21771cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 21781cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 21791cf9b237SStefano Zampini { 21801cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 21811cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 21821cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 21831cf9b237SStefano Zampini } 21843272d46bSStefano Zampini if (*B == A) { 21853272d46bSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 21863272d46bSStefano Zampini } 21871cf9b237SStefano Zampini *B = Bt; 21881cf9b237SStefano Zampini PetscFunctionReturn(0); 21891cf9b237SStefano Zampini } 21901cf9b237SStefano Zampini 21918361f951SStefano Zampini PetscErrorCode PCBDDCDetectDisconnectedComponents(PC pc, PetscBool filter, PetscInt *ncc, IS* cc[], IS* primalv) 21924f1b2e48SStefano Zampini { 2193c80a6c00SStefano Zampini Mat B = NULL; 2194c80a6c00SStefano Zampini DM dm; 21954f1b2e48SStefano Zampini IS is_dummy,*cc_n; 21964f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 21974f1b2e48SStefano Zampini PCBDDCGraph graph; 2198c80a6c00SStefano Zampini PetscInt *xadj_filtered = NULL,*adjncy_filtered = NULL; 21994f1b2e48SStefano Zampini PetscInt i,n; 22004f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2201c80a6c00SStefano Zampini PetscBool isplex = PETSC_FALSE; 22024f1b2e48SStefano Zampini PetscErrorCode ierr; 22034f1b2e48SStefano Zampini 22044f1b2e48SStefano Zampini PetscFunctionBegin; 2205a2eca866SStefano Zampini if (ncc) *ncc = 0; 2206a2eca866SStefano Zampini if (cc) *cc = NULL; 2207a2eca866SStefano Zampini if (primalv) *primalv = NULL; 2208c80a6c00SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 2209c80a6c00SStefano Zampini ierr = MatGetDM(pc->pmat,&dm);CHKERRQ(ierr); 22104f819b78SStefano Zampini if (!dm) { 22114f819b78SStefano Zampini ierr = PCGetDM(pc,&dm);CHKERRQ(ierr); 2212c80a6c00SStefano Zampini } 2213c80a6c00SStefano Zampini if (dm) { 2214c80a6c00SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)dm,DMPLEX,&isplex);CHKERRQ(ierr); 2215c80a6c00SStefano Zampini } 22168361f951SStefano Zampini if (filter) isplex = PETSC_FALSE; 22178361f951SStefano Zampini 2218c80a6c00SStefano Zampini if (isplex) { /* this code has been modified from plexpartition.c */ 2219c80a6c00SStefano Zampini PetscInt p, pStart, pEnd, a, adjSize, idx, size, nroots; 2220c80a6c00SStefano Zampini PetscInt *adj = NULL; 2221c80a6c00SStefano Zampini IS cellNumbering; 2222c80a6c00SStefano Zampini const PetscInt *cellNum; 2223c80a6c00SStefano Zampini PetscBool useCone, useClosure; 2224c80a6c00SStefano Zampini PetscSection section; 2225c80a6c00SStefano Zampini PetscSegBuffer adjBuffer; 2226c80a6c00SStefano Zampini PetscSF sfPoint; 2227c80a6c00SStefano Zampini PetscErrorCode ierr; 2228c80a6c00SStefano Zampini 2229c80a6c00SStefano Zampini ierr = DMPlexGetHeightStratum(dm, 0, &pStart, &pEnd);CHKERRQ(ierr); 2230c80a6c00SStefano Zampini ierr = DMGetPointSF(dm, &sfPoint);CHKERRQ(ierr); 2231c80a6c00SStefano Zampini ierr = PetscSFGetGraph(sfPoint, &nroots, NULL, NULL, NULL);CHKERRQ(ierr); 2232c80a6c00SStefano Zampini /* Build adjacency graph via a section/segbuffer */ 2233c80a6c00SStefano Zampini ierr = PetscSectionCreate(PetscObjectComm((PetscObject) dm), §ion);CHKERRQ(ierr); 2234c80a6c00SStefano Zampini ierr = PetscSectionSetChart(section, pStart, pEnd);CHKERRQ(ierr); 2235c80a6c00SStefano Zampini ierr = PetscSegBufferCreate(sizeof(PetscInt),1000,&adjBuffer);CHKERRQ(ierr); 2236c80a6c00SStefano Zampini /* Always use FVM adjacency to create partitioner graph */ 2237b0441da4SMatthew G. Knepley ierr = DMGetBasicAdjacency(dm, &useCone, &useClosure);CHKERRQ(ierr); 2238b0441da4SMatthew G. Knepley ierr = DMSetBasicAdjacency(dm, PETSC_TRUE, PETSC_FALSE);CHKERRQ(ierr); 2239956e2312SStefano Zampini ierr = DMPlexGetCellNumbering(dm, &cellNumbering);CHKERRQ(ierr); 2240c80a6c00SStefano Zampini ierr = ISGetIndices(cellNumbering, &cellNum);CHKERRQ(ierr); 2241c80a6c00SStefano Zampini for (n = 0, p = pStart; p < pEnd; p++) { 2242c80a6c00SStefano Zampini /* Skip non-owned cells in parallel (ParMetis expects no overlap) */ 2243c80a6c00SStefano Zampini if (nroots > 0) {if (cellNum[p] < 0) continue;} 2244c80a6c00SStefano Zampini adjSize = PETSC_DETERMINE; 2245c80a6c00SStefano Zampini ierr = DMPlexGetAdjacency(dm, p, &adjSize, &adj);CHKERRQ(ierr); 2246c80a6c00SStefano Zampini for (a = 0; a < adjSize; ++a) { 2247c80a6c00SStefano Zampini const PetscInt point = adj[a]; 22485cef3d0dSStefano Zampini if (pStart <= point && point < pEnd) { 2249c80a6c00SStefano Zampini PetscInt *PETSC_RESTRICT pBuf; 2250c80a6c00SStefano Zampini ierr = PetscSectionAddDof(section, p, 1);CHKERRQ(ierr); 2251c80a6c00SStefano Zampini ierr = PetscSegBufferGetInts(adjBuffer, 1, &pBuf);CHKERRQ(ierr); 2252c80a6c00SStefano Zampini *pBuf = point; 2253c80a6c00SStefano Zampini } 2254c80a6c00SStefano Zampini } 2255c80a6c00SStefano Zampini n++; 2256c80a6c00SStefano Zampini } 2257b0441da4SMatthew G. Knepley ierr = DMSetBasicAdjacency(dm, useCone, useClosure);CHKERRQ(ierr); 2258c80a6c00SStefano Zampini /* Derive CSR graph from section/segbuffer */ 2259c80a6c00SStefano Zampini ierr = PetscSectionSetUp(section);CHKERRQ(ierr); 2260c80a6c00SStefano Zampini ierr = PetscSectionGetStorageSize(section, &size);CHKERRQ(ierr); 2261c80a6c00SStefano Zampini ierr = PetscMalloc1(n+1, &xadj);CHKERRQ(ierr); 2262c80a6c00SStefano Zampini for (idx = 0, p = pStart; p < pEnd; p++) { 2263c80a6c00SStefano Zampini if (nroots > 0) {if (cellNum[p] < 0) continue;} 2264c80a6c00SStefano Zampini ierr = PetscSectionGetOffset(section, p, &(xadj[idx++]));CHKERRQ(ierr); 2265c80a6c00SStefano Zampini } 2266c80a6c00SStefano Zampini xadj[n] = size; 2267c80a6c00SStefano Zampini ierr = PetscSegBufferExtractAlloc(adjBuffer, &adjncy);CHKERRQ(ierr); 2268c80a6c00SStefano Zampini /* Clean up */ 2269c80a6c00SStefano Zampini ierr = PetscSegBufferDestroy(&adjBuffer);CHKERRQ(ierr); 2270c80a6c00SStefano Zampini ierr = PetscSectionDestroy(§ion);CHKERRQ(ierr); 2271c80a6c00SStefano Zampini ierr = PetscFree(adj);CHKERRQ(ierr); 2272c80a6c00SStefano Zampini graph->xadj = xadj; 2273c80a6c00SStefano Zampini graph->adjncy = adjncy; 2274c80a6c00SStefano Zampini } else { 2275c80a6c00SStefano Zampini Mat A; 22768361f951SStefano Zampini PetscBool isseqaij, flg_row; 2277c80a6c00SStefano Zampini 2278c80a6c00SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 227963c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 2280a2eca866SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 228163c961adSStefano Zampini PetscFunctionReturn(0); 228263c961adSStefano Zampini } 22837ebab0bbSStefano Zampini ierr = PetscObjectBaseTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 22844f1b2e48SStefano Zampini if (!isseqaij && filter) { 22851cf9b237SStefano Zampini PetscBool isseqdense; 22861cf9b237SStefano Zampini 22871cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 22881cf9b237SStefano Zampini if (!isseqdense) { 22894f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 22901cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 22911cf9b237SStefano Zampini PetscScalar *array; 22921cf9b237SStefano Zampini PetscReal chop=1.e-6; 22931cf9b237SStefano Zampini 22941cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 22951cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 22961cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 22971cf9b237SStefano Zampini for (i=0;i<n;i++) { 22981cf9b237SStefano Zampini PetscInt j; 22991cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 23001cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 23011cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 23021cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 23031cf9b237SStefano Zampini } 23041cf9b237SStefano Zampini } 23051cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 23069d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 23071cf9b237SStefano Zampini } 23084f1b2e48SStefano Zampini } else { 2309c80a6c00SStefano Zampini ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 23104f1b2e48SStefano Zampini B = A; 23114f1b2e48SStefano Zampini } 23124f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 23134f1b2e48SStefano Zampini 23144f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 23154f1b2e48SStefano Zampini if (filter) { 23164f1b2e48SStefano Zampini PetscScalar *data; 23174f1b2e48SStefano Zampini PetscInt j,cum; 23184f1b2e48SStefano Zampini 23194f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 23204f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 23214f1b2e48SStefano Zampini cum = 0; 23224f1b2e48SStefano Zampini for (i=0;i<n;i++) { 23234f1b2e48SStefano Zampini PetscInt t; 23244f1b2e48SStefano Zampini 23254f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 23264f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 23274f1b2e48SStefano Zampini continue; 23284f1b2e48SStefano Zampini } 23294f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 23304f1b2e48SStefano Zampini } 23314f1b2e48SStefano Zampini t = xadj_filtered[i]; 23324f1b2e48SStefano Zampini xadj_filtered[i] = cum; 23334f1b2e48SStefano Zampini cum += t; 23344f1b2e48SStefano Zampini } 23354f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 23364f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 23374f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 23384f1b2e48SStefano Zampini } else { 23394f1b2e48SStefano Zampini graph->xadj = xadj; 23404f1b2e48SStefano Zampini graph->adjncy = adjncy; 23414f1b2e48SStefano Zampini } 2342c80a6c00SStefano Zampini } 2343c80a6c00SStefano Zampini /* compute local connected components using PCBDDCGraph */ 2344c80a6c00SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 2345c80a6c00SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 2346c80a6c00SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2347c80a6c00SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 2348c80a6c00SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 23494f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 23504f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 2351c80a6c00SStefano Zampini 23524f1b2e48SStefano Zampini /* partial clean up */ 23534f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 2354c80a6c00SStefano Zampini if (B) { 2355c80a6c00SStefano Zampini PetscBool flg_row; 23564f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 23574f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 23584f1b2e48SStefano Zampini } 2359c80a6c00SStefano Zampini if (isplex) { 2360c80a6c00SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 2361c80a6c00SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 2362c80a6c00SStefano Zampini } 23634f1b2e48SStefano Zampini 23644f1b2e48SStefano Zampini /* get back data */ 2365c80a6c00SStefano Zampini if (isplex) { 2366c80a6c00SStefano Zampini if (ncc) *ncc = graph->ncc; 2367c80a6c00SStefano Zampini if (cc || primalv) { 2368c80a6c00SStefano Zampini Mat A; 2369c80a6c00SStefano Zampini PetscBT btv,btvt; 2370c80a6c00SStefano Zampini PetscSection subSection; 2371c80a6c00SStefano Zampini PetscInt *ids,cum,cump,*cids,*pids; 2372c80a6c00SStefano Zampini 2373c80a6c00SStefano Zampini ierr = DMPlexGetSubdomainSection(dm,&subSection);CHKERRQ(ierr); 2374c80a6c00SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 2375c80a6c00SStefano Zampini ierr = PetscMalloc3(A->rmap->n,&ids,graph->ncc+1,&cids,A->rmap->n,&pids);CHKERRQ(ierr); 2376c80a6c00SStefano Zampini ierr = PetscBTCreate(A->rmap->n,&btv);CHKERRQ(ierr); 2377c80a6c00SStefano Zampini ierr = PetscBTCreate(A->rmap->n,&btvt);CHKERRQ(ierr); 2378c80a6c00SStefano Zampini 2379c80a6c00SStefano Zampini cids[0] = 0; 2380c80a6c00SStefano Zampini for (i = 0, cump = 0, cum = 0; i < graph->ncc; i++) { 2381c80a6c00SStefano Zampini PetscInt j; 2382c80a6c00SStefano Zampini 2383c80a6c00SStefano Zampini ierr = PetscBTMemzero(A->rmap->n,btvt);CHKERRQ(ierr); 2384c80a6c00SStefano Zampini for (j = graph->cptr[i]; j < graph->cptr[i+1]; j++) { 2385c80a6c00SStefano Zampini PetscInt k, size, *closure = NULL, cell = graph->queue[j]; 2386c80a6c00SStefano Zampini 2387c80a6c00SStefano Zampini ierr = DMPlexGetTransitiveClosure(dm,cell,PETSC_TRUE,&size,&closure);CHKERRQ(ierr); 2388c80a6c00SStefano Zampini for (k = 0; k < 2*size; k += 2) { 238920c3699dSStefano Zampini PetscInt s, pp, p = closure[k], off, dof, cdof; 2390c80a6c00SStefano Zampini 2391c80a6c00SStefano Zampini ierr = PetscSectionGetConstraintDof(subSection,p,&cdof);CHKERRQ(ierr); 2392c80a6c00SStefano Zampini ierr = PetscSectionGetOffset(subSection,p,&off);CHKERRQ(ierr); 2393c80a6c00SStefano Zampini ierr = PetscSectionGetDof(subSection,p,&dof);CHKERRQ(ierr); 2394c80a6c00SStefano Zampini for (s = 0; s < dof-cdof; s++) { 2395c80a6c00SStefano Zampini if (PetscBTLookupSet(btvt,off+s)) continue; 2396c80a6c00SStefano Zampini if (!PetscBTLookup(btv,off+s)) { 2397c80a6c00SStefano Zampini ids[cum++] = off+s; 2398c80a6c00SStefano Zampini } else { /* cross-vertex */ 2399c80a6c00SStefano Zampini pids[cump++] = off+s; 2400c80a6c00SStefano Zampini } 2401c80a6c00SStefano Zampini } 240220c3699dSStefano Zampini ierr = DMPlexGetTreeParent(dm,p,&pp,NULL);CHKERRQ(ierr); 240320c3699dSStefano Zampini if (pp != p) { 240420c3699dSStefano Zampini ierr = PetscSectionGetConstraintDof(subSection,pp,&cdof);CHKERRQ(ierr); 240520c3699dSStefano Zampini ierr = PetscSectionGetOffset(subSection,pp,&off);CHKERRQ(ierr); 240620c3699dSStefano Zampini ierr = PetscSectionGetDof(subSection,pp,&dof);CHKERRQ(ierr); 240720c3699dSStefano Zampini for (s = 0; s < dof-cdof; s++) { 240820c3699dSStefano Zampini if (PetscBTLookupSet(btvt,off+s)) continue; 240920c3699dSStefano Zampini if (!PetscBTLookup(btv,off+s)) { 241020c3699dSStefano Zampini ids[cum++] = off+s; 241120c3699dSStefano Zampini } else { /* cross-vertex */ 241220c3699dSStefano Zampini pids[cump++] = off+s; 241320c3699dSStefano Zampini } 241420c3699dSStefano Zampini } 241520c3699dSStefano Zampini } 2416c80a6c00SStefano Zampini } 2417c80a6c00SStefano Zampini ierr = DMPlexRestoreTransitiveClosure(dm,cell,PETSC_TRUE,&size,&closure);CHKERRQ(ierr); 2418c80a6c00SStefano Zampini } 2419c80a6c00SStefano Zampini cids[i+1] = cum; 2420c80a6c00SStefano Zampini /* mark dofs as already assigned */ 2421c80a6c00SStefano Zampini for (j = cids[i]; j < cids[i+1]; j++) { 2422c80a6c00SStefano Zampini ierr = PetscBTSet(btv,ids[j]);CHKERRQ(ierr); 2423c80a6c00SStefano Zampini } 2424c80a6c00SStefano Zampini } 2425c80a6c00SStefano Zampini if (cc) { 2426c80a6c00SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 2427c80a6c00SStefano Zampini for (i = 0; i < graph->ncc; i++) { 2428c80a6c00SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cids[i+1]-cids[i],ids+cids[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 2429c80a6c00SStefano Zampini } 2430c80a6c00SStefano Zampini *cc = cc_n; 2431c80a6c00SStefano Zampini } 2432c80a6c00SStefano Zampini if (primalv) { 2433c80a6c00SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),cump,pids,PETSC_COPY_VALUES,primalv);CHKERRQ(ierr); 2434c80a6c00SStefano Zampini } 2435c80a6c00SStefano Zampini ierr = PetscFree3(ids,cids,pids);CHKERRQ(ierr); 2436c80a6c00SStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 2437c80a6c00SStefano Zampini ierr = PetscBTDestroy(&btvt);CHKERRQ(ierr); 2438c80a6c00SStefano Zampini } 2439c80a6c00SStefano Zampini } else { 24401cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 24411cf9b237SStefano Zampini if (cc) { 24424f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 24434f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 24444f1b2e48SStefano 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); 24454f1b2e48SStefano Zampini } 24464f1b2e48SStefano Zampini *cc = cc_n; 24471cf9b237SStefano Zampini } 2448c80a6c00SStefano Zampini } 24494f1b2e48SStefano Zampini /* clean up graph */ 24500a545947SLisandro Dalcin graph->xadj = NULL; 24510a545947SLisandro Dalcin graph->adjncy = NULL; 24524f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 24534f1b2e48SStefano Zampini PetscFunctionReturn(0); 24544f1b2e48SStefano Zampini } 24554f1b2e48SStefano Zampini 24565408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 24575408967cSStefano Zampini { 24585408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 24595408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2460dee84bffSStefano Zampini IS dirIS = NULL; 24614f1b2e48SStefano Zampini PetscInt i; 24625408967cSStefano Zampini PetscErrorCode ierr; 24635408967cSStefano Zampini 24645408967cSStefano Zampini PetscFunctionBegin; 2465dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 24665408967cSStefano Zampini if (zerodiag) { 24675408967cSStefano Zampini Mat A; 24685408967cSStefano Zampini Vec vec3_N; 24695408967cSStefano Zampini PetscScalar *vals; 24705408967cSStefano Zampini const PetscInt *idxs; 2471d12d3064SStefano Zampini PetscInt nz,*count; 24725408967cSStefano Zampini 24735408967cSStefano Zampini /* p0 */ 24745408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 24755408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 24765408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 24775408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 24784f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 24795408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24805408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 24815408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 24825408967cSStefano Zampini /* v_I */ 24835408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 24845408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 24855408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24865408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 24875408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 24885408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 24895408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24905408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 24915408967cSStefano Zampini if (dirIS) { 24925408967cSStefano Zampini PetscInt n; 24935408967cSStefano Zampini 24945408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 24955408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 24965408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 24975408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24985408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 24995408967cSStefano Zampini } 25005408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 25015408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 25025408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 25035408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 2504669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 25055408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 25065408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 25079a962809SStefano 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])); 25085408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 25095408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 2510d12d3064SStefano Zampini 2511d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 2512d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 2513d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2514d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 2515d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2516d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 25176080607fSStefano 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]); 2518d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 2519d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 25205408967cSStefano Zampini } 2521dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 25225408967cSStefano Zampini 25235408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 25245408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 25254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 25265408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 25274f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 25285408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 2529f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2530f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 25316080607fSStefano Zampini if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %D instead of %g",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i); 2532f2a566d8SStefano Zampini } 25335408967cSStefano Zampini PetscFunctionReturn(0); 25345408967cSStefano Zampini } 25355408967cSStefano Zampini 25363b03f7bbSStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, PetscBool reuse, IS *zerodiaglocal) 2537339f8db1SStefano Zampini { 2538339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 25393b03f7bbSStefano Zampini IS pressures = NULL,zerodiag = NULL,*bzerodiag = NULL,zerodiag_save,*zerodiag_subs; 25403b03f7bbSStefano Zampini PetscInt nz,n,benign_n,bsp = 1; 25414edc6404Sstefano_zampini PetscInt *interior_dofs,n_interior_dofs,nneu; 25424edc6404Sstefano_zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag,checkb; 2543339f8db1SStefano Zampini PetscErrorCode ierr; 2544339f8db1SStefano Zampini 2545339f8db1SStefano Zampini PetscFunctionBegin; 25463b03f7bbSStefano Zampini if (reuse) goto project_b0; 25479f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 25489f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 2549a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 2550a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 2551a3df083aSStefano Zampini } 2552a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 25533b03f7bbSStefano Zampini has_null_pressures = PETSC_TRUE; 25543b03f7bbSStefano Zampini have_null = PETSC_TRUE; 25553b03f7bbSStefano Zampini /* if a local information on dofs is present, gets pressure dofs from command line (uses the last field is not provided) 25563b03f7bbSStefano Zampini Without local information, it uses only the zerodiagonal dofs (ok if the pressure block is all zero and it is a scalar field) 25574f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 25584f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 25591ae86dd6SStefano Zampini since the local Schur complements are already SPD 25604f1b2e48SStefano Zampini */ 256140fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 25627fbe2174Sstefano_zampini IS iP = NULL; 25633b03f7bbSStefano Zampini PetscInt p,*pp; 25643b03f7bbSStefano Zampini PetscBool flg; 25654f1b2e48SStefano Zampini 25663b03f7bbSStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pp);CHKERRQ(ierr); 25673b03f7bbSStefano Zampini n = pcbddc->n_ISForDofsLocal; 256828b8efb1Sstefano_zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC benign options","PC");CHKERRQ(ierr); 25693b03f7bbSStefano Zampini ierr = PetscOptionsIntArray("-pc_bddc_pressure_field","Field id for pressures",NULL,pp,&n,&flg);CHKERRQ(ierr); 257028b8efb1Sstefano_zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 25713b03f7bbSStefano Zampini if (!flg) { 25723b03f7bbSStefano Zampini n = 1; 25733b03f7bbSStefano Zampini pp[0] = pcbddc->n_ISForDofsLocal-1; 25743b03f7bbSStefano Zampini } 25753b03f7bbSStefano Zampini 25763b03f7bbSStefano Zampini bsp = 0; 25773b03f7bbSStefano Zampini for (p=0;p<n;p++) { 25783b03f7bbSStefano Zampini PetscInt bs; 25793b03f7bbSStefano Zampini 25803b03f7bbSStefano Zampini if (pp[p] < 0 || pp[p] > pcbddc->n_ISForDofsLocal-1) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Invalid field id for pressures %D",pp[p]); 25813b03f7bbSStefano Zampini ierr = ISGetBlockSize(pcbddc->ISForDofsLocal[pp[p]],&bs);CHKERRQ(ierr); 25823b03f7bbSStefano Zampini bsp += bs; 25833b03f7bbSStefano Zampini } 25843b03f7bbSStefano Zampini ierr = PetscMalloc1(bsp,&bzerodiag);CHKERRQ(ierr); 25853b03f7bbSStefano Zampini bsp = 0; 25863b03f7bbSStefano Zampini for (p=0;p<n;p++) { 25873b03f7bbSStefano Zampini const PetscInt *idxs; 25883b03f7bbSStefano Zampini PetscInt b,bs,npl,*bidxs; 25893b03f7bbSStefano Zampini 25903b03f7bbSStefano Zampini ierr = ISGetBlockSize(pcbddc->ISForDofsLocal[pp[p]],&bs);CHKERRQ(ierr); 25913b03f7bbSStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[pp[p]],&npl);CHKERRQ(ierr); 25923b03f7bbSStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[pp[p]],&idxs);CHKERRQ(ierr); 25933b03f7bbSStefano Zampini ierr = PetscMalloc1(npl/bs,&bidxs);CHKERRQ(ierr); 25943b03f7bbSStefano Zampini for (b=0;b<bs;b++) { 25953b03f7bbSStefano Zampini PetscInt i; 25963b03f7bbSStefano Zampini 25973b03f7bbSStefano Zampini for (i=0;i<npl/bs;i++) bidxs[i] = idxs[bs*i+b]; 25983b03f7bbSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl/bs,bidxs,PETSC_COPY_VALUES,&bzerodiag[bsp]);CHKERRQ(ierr); 25993b03f7bbSStefano Zampini bsp++; 26003b03f7bbSStefano Zampini } 26013b03f7bbSStefano Zampini ierr = PetscFree(bidxs);CHKERRQ(ierr); 26023b03f7bbSStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[pp[p]],&idxs);CHKERRQ(ierr); 26033b03f7bbSStefano Zampini } 26043b03f7bbSStefano Zampini ierr = ISConcatenate(PETSC_COMM_SELF,bsp,bzerodiag,&pressures);CHKERRQ(ierr); 26053b03f7bbSStefano Zampini 26067fbe2174Sstefano_zampini /* remove zeroed out pressures if we are setting up a BDDC solver for a saddle-point FETI-DP */ 26077fbe2174Sstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lP",(PetscObject*)&iP);CHKERRQ(ierr); 26087fbe2174Sstefano_zampini if (iP) { 26097fbe2174Sstefano_zampini IS newpressures; 26107fbe2174Sstefano_zampini 26117fbe2174Sstefano_zampini ierr = ISDifference(pressures,iP,&newpressures);CHKERRQ(ierr); 26127fbe2174Sstefano_zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 26137fbe2174Sstefano_zampini pressures = newpressures; 26147fbe2174Sstefano_zampini } 261540fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 261640fa8d13SStefano Zampini if (!sorted) { 261740fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 261840fa8d13SStefano Zampini } 26193b03f7bbSStefano Zampini ierr = PetscFree(pp);CHKERRQ(ierr); 262040fa8d13SStefano Zampini } 26213b03f7bbSStefano Zampini 262297d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 262397d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 262427b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 262597d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 2626339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 2627339f8db1SStefano Zampini if (!sorted) { 2628339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 2629339f8db1SStefano Zampini } 26304edc6404Sstefano_zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 26314edc6404Sstefano_zampini zerodiag_save = zerodiag; 2632339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 26334f1b2e48SStefano Zampini if (!nz) { 26344f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 26354f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 263640fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 263740fa8d13SStefano Zampini } 26384f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 26393b03f7bbSStefano Zampini 26404f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 26414f1b2e48SStefano Zampini zerodiag_subs = NULL; 26423b03f7bbSStefano Zampini benign_n = 0; 26431f4df5f7SStefano Zampini n_interior_dofs = 0; 26441f4df5f7SStefano Zampini interior_dofs = NULL; 26454edc6404Sstefano_zampini nneu = 0; 26464edc6404Sstefano_zampini if (pcbddc->NeumannBoundariesLocal) { 26474edc6404Sstefano_zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nneu);CHKERRQ(ierr); 26484edc6404Sstefano_zampini } 26493369cb78Sstefano_zampini checkb = (PetscBool)(!pcbddc->NeumannBoundariesLocal || pcbddc->current_level); 26504edc6404Sstefano_zampini if (checkb) { /* need to compute interior nodes */ 26511f4df5f7SStefano Zampini PetscInt n,i,j; 26521f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 26531f4df5f7SStefano Zampini PetscInt *iwork; 26541f4df5f7SStefano Zampini 26551f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 26561f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 26571f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 26581f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 265990648384SStefano Zampini for (i=1;i<n_neigh;i++) 26601f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 26611f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 26621f4df5f7SStefano Zampini for (i=0;i<n;i++) 26631f4df5f7SStefano Zampini if (!iwork[i]) 26641f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 26651f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 26661f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 26671f4df5f7SStefano Zampini } 26684f1b2e48SStefano Zampini if (has_null_pressures) { 26694f1b2e48SStefano Zampini IS *subs; 26704edc6404Sstefano_zampini PetscInt nsubs,i,j,nl; 26711f4df5f7SStefano Zampini const PetscInt *idxs; 26721f4df5f7SStefano Zampini PetscScalar *array; 26731f4df5f7SStefano Zampini Vec *work; 26741f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 26754f1b2e48SStefano Zampini 26764f1b2e48SStefano Zampini subs = pcbddc->local_subs; 26774f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 26781f4df5f7SStefano 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) */ 26794edc6404Sstefano_zampini if (checkb) { 26801f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 26811f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 26821f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 26831f4df5f7SStefano Zampini /* work[0] = 1_p */ 26841f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 26851f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 26861f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 26871f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 26881f4df5f7SStefano Zampini /* work[0] = 1_v */ 26891f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 26901f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 26911f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 26921f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 26931f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 26941f4df5f7SStefano Zampini } 26953b03f7bbSStefano Zampini 26963b03f7bbSStefano Zampini if (nsubs > 1 || bsp > 1) { 26973b03f7bbSStefano Zampini IS *is; 26983b03f7bbSStefano Zampini PetscInt b,totb; 26993b03f7bbSStefano Zampini 27003b03f7bbSStefano Zampini totb = bsp; 27013b03f7bbSStefano Zampini is = bsp > 1 ? bzerodiag : &zerodiag; 27023b03f7bbSStefano Zampini nsubs = PetscMax(nsubs,1); 27033b03f7bbSStefano Zampini ierr = PetscCalloc1(nsubs*totb,&zerodiag_subs);CHKERRQ(ierr); 27043b03f7bbSStefano Zampini for (b=0;b<totb;b++) { 27054f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 27064f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 27074f1b2e48SStefano Zampini IS t_zerodiag_subs; 27084f1b2e48SStefano Zampini PetscInt nl; 27094f1b2e48SStefano Zampini 27103b03f7bbSStefano Zampini if (subs) { 27114f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 27123b03f7bbSStefano Zampini } else { 27133b03f7bbSStefano Zampini IS tis; 27143b03f7bbSStefano Zampini 27153b03f7bbSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&nl,NULL);CHKERRQ(ierr); 27163b03f7bbSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,nl,0,1,&tis);CHKERRQ(ierr); 27173b03f7bbSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(tis,&l2g);CHKERRQ(ierr); 27183b03f7bbSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27193b03f7bbSStefano Zampini } 27203b03f7bbSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,is[b],&t_zerodiag_subs);CHKERRQ(ierr); 27214f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 27224f1b2e48SStefano Zampini if (nl) { 27234f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 27244f1b2e48SStefano Zampini 27254edc6404Sstefano_zampini if (checkb) { 27261f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 27271f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 27281f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 27291f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 27301f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 27311f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 27321f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 27331f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 27341f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 27351f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 27361f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 27371f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 27381f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 27391f4df5f7SStefano Zampini valid = PETSC_FALSE; 27401f4df5f7SStefano Zampini break; 27411f4df5f7SStefano Zampini } 27421f4df5f7SStefano Zampini } 27431f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 27441f4df5f7SStefano Zampini } 27456632bad2Sstefano_zampini if (valid && nneu) { 27466632bad2Sstefano_zampini const PetscInt *idxs; 27471f4df5f7SStefano Zampini PetscInt nzb; 27481f4df5f7SStefano Zampini 27496632bad2Sstefano_zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 27506632bad2Sstefano_zampini ierr = ISGlobalToLocalMappingApply(l2g,IS_GTOLM_DROP,nneu,idxs,&nzb,NULL);CHKERRQ(ierr); 27516632bad2Sstefano_zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 27521f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 27531f4df5f7SStefano Zampini } 27541f4df5f7SStefano Zampini if (valid && pressures) { 27553b03f7bbSStefano Zampini IS t_pressure_subs,tmp; 27563b03f7bbSStefano Zampini PetscInt i1,i2; 27573b03f7bbSStefano Zampini 27584f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 27593b03f7bbSStefano Zampini ierr = ISEmbed(t_zerodiag_subs,t_pressure_subs,PETSC_TRUE,&tmp);CHKERRQ(ierr); 27603b03f7bbSStefano Zampini ierr = ISGetLocalSize(tmp,&i1);CHKERRQ(ierr); 27613b03f7bbSStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&i2);CHKERRQ(ierr); 27623b03f7bbSStefano Zampini if (i2 != i1) valid = PETSC_FALSE; 27634f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 27643b03f7bbSStefano Zampini ierr = ISDestroy(&tmp);CHKERRQ(ierr); 27654f1b2e48SStefano Zampini } 27664f1b2e48SStefano Zampini if (valid) { 27673b03f7bbSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[benign_n]);CHKERRQ(ierr); 27683b03f7bbSStefano Zampini benign_n++; 27693b03f7bbSStefano Zampini } else recompute_zerodiag = PETSC_TRUE; 27704f1b2e48SStefano Zampini } 27714f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 27724f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 27734f1b2e48SStefano Zampini } 27743b03f7bbSStefano Zampini } 27754f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 27764f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 27771f4df5f7SStefano Zampini 27786632bad2Sstefano_zampini if (nneu) valid = PETSC_FALSE; 27791f4df5f7SStefano Zampini if (valid && pressures) { 27804f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 27814f1b2e48SStefano Zampini } 27824edc6404Sstefano_zampini if (valid && checkb) { 27831f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 27841f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 27851f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 27861f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 27871f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 27881f4df5f7SStefano Zampini valid = PETSC_FALSE; 27891f4df5f7SStefano Zampini break; 27901f4df5f7SStefano Zampini } 27911f4df5f7SStefano Zampini } 27921f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 27931f4df5f7SStefano Zampini } 27944f1b2e48SStefano Zampini if (valid) { 27953b03f7bbSStefano Zampini benign_n = 1; 27963b03f7bbSStefano Zampini ierr = PetscMalloc1(benign_n,&zerodiag_subs);CHKERRQ(ierr); 27974f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 27984f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 27994f1b2e48SStefano Zampini } 28004f1b2e48SStefano Zampini } 28014edc6404Sstefano_zampini if (checkb) { 28021f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 28034f1b2e48SStefano Zampini } 28041f4df5f7SStefano Zampini } 28051f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 28064f1b2e48SStefano Zampini 28073b03f7bbSStefano Zampini if (!benign_n) { 2808b9b0e38cSStefano Zampini PetscInt n; 2809b9b0e38cSStefano Zampini 28104f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 28114f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2812b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 281376a58201SStefano Zampini if (n) have_null = PETSC_FALSE; 2814b9b0e38cSStefano Zampini } 28154f1b2e48SStefano Zampini 28164f1b2e48SStefano Zampini /* final check for null pressures */ 28174f1b2e48SStefano Zampini if (zerodiag && pressures) { 28183b03f7bbSStefano Zampini ierr = ISEqual(pressures,zerodiag,&have_null);CHKERRQ(ierr); 28194f1b2e48SStefano Zampini } 28204f1b2e48SStefano Zampini 28214f1b2e48SStefano Zampini if (recompute_zerodiag) { 28224f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 28233b03f7bbSStefano Zampini if (benign_n == 1) { 28244f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 28254f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 28264f1b2e48SStefano Zampini } else { 28274f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 28284f1b2e48SStefano Zampini 28294f1b2e48SStefano Zampini nzn = 0; 28303b03f7bbSStefano Zampini for (i=0;i<benign_n;i++) { 28314f1b2e48SStefano Zampini PetscInt ns; 28324f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 28334f1b2e48SStefano Zampini nzn += ns; 28344f1b2e48SStefano Zampini } 28354f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 28364f1b2e48SStefano Zampini nzn = 0; 28373b03f7bbSStefano Zampini for (i=0;i<benign_n;i++) { 28384f1b2e48SStefano Zampini PetscInt ns,*idxs; 28394f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 28404f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2841580bdb30SBarry Smith ierr = PetscArraycpy(new_idxs+nzn,idxs,ns);CHKERRQ(ierr); 28424f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 28434f1b2e48SStefano Zampini nzn += ns; 28444f1b2e48SStefano Zampini } 28454f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 28464f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 28474f1b2e48SStefano Zampini } 28484f1b2e48SStefano Zampini have_null = PETSC_FALSE; 28494f1b2e48SStefano Zampini } 28504f1b2e48SStefano Zampini 28513b03f7bbSStefano Zampini /* determines if the coarse solver will be singular or not */ 2852820f2d46SBarry Smith ierr = MPIU_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 28533b03f7bbSStefano Zampini 2854669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2855a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2856a198735bSStefano Zampini Mat A,loc_divudotp; 2857a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2858a198735bSStefano Zampini IS row,col,isused = NULL; 2859a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2860a198735bSStefano Zampini 28611f4df5f7SStefano Zampini if (pressures) { 28621f4df5f7SStefano Zampini isused = pressures; 28631f4df5f7SStefano Zampini } else { 28644edc6404Sstefano_zampini isused = zerodiag_save; 28651f4df5f7SStefano Zampini } 2866a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2867669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 28681ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 28691ae86dd6SStefano 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"); 2870a198735bSStefano Zampini n_isused = 0; 2871a198735bSStefano Zampini if (isused) { 2872a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2873a198735bSStefano Zampini } 2874ffc4695bSBarry Smith ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 2875a198735bSStefano Zampini st = st-n_isused; 28761ae86dd6SStefano Zampini if (n) { 2877a198735bSStefano Zampini const PetscInt *gidxs; 2878a198735bSStefano Zampini 28797dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2880a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2881a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2882a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2883a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2884a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 28851ae86dd6SStefano Zampini } else { 2886a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2887a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2888a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2889a198735bSStefano Zampini } 2890a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2891a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2892a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2893a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2894a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2895a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2896a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2897a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2898a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2899a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2900a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2901a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2902a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2903a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 29041ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 29051ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 29061ae86dd6SStefano Zampini } 29074edc6404Sstefano_zampini ierr = ISDestroy(&zerodiag_save);CHKERRQ(ierr); 29083b03f7bbSStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 29093b03f7bbSStefano Zampini if (bzerodiag) { 29103b03f7bbSStefano Zampini PetscInt i; 2911b3afcdbeSStefano Zampini 29123b03f7bbSStefano Zampini for (i=0;i<bsp;i++) { 29133b03f7bbSStefano Zampini ierr = ISDestroy(&bzerodiag[i]);CHKERRQ(ierr); 29143b03f7bbSStefano Zampini } 29153b03f7bbSStefano Zampini ierr = PetscFree(bzerodiag);CHKERRQ(ierr); 29163b03f7bbSStefano Zampini } 29173b03f7bbSStefano Zampini pcbddc->benign_n = benign_n; 29183b03f7bbSStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 29193b03f7bbSStefano Zampini 29203b03f7bbSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 29213b03f7bbSStefano Zampini have_null = (PetscBool)(!!pcbddc->benign_n); 2922820f2d46SBarry Smith ierr = MPIU_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 29233b03f7bbSStefano Zampini 29243b03f7bbSStefano Zampini project_b0: 2925aa0d93e9SStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2926b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 29273b03f7bbSStefano Zampini if (pcbddc->benign_n) { 29284f1b2e48SStefano Zampini PetscInt i,s,*nnz; 29294f1b2e48SStefano Zampini 2930339f8db1SStefano Zampini /* local change of basis for pressures */ 2931339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 293297d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2933339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2934339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2935339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 2936aa0d93e9SStefano Zampini for (i=0;i<n;i++) nnz[i] = 1; /* defaults to identity */ 29374f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2938aa0d93e9SStefano Zampini const PetscInt *idxs; 29394f1b2e48SStefano Zampini PetscInt nzs,j; 29404f1b2e48SStefano Zampini 29413b03f7bbSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nzs);CHKERRQ(ierr); 29423b03f7bbSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 29434f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 29444f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 29453b03f7bbSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 29464f1b2e48SStefano Zampini } 2947339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2948e1b21442SStefano Zampini ierr = MatSetOption(pcbddc->benign_change,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2949339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2950aa0d93e9SStefano Zampini /* set identity by default */ 2951aa0d93e9SStefano Zampini for (i=0;i<n;i++) { 2952aa0d93e9SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,i,i,1.,INSERT_VALUES);CHKERRQ(ierr); 2953339f8db1SStefano Zampini } 29549f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 29554f1b2e48SStefano 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); 2956339f8db1SStefano Zampini /* set change on pressures */ 29574f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 29584f1b2e48SStefano Zampini PetscScalar *array; 2959aa0d93e9SStefano Zampini const PetscInt *idxs; 29604f1b2e48SStefano Zampini PetscInt nzs; 29614f1b2e48SStefano Zampini 29623b03f7bbSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[s],&nzs);CHKERRQ(ierr); 29633b03f7bbSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[s],&idxs);CHKERRQ(ierr); 29644f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2965339f8db1SStefano Zampini PetscScalar vals[2]; 2966339f8db1SStefano Zampini PetscInt cols[2]; 2967339f8db1SStefano Zampini 2968339f8db1SStefano Zampini cols[0] = idxs[i]; 29694f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2970339f8db1SStefano Zampini vals[0] = 1.; 2971b0f5fe93SStefano Zampini vals[1] = 1.; 29724f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2973339f8db1SStefano Zampini } 29744f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 29754f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 29764f1b2e48SStefano Zampini array[nzs-1] = 1.; 29774f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 29784f1b2e48SStefano Zampini /* store local idxs for p0 */ 29794f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 29803b03f7bbSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2981339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 29824f1b2e48SStefano Zampini } 2983339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2984339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 29853b03f7bbSStefano Zampini 2986a3df083aSStefano Zampini /* project if needed */ 2987a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 29881dd7afcfSStefano Zampini Mat M; 29891dd7afcfSStefano Zampini 29901dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2991339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 29921dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 29931dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2994a3df083aSStefano Zampini } 29954f1b2e48SStefano Zampini /* store global idxs for p0 */ 29964f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2997339f8db1SStefano Zampini } 2998339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2999339f8db1SStefano Zampini PetscFunctionReturn(0); 3000339f8db1SStefano Zampini } 3001339f8db1SStefano Zampini 3002015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 3003efc2fbd9SStefano Zampini { 3004efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3005de9d7bd0SStefano Zampini PetscScalar *array; 3006efc2fbd9SStefano Zampini PetscErrorCode ierr; 3007efc2fbd9SStefano Zampini 3008efc2fbd9SStefano Zampini PetscFunctionBegin; 3009efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 3010efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 30114f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 3012efc2fbd9SStefano Zampini } 3013de9d7bd0SStefano Zampini if (get) { 3014efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 3015ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0,MPI_REPLACE);CHKERRQ(ierr); 3016ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0,MPI_REPLACE);CHKERRQ(ierr); 3017efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 3018de9d7bd0SStefano Zampini } else { 3019de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 302083df288dSJunchao Zhang ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPI_REPLACE);CHKERRQ(ierr); 302183df288dSJunchao Zhang ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPI_REPLACE);CHKERRQ(ierr); 3022de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 3023efc2fbd9SStefano Zampini } 3024efc2fbd9SStefano Zampini PetscFunctionReturn(0); 3025efc2fbd9SStefano Zampini } 3026efc2fbd9SStefano Zampini 3027c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 3028c263805aSStefano Zampini { 3029c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3030c263805aSStefano Zampini PetscErrorCode ierr; 3031c263805aSStefano Zampini 3032c263805aSStefano Zampini PetscFunctionBegin; 3033c263805aSStefano Zampini /* TODO: add error checking 3034c263805aSStefano Zampini - avoid nested pop (or push) calls. 3035c263805aSStefano Zampini - cannot push before pop. 30361c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 3037c263805aSStefano Zampini */ 30384f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 3039efc2fbd9SStefano Zampini PetscFunctionReturn(0); 3040efc2fbd9SStefano Zampini } 3041c263805aSStefano Zampini if (pop) { 3042a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 30434f1b2e48SStefano Zampini IS is_p0; 30444f1b2e48SStefano Zampini MatReuse reuse; 3045c263805aSStefano Zampini 3046c263805aSStefano Zampini /* extract B_0 */ 30474f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 30484f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 30494f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 30504f1b2e48SStefano Zampini } 30514f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 30527dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 3053c263805aSStefano Zampini /* remove rows and cols from local problem */ 3054c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 305597d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 30564f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 30574f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 3058a3df083aSStefano Zampini } else { 3059a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 3060a3df083aSStefano Zampini PetscScalar *vals; 3061a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 3062a3df083aSStefano Zampini 3063a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 3064a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 3065a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 30660b5adadeSStefano Zampini PetscInt *nnz; 3067a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 3068a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 3069a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 3070331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 3071331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3072331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 3073331e053bSStefano Zampini nnz[i] = n - nnz[i]; 3074331e053bSStefano Zampini } 3075331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 3076e1b21442SStefano Zampini ierr = MatSetOption(pcbddc->benign_B0,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3077331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3078331e053bSStefano Zampini } 3079a3df083aSStefano Zampini 3080a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3081a3df083aSStefano Zampini PetscScalar *array; 3082a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 3083a3df083aSStefano Zampini 3084a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 3085a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3086a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 3087a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 3088a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 3089a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 3090a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 3091a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 3092a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 3093a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 3094a3df083aSStefano Zampini cum = 0; 3095a3df083aSStefano Zampini for (j=0;j<n;j++) { 309622db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 3097a3df083aSStefano Zampini vals[cum] = array[j]; 3098a3df083aSStefano Zampini idxs_ins[cum] = j; 3099a3df083aSStefano Zampini cum++; 3100a3df083aSStefano Zampini } 3101a3df083aSStefano Zampini } 3102a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 3103a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 3104a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 3105a3df083aSStefano Zampini } 3106a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3107a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3108a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 3109a3df083aSStefano Zampini } 3110c263805aSStefano Zampini } else { /* push */ 3111a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 31124f1b2e48SStefano Zampini PetscInt i; 31134f1b2e48SStefano Zampini 31144f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 31154f1b2e48SStefano Zampini PetscScalar *B0_vals; 31164f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 31174f1b2e48SStefano Zampini 31184f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 31194f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 31207b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 31214f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 31224f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 31234f1b2e48SStefano Zampini } 3124c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3125c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 31266080607fSStefano Zampini } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!"); 3127c263805aSStefano Zampini } 3128c263805aSStefano Zampini PetscFunctionReturn(0); 3129c263805aSStefano Zampini } 3130c263805aSStefano Zampini 313108122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 3132b1b3d7a2SStefano Zampini { 3133b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 313408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 313508122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 313608122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 313708122e43SStefano Zampini PetscScalar *work,lwork; 313808122e43SStefano Zampini PetscScalar *St,*S,*eigv; 313908122e43SStefano Zampini PetscScalar *Sarray,*Starray; 3140bd2a564bSStefano Zampini PetscReal *eigs,thresh,lthresh,uthresh; 31411b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 3142f6f667cfSStefano Zampini PetscBool allocated_S_St; 314308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 314408122e43SStefano Zampini PetscReal *rwork; 314508122e43SStefano Zampini #endif 3146b1b3d7a2SStefano Zampini PetscErrorCode ierr; 3147b1b3d7a2SStefano Zampini 3148b1b3d7a2SStefano Zampini PetscFunctionBegin; 3149b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 3150af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 3151bd2a564bSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_symmetric)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for this matrix pencil (herm %d, symm %d, posdef %d)",sub_schurs->is_hermitian,sub_schurs->is_symmetric,sub_schurs->is_posdef); 315243371fb9SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_AdaptiveSetUp[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 315306a4e24aSStefano Zampini 3154fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3155fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3156fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3157fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 31581575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3159fd14bc51SStefano Zampini } 3160fd14bc51SStefano Zampini 3161e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 31626080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %D (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef);CHKERRQ(ierr); 3163e496cd5dSStefano Zampini } 3164e496cd5dSStefano Zampini 316508122e43SStefano Zampini /* max size of subsets */ 316608122e43SStefano Zampini mss = 0; 316708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 316808122e43SStefano Zampini PetscInt subset_size; 3169862806e4SStefano Zampini 317008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 317108122e43SStefano Zampini mss = PetscMax(mss,subset_size); 317208122e43SStefano Zampini } 317308122e43SStefano Zampini 317408122e43SStefano Zampini /* min/max and threshold */ 317508122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 3176f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 317708122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 3178f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 3179bd2a564bSStefano Zampini if (nmin || !sub_schurs->is_posdef) { /* XXX */ 3180f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 3181f6f667cfSStefano Zampini } 318208122e43SStefano Zampini 318308122e43SStefano Zampini /* allocate lapack workspace */ 318408122e43SStefano Zampini cum = cum2 = 0; 318508122e43SStefano Zampini maxneigs = 0; 318608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 318708122e43SStefano Zampini PetscInt n,subset_size; 3188f6f667cfSStefano Zampini 318908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 319008122e43SStefano Zampini n = PetscMin(subset_size,nmax); 31919162d606SStefano Zampini cum += subset_size; 31929162d606SStefano Zampini cum2 += subset_size*n; 319308122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 319408122e43SStefano Zampini } 31957ebab0bbSStefano Zampini lwork = 0; 319608122e43SStefano Zampini if (mss) { 3197bd2a564bSStefano Zampini if (sub_schurs->is_symmetric) { 31987ebab0bbSStefano Zampini PetscScalar sdummy = 0.; 319908122e43SStefano Zampini PetscBLASInt B_itype = 1; 32007ebab0bbSStefano Zampini PetscBLASInt B_N = mss, idummy = 0; 32017ebab0bbSStefano Zampini PetscReal rdummy = 0.,zero = 0.0; 32024c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 320308122e43SStefano Zampini 320408122e43SStefano Zampini B_lwork = -1; 32057ebab0bbSStefano Zampini /* some implementations may complain about NULL pointers, even if we are querying */ 32067ebab0bbSStefano Zampini S = &sdummy; 32077ebab0bbSStefano Zampini St = &sdummy; 32087ebab0bbSStefano Zampini eigs = &rdummy; 32097ebab0bbSStefano Zampini eigv = &sdummy; 32107ebab0bbSStefano Zampini B_iwork = &idummy; 32117ebab0bbSStefano Zampini B_ifail = &idummy; 3212d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32137ebab0bbSStefano Zampini rwork = &rdummy; 3214d1710679SStefano Zampini #endif 32158bec7fa6SStefano Zampini thresh = 1.0; 321608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 321708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 321808122e43SStefano 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)); 321908122e43SStefano Zampini #else 322008122e43SStefano 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)); 322108122e43SStefano Zampini #endif 322208122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 322308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3224bd2a564bSStefano Zampini } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented"); 322508122e43SStefano Zampini } 322608122e43SStefano Zampini 322708122e43SStefano Zampini nv = 0; 3228d62866d3SStefano 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) */ 3229d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 323008122e43SStefano Zampini } 32314c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 3232f6f667cfSStefano Zampini if (allocated_S_St) { 3233f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 3234f6f667cfSStefano Zampini } 3235f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 323608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 323708122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 323808122e43SStefano Zampini #endif 32399162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 32409162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 32419162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 324208122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 32439162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 3244580bdb30SBarry Smith ierr = PetscArrayzero(pcbddc->adaptive_constraints_n,nv+sub_schurs->n_subs);CHKERRQ(ierr); 324508122e43SStefano Zampini 324608122e43SStefano Zampini maxneigs = 0; 324772b8c272SStefano Zampini cum = cumarray = 0; 32489162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 32499162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 3250d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 325108122e43SStefano Zampini const PetscInt *idxs; 325208122e43SStefano Zampini 3253d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 325408122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 325508122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 325608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 325708122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 32589162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 32599162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 326008122e43SStefano Zampini } 3261d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 326208122e43SStefano Zampini } 326308122e43SStefano Zampini 326408122e43SStefano Zampini if (mss) { /* multilevel */ 326508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 326608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 326708122e43SStefano Zampini } 326808122e43SStefano Zampini 3269bd2a564bSStefano Zampini lthresh = pcbddc->adaptive_threshold[0]; 3270bd2a564bSStefano Zampini uthresh = pcbddc->adaptive_threshold[1]; 327108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 327208122e43SStefano Zampini const PetscInt *idxs; 32739d54b7f4SStefano Zampini PetscReal upper,lower; 3274862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 327508122e43SStefano Zampini PetscBLASInt B_N; 3276aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 3277bd2a564bSStefano Zampini PetscBool scal = PETSC_FALSE; 327808122e43SStefano Zampini 32799d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 32809d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 3281bd2a564bSStefano Zampini lower = uthresh; 32829d54b7f4SStefano Zampini } else { 3283bd2a564bSStefano Zampini if (!sub_schurs->is_posdef) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented without deluxe scaling"); 3284bd2a564bSStefano Zampini upper = 1./uthresh; 32859d54b7f4SStefano Zampini lower = 0.; 32869d54b7f4SStefano Zampini } 3287862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 3288ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 3289f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 3290bd2a564bSStefano Zampini /* this is experimental: we assume the dofs have been properly grouped to have 3291bd2a564bSStefano Zampini the diagonal blocks Schur complements either positive or negative definite (true for Stokes) */ 3292bd2a564bSStefano Zampini if (!sub_schurs->is_posdef) { 3293bd2a564bSStefano Zampini Mat T; 3294bd2a564bSStefano Zampini 3295bd2a564bSStefano Zampini for (j=0;j<subset_size;j++) { 3296bd2a564bSStefano Zampini if (PetscRealPart(*(Sarray+cumarray+j*(subset_size+1))) < 0.0) { 3297bd2a564bSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,Sarray+cumarray,&T);CHKERRQ(ierr); 3298bd2a564bSStefano Zampini ierr = MatScale(T,-1.0);CHKERRQ(ierr); 3299bd2a564bSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 3300bd2a564bSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,Starray+cumarray,&T);CHKERRQ(ierr); 3301bd2a564bSStefano Zampini ierr = MatScale(T,-1.0);CHKERRQ(ierr); 3302bd2a564bSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 3303bd2a564bSStefano Zampini if (sub_schurs->change_primal_sub) { 3304bd2a564bSStefano Zampini PetscInt nz,k; 3305bd2a564bSStefano Zampini const PetscInt *idxs; 3306bd2a564bSStefano Zampini 3307bd2a564bSStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nz);CHKERRQ(ierr); 3308bd2a564bSStefano Zampini ierr = ISGetIndices(sub_schurs->change_primal_sub[i],&idxs);CHKERRQ(ierr); 3309bd2a564bSStefano Zampini for (k=0;k<nz;k++) { 3310bd2a564bSStefano Zampini *( Sarray + cumarray + idxs[k]*(subset_size+1)) *= -1.0; 3311bd2a564bSStefano Zampini *(Starray + cumarray + idxs[k]*(subset_size+1)) = 0.0; 3312bd2a564bSStefano Zampini } 3313bd2a564bSStefano Zampini ierr = ISRestoreIndices(sub_schurs->change_primal_sub[i],&idxs);CHKERRQ(ierr); 3314bd2a564bSStefano Zampini } 3315bd2a564bSStefano Zampini scal = PETSC_TRUE; 3316bd2a564bSStefano Zampini break; 3317bd2a564bSStefano Zampini } 3318bd2a564bSStefano Zampini } 3319bd2a564bSStefano Zampini } 3320bd2a564bSStefano Zampini 3321f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 3322bd2a564bSStefano Zampini if (sub_schurs->is_symmetric) { 3323aff50787SStefano Zampini PetscInt j,k; 3324580bdb30SBarry Smith if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscArraycmp() later */ 3325580bdb30SBarry Smith ierr = PetscArrayzero(S,subset_size*subset_size);CHKERRQ(ierr); 3326580bdb30SBarry Smith ierr = PetscArrayzero(St,subset_size*subset_size);CHKERRQ(ierr); 332708122e43SStefano Zampini } 332808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 3329aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 3330aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 3331aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 3332aff50787SStefano Zampini } 333308122e43SStefano Zampini } 333408122e43SStefano Zampini } else { 3335580bdb30SBarry Smith ierr = PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3336580bdb30SBarry Smith ierr = PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 333708122e43SStefano Zampini } 33388bec7fa6SStefano Zampini } else { 3339f6f667cfSStefano Zampini S = Sarray + cumarray; 3340f6f667cfSStefano Zampini St = Starray + cumarray; 33418bec7fa6SStefano Zampini } 3342aff50787SStefano Zampini /* see if we can save some work */ 3343b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 3344580bdb30SBarry Smith ierr = PetscArraycmp(S,St,subset_size*subset_size,&same_data);CHKERRQ(ierr); 3345aff50787SStefano Zampini } 3346aff50787SStefano Zampini 3347b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 3348aff50787SStefano Zampini B_neigs = 0; 3349aff50787SStefano Zampini } else { 3350bd2a564bSStefano Zampini if (sub_schurs->is_symmetric) { 335108122e43SStefano Zampini PetscBLASInt B_itype = 1; 3352f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 33534c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 33549552c7c7SStefano Zampini PetscInt nmin_s; 3355bd2a564bSStefano Zampini PetscBool compute_range; 3356bd2a564bSStefano Zampini 33579036ceccSStefano Zampini B_neigs = 0; 3358bd2a564bSStefano Zampini compute_range = (PetscBool)!same_data; 3359bd2a564bSStefano Zampini if (nmin >= subset_size) compute_range = PETSC_FALSE; 336008122e43SStefano Zampini 3361fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 33629036ceccSStefano Zampini PetscInt nc = 0; 3363d16cbb6bSStefano Zampini 33649036ceccSStefano Zampini if (sub_schurs->change_primal_sub) { 33659036ceccSStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nc);CHKERRQ(ierr); 33669036ceccSStefano Zampini } 33676080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %D/%D size %D count %D fid %D (range %d) (change %D).\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]+1,pcbddc->mat_graph->which_dof[idxs[0]],compute_range,nc);CHKERRQ(ierr); 3368b7ab4a40SStefano Zampini } 3369b7ab4a40SStefano Zampini 337008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3371b7ab4a40SStefano Zampini if (compute_range) { 3372d16cbb6bSStefano Zampini 3373d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 3374bd2a564bSStefano Zampini if (sub_schurs->is_posdef) { 337508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 33769d54b7f4SStefano 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)); 337708122e43SStefano Zampini #else 33789d54b7f4SStefano 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)); 337908122e43SStefano Zampini #endif 338043371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 3381bd2a564bSStefano Zampini } else { /* no theory so far, but it works nicely */ 33829036ceccSStefano Zampini PetscInt recipe = 0,recipe_m = 1; 3383bd2a564bSStefano Zampini PetscReal bb[2]; 3384bd2a564bSStefano Zampini 3385bd2a564bSStefano Zampini ierr = PetscOptionsGetInt(NULL,((PetscObject)pc)->prefix,"-pc_bddc_adaptive_recipe",&recipe,NULL);CHKERRQ(ierr); 3386bd2a564bSStefano Zampini switch (recipe) { 3387bd2a564bSStefano Zampini case 0: 3388bd2a564bSStefano Zampini if (scal) { bb[0] = PETSC_MIN_REAL; bb[1] = lthresh; } 3389bd2a564bSStefano Zampini else { bb[0] = uthresh; bb[1] = PETSC_MAX_REAL; } 3390bd2a564bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3391bd2a564bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 3392bd2a564bSStefano Zampini #else 3393bd2a564bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 3394bd2a564bSStefano Zampini #endif 339543371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 3396bd2a564bSStefano Zampini break; 3397bd2a564bSStefano Zampini case 1: 3398bd2a564bSStefano Zampini bb[0] = PETSC_MIN_REAL; bb[1] = lthresh*lthresh; 3399bd2a564bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3400bd2a564bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 3401bd2a564bSStefano Zampini #else 3402bd2a564bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 3403bd2a564bSStefano Zampini #endif 340443371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 3405bd2a564bSStefano Zampini if (!scal) { 34069036ceccSStefano Zampini PetscBLASInt B_neigs2 = 0; 3407bd2a564bSStefano Zampini 3408aed7e7d0SStefano Zampini bb[0] = PetscMax(lthresh*lthresh,uthresh); bb[1] = PETSC_MAX_REAL; 3409580bdb30SBarry Smith ierr = PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3410580bdb30SBarry Smith ierr = PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3411bd2a564bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3412bd2a564bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 3413bd2a564bSStefano Zampini #else 3414bd2a564bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 3415bd2a564bSStefano Zampini #endif 341643371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 3417bd2a564bSStefano Zampini B_neigs += B_neigs2; 3418bd2a564bSStefano Zampini } 3419bd2a564bSStefano Zampini break; 34209036ceccSStefano Zampini case 2: 34219036ceccSStefano Zampini if (scal) { 34229036ceccSStefano Zampini bb[0] = PETSC_MIN_REAL; 34239036ceccSStefano Zampini bb[1] = 0; 34249036ceccSStefano Zampini #if defined(PETSC_USE_COMPLEX) 34259036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 34269036ceccSStefano Zampini #else 34279036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 34289036ceccSStefano Zampini #endif 342943371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 34309036ceccSStefano Zampini } else { 34319036ceccSStefano Zampini PetscBLASInt B_neigs2 = 0; 34329036ceccSStefano Zampini PetscBool import = PETSC_FALSE; 34339036ceccSStefano Zampini 34349036ceccSStefano Zampini lthresh = PetscMax(lthresh,0.0); 34359036ceccSStefano Zampini if (lthresh > 0.0) { 34369036ceccSStefano Zampini bb[0] = PETSC_MIN_REAL; 34379036ceccSStefano Zampini bb[1] = lthresh*lthresh; 34389036ceccSStefano Zampini 34399036ceccSStefano Zampini import = PETSC_TRUE; 34409036ceccSStefano Zampini #if defined(PETSC_USE_COMPLEX) 34419036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 34429036ceccSStefano Zampini #else 34439036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 34449036ceccSStefano Zampini #endif 344543371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 34469036ceccSStefano Zampini } 34479036ceccSStefano Zampini bb[0] = PetscMax(lthresh*lthresh,uthresh); 34489036ceccSStefano Zampini bb[1] = PETSC_MAX_REAL; 34499036ceccSStefano Zampini if (import) { 3450580bdb30SBarry Smith ierr = PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3451580bdb30SBarry Smith ierr = PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 34529036ceccSStefano Zampini } 34539036ceccSStefano Zampini #if defined(PETSC_USE_COMPLEX) 34549036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 34559036ceccSStefano Zampini #else 34569036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 34579036ceccSStefano Zampini #endif 345843371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 34599036ceccSStefano Zampini B_neigs += B_neigs2; 34609036ceccSStefano Zampini } 34619036ceccSStefano Zampini break; 34629036ceccSStefano Zampini case 3: 34639036ceccSStefano Zampini if (scal) { 34649036ceccSStefano Zampini ierr = PetscOptionsGetInt(NULL,((PetscObject)pc)->prefix,"-pc_bddc_adaptive_recipe3_min_scal",&recipe_m,NULL);CHKERRQ(ierr); 34659036ceccSStefano Zampini } else { 34669036ceccSStefano Zampini ierr = PetscOptionsGetInt(NULL,((PetscObject)pc)->prefix,"-pc_bddc_adaptive_recipe3_min",&recipe_m,NULL);CHKERRQ(ierr); 34679036ceccSStefano Zampini } 34689036ceccSStefano Zampini if (!scal) { 34699036ceccSStefano Zampini bb[0] = uthresh; 34709036ceccSStefano Zampini bb[1] = PETSC_MAX_REAL; 34719036ceccSStefano Zampini #if defined(PETSC_USE_COMPLEX) 34729036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 34739036ceccSStefano Zampini #else 34749036ceccSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 34759036ceccSStefano Zampini #endif 347643371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 34779036ceccSStefano Zampini } 34789036ceccSStefano Zampini if (recipe_m > 0 && B_N - B_neigs > 0) { 34799036ceccSStefano Zampini PetscBLASInt B_neigs2 = 0; 34809036ceccSStefano Zampini 34819036ceccSStefano Zampini B_IL = 1; 34829036ceccSStefano Zampini ierr = PetscBLASIntCast(PetscMin(recipe_m,B_N - B_neigs),&B_IU);CHKERRQ(ierr); 3483580bdb30SBarry Smith ierr = PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3484580bdb30SBarry Smith ierr = PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 34859036ceccSStefano Zampini #if defined(PETSC_USE_COMPLEX) 34869036ceccSStefano 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*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 34879036ceccSStefano Zampini #else 34889036ceccSStefano 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*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 34899036ceccSStefano Zampini #endif 349043371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 34919036ceccSStefano Zampini B_neigs += B_neigs2; 34929036ceccSStefano Zampini } 34939036ceccSStefano Zampini break; 349448cebe81SStefano Zampini case 4: 349548cebe81SStefano Zampini bb[0] = PETSC_MIN_REAL; bb[1] = lthresh; 349648cebe81SStefano Zampini #if defined(PETSC_USE_COMPLEX) 349748cebe81SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 349848cebe81SStefano Zampini #else 349948cebe81SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 350048cebe81SStefano Zampini #endif 350143371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 350248cebe81SStefano Zampini { 350348cebe81SStefano Zampini PetscBLASInt B_neigs2 = 0; 350448cebe81SStefano Zampini 350548cebe81SStefano Zampini bb[0] = PetscMax(lthresh+PETSC_SMALL,uthresh); bb[1] = PETSC_MAX_REAL; 3506580bdb30SBarry Smith ierr = PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3507580bdb30SBarry Smith ierr = PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 350848cebe81SStefano Zampini #if defined(PETSC_USE_COMPLEX) 350948cebe81SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 351048cebe81SStefano Zampini #else 351148cebe81SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 351248cebe81SStefano Zampini #endif 351343371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 351448cebe81SStefano Zampini B_neigs += B_neigs2; 351548cebe81SStefano Zampini } 351648cebe81SStefano Zampini break; 351780db8efeSStefano Zampini case 5: /* same as before: first compute all eigenvalues, then filter */ 351880db8efeSStefano Zampini #if defined(PETSC_USE_COMPLEX) 351980db8efeSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","A","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 352080db8efeSStefano Zampini #else 352180db8efeSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","A","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 352280db8efeSStefano Zampini #endif 352343371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 352480db8efeSStefano Zampini { 352580db8efeSStefano Zampini PetscInt e,k,ne; 352680db8efeSStefano Zampini for (e=0,ne=0;e<B_neigs;e++) { 352780db8efeSStefano Zampini if (eigs[e] < lthresh || eigs[e] > uthresh) { 352880db8efeSStefano Zampini for (k=0;k<B_N;k++) S[ne*B_N+k] = eigv[e*B_N+k]; 352980db8efeSStefano Zampini eigs[ne] = eigs[e]; 353080db8efeSStefano Zampini ne++; 353180db8efeSStefano Zampini } 353280db8efeSStefano Zampini } 3533580bdb30SBarry Smith ierr = PetscArraycpy(eigv,S,B_N*ne);CHKERRQ(ierr); 353480db8efeSStefano Zampini B_neigs = ne; 353580db8efeSStefano Zampini } 353680db8efeSStefano Zampini break; 3537bd2a564bSStefano Zampini default: 3538bd2a564bSStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Unknown recipe %D",recipe); 3539bd2a564bSStefano Zampini } 3540bd2a564bSStefano Zampini } 3541bd2a564bSStefano Zampini } else if (!same_data) { /* this is just to see all the eigenvalues */ 3542d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 3543d16cbb6bSStefano Zampini B_IL = 1; 3544d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 35459d54b7f4SStefano 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)); 3546d16cbb6bSStefano Zampini #else 35479d54b7f4SStefano 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)); 3548d16cbb6bSStefano Zampini #endif 354943371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 3550b03ebc13SStefano Zampini } else { /* same_data is true, so just get the adaptive functional requested by the user */ 3551b7ab4a40SStefano Zampini PetscInt k; 3552b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 3553b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 3554b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 3555b7ab4a40SStefano Zampini nmin = nmax; 3556580bdb30SBarry Smith ierr = PetscArrayzero(eigv,subset_size*nmax);CHKERRQ(ierr); 3557b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 3558b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 3559b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 3560b7ab4a40SStefano Zampini } 3561d16cbb6bSStefano Zampini } 356208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 356308122e43SStefano Zampini if (B_ierr) { 35646c4ed002SBarry 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); 35656c4ed002SBarry 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); 35666c4ed002SBarry 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); 356708122e43SStefano Zampini } 356808122e43SStefano Zampini 356908122e43SStefano Zampini if (B_neigs > nmax) { 3570fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 35716080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %D.\n",B_neigs,nmax);CHKERRQ(ierr); 3572fd14bc51SStefano Zampini } 3573bd2a564bSStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = scal ? 0 : B_neigs-nmax; 357408122e43SStefano Zampini B_neigs = nmax; 357508122e43SStefano Zampini } 357608122e43SStefano Zampini 35779552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 35789552c7c7SStefano Zampini if (B_neigs < nmin_s) { 35799036ceccSStefano Zampini PetscBLASInt B_neigs2 = 0; 358008122e43SStefano Zampini 35819d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3582bd2a564bSStefano Zampini if (scal) { 3583bd2a564bSStefano Zampini B_IU = nmin_s; 3584bd2a564bSStefano Zampini B_IL = B_neigs + 1; 3585bd2a564bSStefano Zampini } else { 3586f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 35879d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 3588bd2a564bSStefano Zampini } 35899d54b7f4SStefano Zampini } else { 35909d54b7f4SStefano Zampini B_IL = B_neigs + 1; 35919d54b7f4SStefano Zampini B_IU = nmin_s; 35929d54b7f4SStefano Zampini } 3593fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 35946080607fSStefano Zampini ierr = 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);CHKERRQ(ierr); 3595fd14bc51SStefano Zampini } 3596bd2a564bSStefano Zampini if (sub_schurs->is_symmetric) { 35971ae86dd6SStefano Zampini PetscInt j,k; 359808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 35991ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 36001ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 36011ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 360208122e43SStefano Zampini } 360308122e43SStefano Zampini } 360408122e43SStefano Zampini } else { 3605580bdb30SBarry Smith ierr = PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 3606580bdb30SBarry Smith ierr = PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);CHKERRQ(ierr); 360708122e43SStefano Zampini } 360808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 360908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 36109d54b7f4SStefano 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)); 361108122e43SStefano Zampini #else 36129d54b7f4SStefano 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)); 361308122e43SStefano Zampini #endif 361443371fb9SStefano Zampini ierr = PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);CHKERRQ(ierr); 361508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 361608122e43SStefano Zampini B_neigs += B_neigs2; 361708122e43SStefano Zampini } 361808122e43SStefano Zampini if (B_ierr) { 36196c4ed002SBarry 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); 36206c4ed002SBarry 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); 36216c4ed002SBarry 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); 362208122e43SStefano Zampini } 3623fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3624ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 362508122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 362608122e43SStefano Zampini if (eigs[j] == 0.0) { 3627ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 362808122e43SStefano Zampini } else { 36299d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3630ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 36319d54b7f4SStefano Zampini } else { 36329d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 36339d54b7f4SStefano Zampini } 3634fd14bc51SStefano Zampini } 363508122e43SStefano Zampini } 363608122e43SStefano Zampini } 3637bd2a564bSStefano Zampini } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented"); 3638aff50787SStefano Zampini } 36396c3e6151SStefano Zampini /* change the basis back to the original one */ 36406c3e6151SStefano Zampini if (sub_schurs->change) { 364172b8c272SStefano Zampini Mat change,phi,phit; 36426c3e6151SStefano Zampini 364303dfb2d7SStefano Zampini if (pcbddc->dbg_flag > 2) { 36446c3e6151SStefano Zampini PetscInt ii; 36456c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 36466c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 36476c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 3648684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3649684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 3650684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 3651684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3652684229deSStefano Zampini #else 36536c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 3654684229deSStefano Zampini #endif 36556c3e6151SStefano Zampini } 36566c3e6151SStefano Zampini } 36576c3e6151SStefano Zampini } 365872b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 36596c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 366072b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 36616c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 36626c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 36636c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 36646c3e6151SStefano Zampini } 36658bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 36668bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 36679162d606SStefano Zampini if (B_neigs) { 3668580bdb30SBarry Smith ierr = PetscArraycpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size);CHKERRQ(ierr); 3669fd14bc51SStefano Zampini 3670fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 36719552c7c7SStefano Zampini PetscInt ii; 36729552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 3673ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 36749552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 3675ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3676ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3677ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3678ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3679ac47001eSStefano Zampini #else 3680ac47001eSStefano 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); 3681ac47001eSStefano Zampini #endif 36829552c7c7SStefano Zampini } 36839552c7c7SStefano Zampini } 3684fd14bc51SStefano Zampini } 3685580bdb30SBarry Smith ierr = PetscArraycpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size);CHKERRQ(ierr); 36869162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 36879162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 36889162d606SStefano Zampini cum++; 368908122e43SStefano Zampini } 369008122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 369108122e43SStefano Zampini /* shift for next computation */ 369208122e43SStefano Zampini cumarray += subset_size*subset_size; 369308122e43SStefano Zampini } 3694fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3695fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3696fd14bc51SStefano Zampini } 369708122e43SStefano Zampini 369808122e43SStefano Zampini if (mss) { 369908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 370008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 3701f6f667cfSStefano Zampini /* destroy matrices (junk) */ 3702f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 3703f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 370408122e43SStefano Zampini } 3705f6f667cfSStefano Zampini if (allocated_S_St) { 3706f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 3707f6f667cfSStefano Zampini } 3708f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 370908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 371008122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 371108122e43SStefano Zampini #endif 371208122e43SStefano Zampini if (pcbddc->dbg_flag) { 37131b968477SStefano Zampini PetscInt maxneigs_r; 3714820f2d46SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 37156080607fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %D\n",maxneigs_r);CHKERRQ(ierr); 371608122e43SStefano Zampini } 371743371fb9SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_AdaptiveSetUp[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 371808122e43SStefano Zampini PetscFunctionReturn(0); 371908122e43SStefano Zampini } 3720b1b3d7a2SStefano Zampini 3721c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 3722c8587f34SStefano Zampini { 37238629588bSStefano Zampini PetscScalar *coarse_submat_vals; 3724c8587f34SStefano Zampini PetscErrorCode ierr; 3725c8587f34SStefano Zampini 3726c8587f34SStefano Zampini PetscFunctionBegin; 3727f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 37285e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 3729c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 3730c8587f34SStefano Zampini 3731684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 37320fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 3733684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3734c8587f34SStefano Zampini 37358629588bSStefano Zampini /* 37368629588bSStefano Zampini Setup local correction and local part of coarse basis. 37378629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 37388629588bSStefano Zampini */ 373947f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 37408629588bSStefano Zampini 37418629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 37428629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 37438629588bSStefano Zampini 37448629588bSStefano Zampini /* free */ 37458629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 3746c8587f34SStefano Zampini PetscFunctionReturn(0); 3747c8587f34SStefano Zampini } 3748c8587f34SStefano Zampini 3749674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 3750674ae819SStefano Zampini { 3751674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3752674ae819SStefano Zampini PetscErrorCode ierr; 3753674ae819SStefano Zampini 3754674ae819SStefano Zampini PetscFunctionBegin; 3755674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 375630368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3757674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 3758785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3759674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 3760f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3761f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3762785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 376363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 376463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 3765674ae819SStefano Zampini PetscFunctionReturn(0); 3766674ae819SStefano Zampini } 3767674ae819SStefano Zampini 3768674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 3769674ae819SStefano Zampini { 3770674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 37714f1b2e48SStefano Zampini PetscInt i; 3772674ae819SStefano Zampini PetscErrorCode ierr; 3773674ae819SStefano Zampini 3774674ae819SStefano Zampini PetscFunctionBegin; 37751e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 37761e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 3777a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 3778b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3779674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 378016909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 37811dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 3782674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3783669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 3784fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 37859326c5c6Sstefano_zampini ierr = PCBDDCGraphDestroy(&pcbddc->mat_graph);CHKERRQ(ierr); 37864f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 37874f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 37884f1b2e48SStefano Zampini } 3789e68a0315Sstefano_zampini pcbddc->n_local_subs = 0; 37904f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 3791e68a0315Sstefano_zampini ierr = PCBDDCSubSchursDestroy(&pcbddc->sub_schurs);CHKERRQ(ierr); 3792c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 37938af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 37941c7a958bSStefano Zampini pcbddc->corner_selected = PETSC_FALSE; 3795674ae819SStefano Zampini PetscFunctionReturn(0); 3796674ae819SStefano Zampini } 3797674ae819SStefano Zampini 3798674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 3799674ae819SStefano Zampini { 3800674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3801674ae819SStefano Zampini PetscErrorCode ierr; 3802674ae819SStefano Zampini 3803674ae819SStefano Zampini PetscFunctionBegin; 3804674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 380558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 3806ca92afb2SStefano Zampini PetscScalar *array; 380706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 380806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 380958da7f69SStefano Zampini } 3810674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3811674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 381215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 381315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3814674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 3815674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 3816674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 381706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 3818674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3819674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 38208ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 3821674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 3822674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 3823674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 38249326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_D);CHKERRQ(ierr); 38259326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 38269326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3827f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3828727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 38290e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3830f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 383170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 383281d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 38330369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 38341dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 38354f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 38368b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 3837ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 3838ca92afb2SStefano Zampini PetscInt i; 3839ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3840ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 3841ca92afb2SStefano Zampini } 3842ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 3843ca92afb2SStefano Zampini } 38444f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 3845674ae819SStefano Zampini PetscFunctionReturn(0); 3846674ae819SStefano Zampini } 3847674ae819SStefano Zampini 3848f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 38496bfb1811SStefano Zampini { 38506bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 38516bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 38526bfb1811SStefano Zampini VecType impVecType; 38534f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 38546bfb1811SStefano Zampini PetscErrorCode ierr; 38556bfb1811SStefano Zampini 38566bfb1811SStefano Zampini PetscFunctionBegin; 38574f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 3858b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 38596bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 3860e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 3861e7b262bdSStefano Zampini /* R nodes */ 3862e7b262bdSStefano Zampini old_size = -1; 3863e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 3864e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 3865e7b262bdSStefano Zampini } 3866e7b262bdSStefano Zampini if (n_R != old_size) { 3867e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3868e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 38696bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 38706bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 38716bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 38726bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 3873e7b262bdSStefano Zampini } 3874e7b262bdSStefano Zampini /* local primal dofs */ 3875e7b262bdSStefano Zampini old_size = -1; 3876e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 3877e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 3878e7b262bdSStefano Zampini } 3879e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 3880e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 388183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 3882e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 38836bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 3884e7b262bdSStefano Zampini } 3885e7b262bdSStefano Zampini /* local explicit constraints */ 3886e7b262bdSStefano Zampini old_size = -1; 3887e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 3888e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 3889e7b262bdSStefano Zampini } 3890e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 3891e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 389283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 389383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 389483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 389583b7ccabSStefano Zampini } 38966bfb1811SStefano Zampini PetscFunctionReturn(0); 38976bfb1811SStefano Zampini } 38986bfb1811SStefano Zampini 389947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 390088ebb749SStefano Zampini { 390125084f0cSStefano Zampini PetscErrorCode ierr; 390225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 390388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 390488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3905d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 390625084f0cSStefano Zampini /* submatrices of local problem */ 390780677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 390806656605SStefano Zampini /* submatrices of local coarse problem */ 390906656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 391025084f0cSStefano Zampini /* working matrices */ 391106656605SStefano Zampini Mat C_CR; 391225084f0cSStefano Zampini /* additional working stuff */ 391306656605SStefano Zampini PC pc_R; 3914c58f9fdbSStefano Zampini Mat F,Brhs = NULL; 39155cbda25cSStefano Zampini Vec dummy_vec; 39167ebab0bbSStefano Zampini PetscBool isLU,isCHOL,need_benign_correction,sparserhs; 391725084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 391806656605SStefano Zampini PetscScalar *work; 391906656605SStefano Zampini PetscInt *idx_V_B; 3920ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 392106656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 392206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 392388ebb749SStefano Zampini 392488ebb749SStefano Zampini PetscFunctionBegin; 39259a962809SStefano 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"); 392643371fb9SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_CorrectionSetUp[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 3927ffd830a3SStefano Zampini 3928ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 3929b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 39304f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 3931b371cd4fSStefano Zampini n_B = pcis->n_B; 3932b371cd4fSStefano Zampini n_D = pcis->n - n_B; 393388ebb749SStefano Zampini n_R = pcis->n - n_vertices; 393488ebb749SStefano Zampini 393588ebb749SStefano Zampini /* vertices in boundary numbering */ 3936785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 39370e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 39386080607fSStefano Zampini if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D",n_vertices,i); 393988ebb749SStefano Zampini 394006656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 3941019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 394206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 3943ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 394406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 3945ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 394606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 3947ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 394806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 3949ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 395006656605SStefano Zampini 395106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 395206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 39532958b453SStefano Zampini ierr = PCSetUp(pc_R);CHKERRQ(ierr); 395406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 395506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 3956ffd830a3SStefano Zampini lda_rhs = n_R; 3957a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 39587ebab0bbSStefano Zampini if (isLU || isCHOL) { 395906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 3960b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 3961df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3962d62866d3SStefano Zampini MatFactorType type; 3963d62866d3SStefano Zampini 3964df4d28bfSStefano Zampini F = reuse_solver->F; 39656816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 3966d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 39677ebab0bbSStefano Zampini if (type == MAT_FACTOR_LU) isLU = PETSC_TRUE; 3968ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 396922db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 39707ebab0bbSStefano Zampini } else F = NULL; 397106656605SStefano Zampini 3972c58f9fdbSStefano Zampini /* determine if we can use a sparse right-hand side */ 3973c58f9fdbSStefano Zampini sparserhs = PETSC_FALSE; 3974c58f9fdbSStefano Zampini if (F) { 3975ea799195SBarry Smith MatSolverType solver; 3976c58f9fdbSStefano Zampini 39773ca39a21SBarry Smith ierr = MatFactorGetSolverType(F,&solver);CHKERRQ(ierr); 3978c58f9fdbSStefano Zampini ierr = PetscStrcmp(solver,MATSOLVERMUMPS,&sparserhs);CHKERRQ(ierr); 3979c58f9fdbSStefano Zampini } 3980c58f9fdbSStefano Zampini 3981ffd830a3SStefano Zampini /* allocate workspace */ 3982ffd830a3SStefano Zampini n = 0; 3983ffd830a3SStefano Zampini if (n_constraints) { 3984ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 3985ffd830a3SStefano Zampini } 3986ffd830a3SStefano Zampini if (n_vertices) { 3987ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 3988ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 3989ffd830a3SStefano Zampini } 39902a3a6641Sstefano_zampini if (!pcbddc->symmetric_primal) { 39912a3a6641Sstefano_zampini n = PetscMax(2*lda_rhs*pcbddc->local_primal_size,n); 39922a3a6641Sstefano_zampini } 3993ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 3994ffd830a3SStefano Zampini 39955cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 39965cbda25cSStefano Zampini dummy_vec = NULL; 39975cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 3998c72cccf8SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&dummy_vec);CHKERRQ(ierr); 3999c72cccf8SStefano Zampini ierr = VecSetSizes(dummy_vec,lda_rhs,PETSC_DECIDE);CHKERRQ(ierr); 4000c72cccf8SStefano Zampini ierr = VecSetType(dummy_vec,((PetscObject)pcis->vec1_N)->type_name);CHKERRQ(ierr); 40015cbda25cSStefano Zampini } 40025cbda25cSStefano Zampini 40037ebab0bbSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 40047ebab0bbSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 40057ebab0bbSStefano Zampini 400688ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 400788ebb749SStefano Zampini if (n_constraints) { 4008837cedc9SStefano Zampini Mat M3,C_B; 400906656605SStefano Zampini IS is_aux; 401080677318SStefano Zampini PetscScalar *array,*array2; 401106656605SStefano Zampini 401225084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 401325084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 40147dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 40157dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 401688ebb749SStefano Zampini 401780677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 401880677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 4019c58f9fdbSStefano Zampini if (!sparserhs) { 4020580bdb30SBarry Smith ierr = PetscArrayzero(work,lda_rhs*n_constraints);CHKERRQ(ierr); 402188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 402206656605SStefano Zampini const PetscScalar *row_cmat_values; 402306656605SStefano Zampini const PetscInt *row_cmat_indices; 402406656605SStefano Zampini PetscInt size_of_constraint,j; 402588ebb749SStefano Zampini 402606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 402706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 4028ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 402906656605SStefano Zampini } 403006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 403106656605SStefano Zampini } 4032c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&Brhs);CHKERRQ(ierr); 4033c58f9fdbSStefano Zampini } else { 4034c58f9fdbSStefano Zampini Mat tC_CR; 4035c58f9fdbSStefano Zampini 4036c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 4037c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 4038c58f9fdbSStefano Zampini PetscScalar *aa; 4039c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 4040c58f9fdbSStefano Zampini PetscBool done; 4041c58f9fdbSStefano Zampini 4042c58f9fdbSStefano Zampini ierr = MatGetRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 404313903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed"); 4044c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(C_CR,&aa);CHKERRQ(ierr); 4045c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_constraints,lda_rhs,ii,jj,aa,&tC_CR);CHKERRQ(ierr); 4046c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 404713903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed"); 4048c58f9fdbSStefano Zampini } else { 4049c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)C_CR);CHKERRQ(ierr); 4050c58f9fdbSStefano Zampini tC_CR = C_CR; 4051c58f9fdbSStefano Zampini } 4052c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tC_CR,&Brhs);CHKERRQ(ierr); 4053c58f9fdbSStefano Zampini ierr = MatDestroy(&tC_CR);CHKERRQ(ierr); 4054c58f9fdbSStefano Zampini } 4055ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 405606656605SStefano Zampini if (F) { 4057a3df083aSStefano Zampini if (need_benign_correction) { 4058df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4059a3df083aSStefano Zampini 406072b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 4061580bdb30SBarry Smith ierr = PetscArrayzero(reuse_solver->benign_save_vals,pcbddc->benign_n);CHKERRQ(ierr); 4062a3df083aSStefano Zampini } 4063c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,local_auxmat2_R);CHKERRQ(ierr); 4064a3df083aSStefano Zampini if (need_benign_correction) { 4065a3df083aSStefano Zampini PetscScalar *marr; 4066df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4067a3df083aSStefano Zampini 4068a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 40695cbda25cSStefano Zampini if (lda_rhs != n_R) { 40705cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 40715cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 40725cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 40735cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 40745cbda25cSStefano Zampini } 40755cbda25cSStefano Zampini } else { 4076a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 4077a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 40785cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 4079a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4080a3df083aSStefano Zampini } 40815cbda25cSStefano Zampini } 4082a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 4083a3df083aSStefano Zampini } 408406656605SStefano Zampini } else { 408580677318SStefano Zampini PetscScalar *marr; 408680677318SStefano Zampini 408780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 408806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 4089ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 4090ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 409106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4092c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);CHKERRQ(ierr); 409306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 409406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 409506656605SStefano Zampini } 409680677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 409706656605SStefano Zampini } 4098c58f9fdbSStefano Zampini if (sparserhs) { 4099c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 4100c58f9fdbSStefano Zampini } 4101c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 410280677318SStefano Zampini if (!pcbddc->switch_static) { 410380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 410480677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 410580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 410680677318SStefano Zampini for (i=0;i<n_constraints;i++) { 4107ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 410880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 410980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 411080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 411180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 411280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 411380677318SStefano Zampini } 411480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 411580677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 411672b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 411780677318SStefano Zampini } else { 4118ffd830a3SStefano Zampini if (lda_rhs != n_R) { 4119ffd830a3SStefano Zampini IS dummy; 4120ffd830a3SStefano Zampini 4121ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 41227dae84e0SHong Zhang ierr = MatCreateSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 4123ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 4124ffd830a3SStefano Zampini } else { 412580677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 412680677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 4127ffd830a3SStefano Zampini } 412825084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 412980677318SStefano Zampini } 413080677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 413180677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR})^{-1} */ 413280677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 413380677318SStefano Zampini if (isCHOL) { 413480677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 413580677318SStefano Zampini } else { 413625084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 413780677318SStefano Zampini } 4138837cedc9SStefano Zampini ierr = MatSeqDenseInvertFactors_Private(M3);CHKERRQ(ierr); 413980677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 4140837cedc9SStefano Zampini ierr = MatMatMult(M3,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 414172b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 4142837cedc9SStefano Zampini ierr = MatCopy(M3,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 4143837cedc9SStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 4144f4ddd8eeSStefano Zampini } 4145fc227af8SStefano Zampini 4146fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 414788ebb749SStefano Zampini if (n_vertices) { 41487ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 41497ebab0bbSStefano Zampini PetscBool oldpin; 41507ebab0bbSStefano Zampini #endif 41517ebab0bbSStefano Zampini PetscBool isaij; 415206656605SStefano Zampini IS is_aux; 41533a50541eSStefano Zampini 4154b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 41556816873aSStefano Zampini IS tis; 41566816873aSStefano Zampini 41576816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 41586816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 41596816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 41606816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 41616816873aSStefano Zampini } else { 41623a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 41636816873aSStefano Zampini } 41647ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 4165b470e4b4SRichard Tran Mills oldpin = pcbddc->local_mat->boundtocpu; 41667ebab0bbSStefano Zampini #endif 4167b470e4b4SRichard Tran Mills ierr = MatBindToCPU(pcbddc->local_mat,PETSC_TRUE);CHKERRQ(ierr); 41687dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 41697dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 41707ebab0bbSStefano Zampini ierr = PetscObjectBaseTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isaij);CHKERRQ(ierr); 41717ebab0bbSStefano Zampini if (!isaij) { /* TODO REMOVE: MatMatMult(A_VR,A_RRmA_RV) below may raise an error */ 4172c58f9fdbSStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 41737ebab0bbSStefano Zampini } 41747dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 41757ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 4176b470e4b4SRichard Tran Mills ierr = MatBindToCPU(pcbddc->local_mat,oldpin);CHKERRQ(ierr); 41777ebab0bbSStefano Zampini #endif 417825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 417988ebb749SStefano Zampini } 418088ebb749SStefano Zampini 418188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 4182f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 418306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 418406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 418506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 418606656605SStefano Zampini } 4187f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 418806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 418906656605SStefano Zampini PetscScalar *marray; 419006656605SStefano Zampini 419106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 419206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 4193f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 4194f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 4195f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 4196f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 4197f4ddd8eeSStefano Zampini } 4198f4ddd8eeSStefano Zampini } 419906656605SStefano Zampini 4200f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 4201a6e023c1Sstefano_zampini PetscScalar *marr; 420288ebb749SStefano Zampini 4203a6e023c1Sstefano_zampini /* memory size */ 420406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 4205a6e023c1Sstefano_zampini if (pcbddc->switch_static || pcbddc->dbg_flag) n += n_D*pcbddc->local_primal_size; 4206a6e023c1Sstefano_zampini if (!pcbddc->symmetric_primal) n *= 2; 4207a6e023c1Sstefano_zampini ierr = PetscCalloc1(n,&marr);CHKERRQ(ierr); 4208a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 4209a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 42108eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4211a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 4212a6e023c1Sstefano_zampini marr += n_D*pcbddc->local_primal_size; 421388ebb749SStefano Zampini } 42143301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 4215a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 4216a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 42178eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4218a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 421988ebb749SStefano Zampini } 422088ebb749SStefano Zampini } else { 4221c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 4222c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 42231b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4224c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 4225c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 4226c0553b1fSStefano Zampini } 422788ebb749SStefano Zampini } 422806656605SStefano Zampini } 4229019a44ceSStefano Zampini 423006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 42314f1b2e48SStefano Zampini p0_lidx_I = NULL; 42324f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 4233d12edf2fSStefano Zampini const PetscInt *idxs; 4234d12edf2fSStefano Zampini 4235d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 42364f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 42374f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 42384f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 42394f1b2e48SStefano Zampini } 4240d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 4241d12edf2fSStefano Zampini } 4242d16cbb6bSStefano Zampini 424306656605SStefano Zampini /* vertices */ 424406656605SStefano Zampini if (n_vertices) { 4245c58f9fdbSStefano Zampini PetscBool restoreavr = PETSC_FALSE; 424616f15bc4SStefano Zampini 4247af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 424804708bb6SStefano Zampini 424916f15bc4SStefano Zampini if (n_R) { 425014393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 425106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 42521683a169SBarry Smith const PetscScalar *x; 42531683a169SBarry Smith PetscScalar *y; 425406656605SStefano Zampini 425521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 425614393ed6SStefano Zampini if (need_benign_correction) { 425714393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 425814393ed6SStefano Zampini IS is_p0; 425914393ed6SStefano Zampini PetscInt *idxs_p0,n; 426014393ed6SStefano Zampini 426114393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 426214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 426314393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 42646080607fSStefano Zampini if (n != pcbddc->benign_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %D != %D",n,pcbddc->benign_n); 426514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 426614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 42677dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 426814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 426914393ed6SStefano Zampini } 427014393ed6SStefano Zampini 4271c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 4272c58f9fdbSStefano Zampini if (!sparserhs || need_benign_correction) { 4273ffd830a3SStefano Zampini if (lda_rhs == n_R) { 4274af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 4275ffd830a3SStefano Zampini } else { 4276ca92afb2SStefano Zampini PetscScalar *av,*array; 4277ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 4278ca92afb2SStefano Zampini PetscInt n; 4279ca92afb2SStefano Zampini PetscBool flg_row; 4280ffd830a3SStefano Zampini 4281ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 4282580bdb30SBarry Smith ierr = PetscArrayzero(array,lda_rhs*n_vertices);CHKERRQ(ierr); 42839d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 4284ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4285ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 4286ca92afb2SStefano Zampini for (i=0;i<n;i++) { 4287ca92afb2SStefano Zampini PetscInt j; 4288ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 4289ffd830a3SStefano Zampini } 4290ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4291ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 4292ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 4293ffd830a3SStefano Zampini } 4294a3df083aSStefano Zampini if (need_benign_correction) { 4295df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4296a3df083aSStefano Zampini PetscScalar *marr; 4297a3df083aSStefano Zampini 4298a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 429914393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 430014393ed6SStefano Zampini 430114393ed6SStefano Zampini | 0 0 0 | (V) 430214393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 430314393ed6SStefano Zampini | 0 0 -1 | (p0) 430414393ed6SStefano Zampini 430514393ed6SStefano Zampini */ 4306df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 430714393ed6SStefano Zampini const PetscScalar *vals; 430814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 430914393ed6SStefano Zampini PetscInt n,j,nz; 431014393ed6SStefano Zampini 4311df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 4312df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 431314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 431414393ed6SStefano Zampini for (j=0;j<n;j++) { 431514393ed6SStefano Zampini PetscScalar val = vals[j]; 431614393ed6SStefano Zampini PetscInt k,col = idxs[j]; 431714393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 431814393ed6SStefano Zampini } 431914393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 4320df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 432114393ed6SStefano Zampini } 432272b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 432372b8c272SStefano Zampini } 4324c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RV);CHKERRQ(ierr); 4325c58f9fdbSStefano Zampini Brhs = A_RV; 4326c58f9fdbSStefano Zampini } else { 4327c58f9fdbSStefano Zampini Mat tA_RVT,A_RVT; 4328c58f9fdbSStefano Zampini 4329c58f9fdbSStefano Zampini if (!pcbddc->symmetric_primal) { 4330fb6280fbSStefano Zampini /* A_RV already scaled by -1 */ 4331c58f9fdbSStefano Zampini ierr = MatTranspose(A_RV,MAT_INITIAL_MATRIX,&A_RVT);CHKERRQ(ierr); 4332c58f9fdbSStefano Zampini } else { 4333c58f9fdbSStefano Zampini restoreavr = PETSC_TRUE; 4334c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 4335c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_VR);CHKERRQ(ierr); 4336c58f9fdbSStefano Zampini A_RVT = A_VR; 4337c58f9fdbSStefano Zampini } 4338c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 4339c58f9fdbSStefano Zampini PetscScalar *aa; 4340c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 4341c58f9fdbSStefano Zampini PetscBool done; 4342c58f9fdbSStefano Zampini 4343c58f9fdbSStefano Zampini ierr = MatGetRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 434413903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed"); 4345c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(A_RVT,&aa);CHKERRQ(ierr); 4346c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_vertices,lda_rhs,ii,jj,aa,&tA_RVT);CHKERRQ(ierr); 4347c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 434813903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed"); 4349c58f9fdbSStefano Zampini } else { 4350c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RVT);CHKERRQ(ierr); 4351c58f9fdbSStefano Zampini tA_RVT = A_RVT; 4352c58f9fdbSStefano Zampini } 4353c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tA_RVT,&Brhs);CHKERRQ(ierr); 4354c58f9fdbSStefano Zampini ierr = MatDestroy(&tA_RVT);CHKERRQ(ierr); 4355c58f9fdbSStefano Zampini ierr = MatDestroy(&A_RVT);CHKERRQ(ierr); 4356c58f9fdbSStefano Zampini } 435772b8c272SStefano Zampini if (F) { 435814393ed6SStefano Zampini /* need to correct the rhs */ 435972b8c272SStefano Zampini if (need_benign_correction) { 436072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 436172b8c272SStefano Zampini PetscScalar *marr; 436272b8c272SStefano Zampini 4363c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&marr);CHKERRQ(ierr); 43645cbda25cSStefano Zampini if (lda_rhs != n_R) { 43655cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 43665cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 43675cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 43685cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 43695cbda25cSStefano Zampini } 43705cbda25cSStefano Zampini } else { 4371a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 4372a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 43735cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 4374a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4375a3df083aSStefano Zampini } 43765cbda25cSStefano Zampini } 4377c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&marr);CHKERRQ(ierr); 4378a3df083aSStefano Zampini } 4379c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,A_RRmA_RV);CHKERRQ(ierr); 4380c58f9fdbSStefano Zampini if (restoreavr) { 4381c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 4382c58f9fdbSStefano Zampini } 438314393ed6SStefano Zampini /* need to correct the solution */ 4384a3df083aSStefano Zampini if (need_benign_correction) { 4385df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4386a3df083aSStefano Zampini PetscScalar *marr; 4387a3df083aSStefano Zampini 4388a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 43895cbda25cSStefano Zampini if (lda_rhs != n_R) { 43905cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 43915cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 43925cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 43935cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 43945cbda25cSStefano Zampini } 43955cbda25cSStefano Zampini } else { 4396a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 4397a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 43985cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 4399a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4400a3df083aSStefano Zampini } 44015cbda25cSStefano Zampini } 4402a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 4403a3df083aSStefano Zampini } 440406656605SStefano Zampini } else { 4405c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&y);CHKERRQ(ierr); 440606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 4407ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 4408ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 440906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4410c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);CHKERRQ(ierr); 441106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 441206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 441306656605SStefano Zampini } 4414c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&y);CHKERRQ(ierr); 441506656605SStefano Zampini } 441680677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 4417c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 4418ffd830a3SStefano Zampini /* S_VV and S_CV */ 441906656605SStefano Zampini if (n_constraints) { 442006656605SStefano Zampini Mat B; 442180677318SStefano Zampini 4422580bdb30SBarry Smith ierr = PetscArrayzero(work+lda_rhs*n_vertices,n_B*n_vertices);CHKERRQ(ierr); 442380677318SStefano Zampini for (i=0;i<n_vertices;i++) { 4424ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 4425ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 442680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 442780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 442880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 442980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 443080677318SStefano Zampini } 4431ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 44324222ddf1SHong Zhang /* Reuse dense S_C = pcbddc->local_auxmat1 * B */ 44334222ddf1SHong Zhang ierr = MatProductCreateWithMat(pcbddc->local_auxmat1,B,NULL,S_CV);CHKERRQ(ierr); 44344222ddf1SHong Zhang ierr = MatProductSetType(S_CV,MATPRODUCT_AB);CHKERRQ(ierr); 44354222ddf1SHong Zhang ierr = MatProductSetFromOptions(S_CV);CHKERRQ(ierr); 44366718818eSStefano Zampini ierr = MatProductSymbolic(S_CV);CHKERRQ(ierr); 44374222ddf1SHong Zhang ierr = MatProductNumeric(S_CV);CHKERRQ(ierr); 4438a961b312SStefano Zampini ierr = MatProductClear(S_CV);CHKERRQ(ierr); 44394222ddf1SHong Zhang 444080677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 4441ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 44424222ddf1SHong Zhang /* Reuse B = local_auxmat2_R * S_CV */ 44434222ddf1SHong Zhang ierr = MatProductCreateWithMat(local_auxmat2_R,S_CV,NULL,B);CHKERRQ(ierr); 44444222ddf1SHong Zhang ierr = MatProductSetType(B,MATPRODUCT_AB);CHKERRQ(ierr); 44454222ddf1SHong Zhang ierr = MatProductSetFromOptions(B);CHKERRQ(ierr); 44466718818eSStefano Zampini ierr = MatProductSymbolic(B);CHKERRQ(ierr); 44474222ddf1SHong Zhang ierr = MatProductNumeric(B);CHKERRQ(ierr); 44484222ddf1SHong Zhang 444906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 4450ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 4451ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 445206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 445306656605SStefano Zampini } 4454ffd830a3SStefano Zampini if (lda_rhs != n_R) { 4455ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 4456ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 4457ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 4458ffd830a3SStefano Zampini } 445906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 446014393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 446114393ed6SStefano Zampini if (need_benign_correction) { 4462df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 446314393ed6SStefano Zampini PetscScalar *marr,*sums; 446414393ed6SStefano Zampini 446514393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 4466f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 4467df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 446814393ed6SStefano Zampini const PetscScalar *vals; 446914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 447014393ed6SStefano Zampini PetscInt n,j,nz; 447114393ed6SStefano Zampini 4472df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 4473df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 447414393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 447514393ed6SStefano Zampini PetscInt k; 447614393ed6SStefano Zampini sums[j] = 0.; 447714393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 447814393ed6SStefano Zampini } 447914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 448014393ed6SStefano Zampini for (j=0;j<n;j++) { 448114393ed6SStefano Zampini PetscScalar val = vals[j]; 448214393ed6SStefano Zampini PetscInt k; 448314393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 448414393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 448514393ed6SStefano Zampini } 448614393ed6SStefano Zampini } 448714393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 4488df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 448914393ed6SStefano Zampini } 449014393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 4491f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 449214393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 449314393ed6SStefano Zampini } 449480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 449506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 44961683a169SBarry Smith ierr = MatDenseGetArrayRead(A_VV,&x);CHKERRQ(ierr); 449706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 449806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 44991683a169SBarry Smith ierr = MatDenseRestoreArrayRead(A_VV,&x);CHKERRQ(ierr); 450006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 450106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4502d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 4503019a44ceSStefano Zampini } else { 4504d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4505d16cbb6bSStefano Zampini } 450621eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 4507d16cbb6bSStefano Zampini 450806656605SStefano Zampini /* coarse basis functions */ 450906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 451016f15bc4SStefano Zampini PetscScalar *y; 451116f15bc4SStefano Zampini 4512ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 451306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 451406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 451506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 451606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 451706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 451806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 451906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 452006656605SStefano Zampini 452106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 45224f1b2e48SStefano Zampini PetscInt j; 45234f1b2e48SStefano Zampini 452406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 452506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 452606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 452706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 452806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 45294f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 453006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 453106656605SStefano Zampini } 453206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 453306656605SStefano Zampini } 453404708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 453504708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 453606656605SStefano Zampini } 45375cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 453806656605SStefano Zampini 453906656605SStefano Zampini if (n_constraints) { 454006656605SStefano Zampini Mat B; 454106656605SStefano Zampini 4542ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 454306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 4544a961b312SStefano Zampini ierr = MatProductCreateWithMat(local_auxmat2_R,S_CC,NULL,B);CHKERRQ(ierr); 4545a961b312SStefano Zampini ierr = MatProductSetType(B,MATPRODUCT_AB);CHKERRQ(ierr); 4546a961b312SStefano Zampini ierr = MatProductSetFromOptions(B);CHKERRQ(ierr); 45476718818eSStefano Zampini ierr = MatProductSymbolic(B);CHKERRQ(ierr); 4548a961b312SStefano Zampini ierr = MatProductNumeric(B);CHKERRQ(ierr); 4549a961b312SStefano Zampini 455006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 455106656605SStefano Zampini if (n_vertices) { 455203dfb2d7SStefano Zampini if (isCHOL || need_benign_correction) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 455380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 455480677318SStefano Zampini } else { 455580677318SStefano Zampini Mat S_VCt; 455680677318SStefano Zampini 4557ffd830a3SStefano Zampini if (lda_rhs != n_R) { 4558ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 455972b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 4560ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 4561ffd830a3SStefano Zampini } 456280677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 456380677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 456480677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 456580677318SStefano Zampini } 456606656605SStefano Zampini } 456706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 456806656605SStefano Zampini /* coarse basis functions */ 456906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 457006656605SStefano Zampini PetscScalar *y; 457106656605SStefano Zampini 4572ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 457306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 457406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 457506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 457606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 457706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 457806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 457906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 45804f1b2e48SStefano Zampini PetscInt j; 45814f1b2e48SStefano Zampini 458206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 458306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 458406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 458506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 458606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 45874f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 458806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 458906656605SStefano Zampini } 459006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 459106656605SStefano Zampini } 459206656605SStefano Zampini } 459380677318SStefano Zampini if (n_constraints) { 459480677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 459580677318SStefano Zampini } 45964f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 459772b8c272SStefano Zampini 459872b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 459972b8c272SStefano Zampini if (pcbddc->benign_n) { 460072b8c272SStefano Zampini Mat B0_B,B0_BPHI; 460172b8c272SStefano Zampini IS is_dummy; 46021683a169SBarry Smith const PetscScalar *data; 460372b8c272SStefano Zampini PetscInt j; 460472b8c272SStefano Zampini 460572b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 46067dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 460772b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 460872b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 460986c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 46101683a169SBarry Smith ierr = MatDenseGetArrayRead(B0_BPHI,&data);CHKERRQ(ierr); 461172b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 461272b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 461372b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 461472b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 461572b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 461672b8c272SStefano Zampini } 461772b8c272SStefano Zampini } 46181683a169SBarry Smith ierr = MatDenseRestoreArrayRead(B0_BPHI,&data);CHKERRQ(ierr); 461972b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 462072b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 462172b8c272SStefano Zampini } 4622019a44ceSStefano Zampini 462306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 46243301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 4625ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 4626ffd830a3SStefano Zampini PetscScalar *marray; 462706656605SStefano Zampini 462806656605SStefano Zampini if (n_constraints) { 4629ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 463006656605SStefano Zampini 4631abc8f43dSstefano_zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 463206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 4633ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 463416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 463506656605SStefano Zampini if (n_vertices) { 4636ffd830a3SStefano Zampini Mat S_VCT; 463706656605SStefano Zampini 463806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 4639ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 464016f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 464106656605SStefano Zampini } 4642ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 46435b782168SStefano Zampini } else { 46445b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 464506656605SStefano Zampini } 464616f15bc4SStefano Zampini if (n_vertices && n_R) { 4647ffd830a3SStefano Zampini PetscScalar *av,*marray; 4648ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 4649ffd830a3SStefano Zampini PetscInt n; 4650ffd830a3SStefano Zampini PetscBool flg_row; 465106656605SStefano Zampini 4652ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 4653af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 4654ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4655ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 4656ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4657ffd830a3SStefano Zampini for (i=0;i<n;i++) { 4658ffd830a3SStefano Zampini PetscInt j; 4659ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 4660ffd830a3SStefano Zampini } 4661ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4662ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4663ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 466406656605SStefano Zampini } 466506656605SStefano Zampini 4666ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 4667abc8f43dSstefano_zampini if (n_vertices) { 4668ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4669ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 4670ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 4671ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 467206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4673c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);CHKERRQ(ierr); 467406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 467506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 467606656605SStefano Zampini } 4677ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4678abc8f43dSstefano_zampini } 46795b782168SStefano Zampini if (B_C) { 4680ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 4681ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 4682ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 4683ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 4684ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4685c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);CHKERRQ(ierr); 4686ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4687ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 468806656605SStefano Zampini } 4689ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 46905b782168SStefano Zampini } 469106656605SStefano Zampini /* coarse basis functions */ 469206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 469306656605SStefano Zampini PetscScalar *y; 469406656605SStefano Zampini 4695ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 469606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 469706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 469806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 469906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 470006656605SStefano Zampini if (i<n_vertices) { 470106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 470206656605SStefano Zampini } 470306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 470406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 470506656605SStefano Zampini 470606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 470706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 470806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 470906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 471006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 471106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 471206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 471306656605SStefano Zampini } 471406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 471506656605SStefano Zampini } 4716ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 4717ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 471806656605SStefano Zampini } 4719a6e023c1Sstefano_zampini 4720d62866d3SStefano Zampini /* free memory */ 472188ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 472206656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 472306656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 472406656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 472506656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 4726d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4727d62866d3SStefano Zampini if (n_vertices) { 4728d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 4729d62866d3SStefano Zampini } 4730d62866d3SStefano Zampini if (n_constraints) { 4731d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 4732d62866d3SStefano Zampini } 47338ead10e4SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_CorrectionSetUp[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 47348ead10e4SStefano Zampini 473588ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 473688ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 473788ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 4738d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 473988ebb749SStefano Zampini Mat coarse_sub_mat; 474025084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 474188ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 474288ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 474388ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 47448bec7fa6SStefano Zampini Mat C_B,CPHI; 47458bec7fa6SStefano Zampini IS is_dummy; 47468bec7fa6SStefano Zampini Vec mones; 474788ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 474888ebb749SStefano Zampini PetscReal real_value; 474988ebb749SStefano Zampini 4750a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 4751a3df083aSStefano Zampini Mat A; 4752a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 47537dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 47547dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 47557dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 47567dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4757a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 4758a3df083aSStefano Zampini } else { 475988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 476088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 476188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 476288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4763a3df083aSStefano Zampini } 476488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 476588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 4766ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 476788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 476888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 476988ebb749SStefano Zampini } 477088ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 477188ebb749SStefano Zampini 477225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 47733301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 477425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4775ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 477688ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 477788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 477888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 477988ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 478088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 478188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 478288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 478388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 478488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 478588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 478688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 478788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 478888ebb749SStefano Zampini } else { 478988ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 479088ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 479188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 479288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 479388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 479488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 479588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 479688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 479788ebb749SStefano Zampini } 479888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 479988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 480088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 4801511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 48024f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4803fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 48041683a169SBarry Smith const PetscScalar *data2; 48051683a169SBarry Smith PetscScalar *data; 48064f1b2e48SStefano Zampini PetscInt j; 4807d12edf2fSStefano Zampini 48084f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 48097dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 4810d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 481186c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 4812d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 48131683a169SBarry Smith ierr = MatDenseGetArrayRead(B0_BPHI,&data2);CHKERRQ(ierr); 48144f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 48154f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 4816d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 48174f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 48184f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 48194f1b2e48SStefano Zampini } 4820d12edf2fSStefano Zampini } 4821d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 48221683a169SBarry Smith ierr = MatDenseRestoreArrayRead(B0_BPHI,&data2);CHKERRQ(ierr); 4823d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 4824d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 4825d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 4826d12edf2fSStefano Zampini } 4827d12edf2fSStefano Zampini #if 0 4828d12edf2fSStefano Zampini { 4829d12edf2fSStefano Zampini PetscViewer viewer; 4830d12edf2fSStefano Zampini char filename[256]; 4831ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4832d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4833a7414863SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4834ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 4835ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 4836ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 4837d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 4838a7414863SStefano Zampini if (pcbddc->coarse_phi_B) { 4839ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 4840ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 484172b8c272SStefano Zampini } 4842ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 4843ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 4844ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 4845ffd830a3SStefano Zampini } 4846ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 4847ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 4848ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 4849ffd830a3SStefano Zampini } 485072b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 4851ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 4852ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 4853ffd830a3SStefano Zampini } 4854fb6280fbSStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->local_mat,"A");CHKERRQ(ierr); 4855fb6280fbSStefano Zampini ierr = MatView(pcbddc->local_mat,viewer);CHKERRQ(ierr); 4856fb6280fbSStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->ConstraintMatrix,"C");CHKERRQ(ierr); 4857fb6280fbSStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,viewer);CHKERRQ(ierr); 4858fb6280fbSStefano Zampini ierr = PetscObjectSetName((PetscObject)pcis->is_I_local,"I");CHKERRQ(ierr); 4859fb6280fbSStefano Zampini ierr = ISView(pcis->is_I_local,viewer);CHKERRQ(ierr); 4860fb6280fbSStefano Zampini ierr = PetscObjectSetName((PetscObject)pcis->is_B_local,"B");CHKERRQ(ierr); 4861fb6280fbSStefano Zampini ierr = ISView(pcis->is_B_local,viewer);CHKERRQ(ierr); 4862fb6280fbSStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->is_R_local,"R");CHKERRQ(ierr); 4863fb6280fbSStefano Zampini ierr = ISView(pcbddc->is_R_local,viewer);CHKERRQ(ierr); 4864d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4865d12edf2fSStefano Zampini } 4866d12edf2fSStefano Zampini #endif 486781d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 48688bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 48691575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 487006656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 48718bec7fa6SStefano Zampini 48728bec7fa6SStefano Zampini /* check constraints */ 4873a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 48747dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 48754f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 48768bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4877a00504b5SStefano Zampini } else { 4878a00504b5SStefano Zampini PetscScalar *data; 4879a00504b5SStefano Zampini Mat tmat; 4880a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4881a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 4882a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4883a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4884a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4885a00504b5SStefano Zampini } 48868bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 48878bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 48888bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 48898bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4890bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 4891ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 4892bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4893bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 4894bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 4895bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4896bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 489788ebb749SStefano Zampini } 48988bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 48998bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 49008bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 49018bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 490225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 490388ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 490488ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 490588ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 490688ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 490788ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 490888ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 490988ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 491088ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 491188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 491288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 4913ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 491488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 491588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 491688ebb749SStefano Zampini } 491788ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 491888ebb749SStefano Zampini } 49197ebab0bbSStefano Zampini /* FINAL CUDA support (we cannot currently mix viennacl and cuda vectors */ 49207ebab0bbSStefano Zampini { 49217ebab0bbSStefano Zampini PetscBool gpu; 49227ebab0bbSStefano Zampini 49237ebab0bbSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->vec1_N,VECSEQCUDA,&gpu);CHKERRQ(ierr); 49247ebab0bbSStefano Zampini if (gpu) { 49257ebab0bbSStefano Zampini if (pcbddc->local_auxmat1) { 49267ebab0bbSStefano Zampini ierr = MatConvert(pcbddc->local_auxmat1,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->local_auxmat1);CHKERRQ(ierr); 49277ebab0bbSStefano Zampini } 49287ebab0bbSStefano Zampini if (pcbddc->local_auxmat2) { 49297ebab0bbSStefano Zampini ierr = MatConvert(pcbddc->local_auxmat2,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 49307ebab0bbSStefano Zampini } 49317ebab0bbSStefano Zampini if (pcbddc->coarse_phi_B) { 49327ebab0bbSStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 49337ebab0bbSStefano Zampini } 49347ebab0bbSStefano Zampini if (pcbddc->coarse_phi_D) { 49357ebab0bbSStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 49367ebab0bbSStefano Zampini } 49377ebab0bbSStefano Zampini if (pcbddc->coarse_psi_B) { 49387ebab0bbSStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 49397ebab0bbSStefano Zampini } 49407ebab0bbSStefano Zampini if (pcbddc->coarse_psi_D) { 49417ebab0bbSStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 49427ebab0bbSStefano Zampini } 49437ebab0bbSStefano Zampini } 49447ebab0bbSStefano Zampini } 49458629588bSStefano Zampini /* get back data */ 49468629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 494788ebb749SStefano Zampini PetscFunctionReturn(0); 494888ebb749SStefano Zampini } 494988ebb749SStefano Zampini 49507dae84e0SHong Zhang PetscErrorCode MatCreateSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 4951aa0d41d4SStefano Zampini { 4952d65f70fdSStefano Zampini Mat *work_mat; 4953d65f70fdSStefano Zampini IS isrow_s,iscol_s; 4954d65f70fdSStefano Zampini PetscBool rsorted,csorted; 4955c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 4956aa0d41d4SStefano Zampini PetscErrorCode ierr; 4957aa0d41d4SStefano Zampini 4958aa0d41d4SStefano Zampini PetscFunctionBegin; 4959d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 4960d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 4961d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 4962d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 4963aa0d41d4SStefano Zampini 4964d65f70fdSStefano Zampini if (!rsorted) { 4965906d46d4SStefano Zampini const PetscInt *idxs; 4966906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 4967aa0d41d4SStefano Zampini 4968d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 4969d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 4970d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4971d65f70fdSStefano Zampini idxs_perm_r[i] = i; 4972aa0d41d4SStefano Zampini } 4973d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 4974d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 4975d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4976d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 4977aa0d41d4SStefano Zampini } 4978d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 4979d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 4980d65f70fdSStefano Zampini } else { 4981d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 4982d65f70fdSStefano Zampini isrow_s = isrow; 4983aa0d41d4SStefano Zampini } 4984906d46d4SStefano Zampini 4985d65f70fdSStefano Zampini if (!csorted) { 4986d65f70fdSStefano Zampini if (isrow == iscol) { 4987d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 4988d65f70fdSStefano Zampini iscol_s = isrow_s; 4989d65f70fdSStefano Zampini } else { 4990d65f70fdSStefano Zampini const PetscInt *idxs; 4991d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 4992906d46d4SStefano Zampini 4993d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 4994d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 4995d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4996d65f70fdSStefano Zampini idxs_perm_c[i] = i; 4997d65f70fdSStefano Zampini } 4998d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 4999d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 5000d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 5001d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 5002d65f70fdSStefano Zampini } 5003d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 5004d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 5005d65f70fdSStefano Zampini } 5006d65f70fdSStefano Zampini } else { 5007d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 5008d65f70fdSStefano Zampini iscol_s = iscol; 5009d65f70fdSStefano Zampini } 5010d65f70fdSStefano Zampini 50117dae84e0SHong Zhang ierr = MatCreateSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 5012d65f70fdSStefano Zampini 5013d65f70fdSStefano Zampini if (!rsorted || !csorted) { 5014906d46d4SStefano Zampini Mat new_mat; 5015d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 5016906d46d4SStefano Zampini 5017d65f70fdSStefano Zampini if (!rsorted) { 5018d65f70fdSStefano Zampini PetscInt *idxs_r,i; 5019d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 5020d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 5021d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 5022906d46d4SStefano Zampini } 5023d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 5024d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 5025d65f70fdSStefano Zampini } else { 5026d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 5027906d46d4SStefano Zampini } 5028d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 5029d65f70fdSStefano Zampini 5030d65f70fdSStefano Zampini if (!csorted) { 5031d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 5032d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 5033d65f70fdSStefano Zampini is_perm_c = is_perm_r; 5034d65f70fdSStefano Zampini } else { 5035d65f70fdSStefano Zampini PetscInt *idxs_c,i; 5036f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 5037d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 5038d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 5039d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 5040d65f70fdSStefano Zampini } 5041d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 5042d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 5043d65f70fdSStefano Zampini } 5044d65f70fdSStefano Zampini } else { 5045d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 5046d65f70fdSStefano Zampini } 5047d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 5048d65f70fdSStefano Zampini 5049d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 5050d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 5051d65f70fdSStefano Zampini work_mat[0] = new_mat; 5052d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 5053d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 5054d65f70fdSStefano Zampini } 5055d65f70fdSStefano Zampini 5056d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 5057d65f70fdSStefano Zampini *B = work_mat[0]; 5058d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 5059d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 5060d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 5061d65f70fdSStefano Zampini PetscFunctionReturn(0); 5062d65f70fdSStefano Zampini } 5063d65f70fdSStefano Zampini 50645e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 5065aa0d41d4SStefano Zampini { 5066aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 50675e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5068022d8d2bSstefano_zampini Mat new_mat,lA; 50695e8657edSStefano Zampini IS is_local,is_global; 5070d65f70fdSStefano Zampini PetscInt local_size; 5071d65f70fdSStefano Zampini PetscBool isseqaij; 5072aa0d41d4SStefano Zampini PetscErrorCode ierr; 5073aa0d41d4SStefano Zampini 5074aa0d41d4SStefano Zampini PetscFunctionBegin; 5075aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 50765e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 50775e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 5078b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 5079aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 50807dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 5081aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 5082906d46d4SStefano Zampini 5083906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 5084906d46d4SStefano Zampini Vec x,x_change; 5085906d46d4SStefano Zampini PetscReal error; 5086906d46d4SStefano Zampini 50875e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 5088906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 50895e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 5090e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5091e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5092d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 509388428137SStefano Zampini if (!pcbddc->change_interior) { 509488428137SStefano Zampini const PetscScalar *x,*y,*v; 509588428137SStefano Zampini PetscReal lerror = 0.; 509688428137SStefano Zampini PetscInt i; 509788428137SStefano Zampini 509888428137SStefano Zampini ierr = VecGetArrayRead(matis->x,&x);CHKERRQ(ierr); 509988428137SStefano Zampini ierr = VecGetArrayRead(matis->y,&y);CHKERRQ(ierr); 510088428137SStefano Zampini ierr = VecGetArrayRead(matis->counter,&v);CHKERRQ(ierr); 510188428137SStefano Zampini for (i=0;i<local_size;i++) 510288428137SStefano Zampini if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror) 510388428137SStefano Zampini lerror = PetscAbsScalar(x[i]-y[i]); 510488428137SStefano Zampini ierr = VecRestoreArrayRead(matis->x,&x);CHKERRQ(ierr); 510588428137SStefano Zampini ierr = VecRestoreArrayRead(matis->y,&y);CHKERRQ(ierr); 510688428137SStefano Zampini ierr = VecRestoreArrayRead(matis->counter,&v);CHKERRQ(ierr); 5107820f2d46SBarry Smith ierr = MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 5108637e8532SStefano Zampini if (error > PETSC_SMALL) { 5109637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 51106080607fSStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on I: %1.6e",error); 5111637e8532SStefano Zampini } else { 51126080607fSStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on I: %1.6e",error); 5113637e8532SStefano Zampini } 5114637e8532SStefano Zampini } 511588428137SStefano Zampini } 5116e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5117e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5118906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 5119906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 5120637e8532SStefano Zampini if (error > PETSC_SMALL) { 5121637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 51226080607fSStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e",error); 5123637e8532SStefano Zampini } else { 51246080607fSStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on N: %1.6e",error); 5125637e8532SStefano Zampini } 5126637e8532SStefano Zampini } 5127906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 5128906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 5129906d46d4SStefano Zampini } 5130906d46d4SStefano Zampini 5131022d8d2bSstefano_zampini /* lA is present if we are setting up an inner BDDC for a saddle point FETI-DP */ 5132022d8d2bSstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject*)&lA);CHKERRQ(ierr); 5133022d8d2bSstefano_zampini 513422d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 51357ebab0bbSStefano Zampini ierr = PetscObjectBaseTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 513622d5777bSStefano Zampini if (isseqaij) { 5137a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 5138a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 5139022d8d2bSstefano_zampini if (lA) { 5140022d8d2bSstefano_zampini Mat work; 5141022d8d2bSstefano_zampini ierr = MatPtAP(lA,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 5142022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 5143022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 5144022d8d2bSstefano_zampini } 5145aa0d41d4SStefano Zampini } else { 5146a00504b5SStefano Zampini Mat work_mat; 51471cf9b237SStefano Zampini 5148a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 5149aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 5150a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 51511d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 5152022d8d2bSstefano_zampini if (lA) { 5153022d8d2bSstefano_zampini Mat work; 5154022d8d2bSstefano_zampini ierr = MatConvert(lA,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 5155022d8d2bSstefano_zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 5156022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 5157022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 5158022d8d2bSstefano_zampini } 5159aa0d41d4SStefano Zampini } 51603301b35fSStefano Zampini if (matis->A->symmetric_set) { 51613301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 5162e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 51633301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 5164e496cd5dSStefano Zampini #endif 51653301b35fSStefano Zampini } 5166d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 5167aa0d41d4SStefano Zampini PetscFunctionReturn(0); 5168aa0d41d4SStefano Zampini } 5169aa0d41d4SStefano Zampini 51708ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 5171a64d13efSStefano Zampini { 5172a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 5173a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5174d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 517553892102SStefano Zampini PetscInt *idx_R_local=NULL; 51763a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 51773a50541eSStefano Zampini PetscInt vbs,bs; 51786816873aSStefano Zampini PetscBT bitmask=NULL; 5179a64d13efSStefano Zampini PetscErrorCode ierr; 5180a64d13efSStefano Zampini 5181a64d13efSStefano Zampini PetscFunctionBegin; 5182b23d619eSStefano Zampini /* 5183b23d619eSStefano Zampini No need to setup local scatters if 5184b23d619eSStefano Zampini - primal space is unchanged 5185b23d619eSStefano Zampini AND 5186b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 5187b23d619eSStefano Zampini AND 5188b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 5189b23d619eSStefano Zampini */ 5190b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 5191f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 5192f4ddd8eeSStefano Zampini } 5193f4ddd8eeSStefano Zampini /* destroy old objects */ 5194f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 5195f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 5196f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 5197a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 5198b371cd4fSStefano Zampini n_B = pcis->n_B; 5199b371cd4fSStefano Zampini n_D = pcis->n - n_B; 5200b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 52013a50541eSStefano Zampini 5202a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 52036816873aSStefano Zampini 520453892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 5205b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 5206854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 5207a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 5208a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 52090e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 5210a64d13efSStefano Zampini } 5211a64d13efSStefano Zampini 5212a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 52134641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 52146816873aSStefano Zampini idx_R_local[n_R++] = i; 5215a64d13efSStefano Zampini } 5216a64d13efSStefano Zampini } 5217df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 5218df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 52196816873aSStefano Zampini 5220df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 5221df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 52226816873aSStefano Zampini } 52233a50541eSStefano Zampini 52243a50541eSStefano Zampini /* Block code */ 52253a50541eSStefano Zampini vbs = 1; 52263a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 52273a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 52283a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 52293a50541eSStefano Zampini PetscInt *vary; 5230b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 5231785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 5232580bdb30SBarry Smith ierr = PetscArrayzero(vary,pcis->n/bs);CHKERRQ(ierr); 5233d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 5234d3df7717SStefano 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 */ 52350e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 5236d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 52373a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 52383a50541eSStefano Zampini is_blocked = PETSC_FALSE; 52393a50541eSStefano Zampini break; 52403a50541eSStefano Zampini } 52413a50541eSStefano Zampini } 5242d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 5243d3df7717SStefano Zampini } else { 5244d3df7717SStefano Zampini /* Verify directly the R set */ 5245d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 5246d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 5247d3df7717SStefano Zampini for (j=1; j<bs; j++) { 5248d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 5249d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 5250d3df7717SStefano Zampini break; 5251d3df7717SStefano Zampini } 5252d3df7717SStefano Zampini } 5253d3df7717SStefano Zampini } 5254d3df7717SStefano Zampini } 52553a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 52563a50541eSStefano Zampini vbs = bs; 52573a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 52583a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 52593a50541eSStefano Zampini } 52603a50541eSStefano Zampini } 52613a50541eSStefano Zampini } 52623a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 5263b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 5264df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 526553892102SStefano Zampini 5266df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 5267df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 526853892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 5269df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 527053892102SStefano Zampini } else { 52713a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 527253892102SStefano Zampini } 5273a64d13efSStefano Zampini 5274a64d13efSStefano Zampini /* print some info if requested */ 5275a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 5276a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5277a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 52781575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5279a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 52806080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %D, dirichlet_size = %D, boundary_size = %D\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 52816080607fSStefano 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); 5282a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5283a64d13efSStefano Zampini } 5284a64d13efSStefano Zampini 5285a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 5286b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 52876816873aSStefano Zampini IS is_aux1,is_aux2; 52886816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 52896816873aSStefano Zampini 52903a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 5291854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 5292854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 5293a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 52944641a718SStefano Zampini for (i=0; i<n_D; i++) { 52954641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 52964641a718SStefano Zampini } 5297a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5298a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 52994641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 53004641a718SStefano Zampini aux_array1[j++] = i; 5301a64d13efSStefano Zampini } 5302a64d13efSStefano Zampini } 5303a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 5304a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5305a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 53064641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 53074641a718SStefano Zampini aux_array2[j++] = i; 5308a64d13efSStefano Zampini } 5309a64d13efSStefano Zampini } 5310a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5311a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 53129448b7f1SJunchao Zhang ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 5313a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 5314a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 5315a64d13efSStefano Zampini 53168eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 5317785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 5318a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 53194641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 53204641a718SStefano Zampini aux_array1[j++] = i; 5321a64d13efSStefano Zampini } 5322a64d13efSStefano Zampini } 5323a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 53249448b7f1SJunchao Zhang ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 5325a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 5326a64d13efSStefano Zampini } 53274641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 53283a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 5329d62866d3SStefano Zampini } else { 5330df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 53316816873aSStefano Zampini IS tis; 53326816873aSStefano Zampini PetscInt schur_size; 53336816873aSStefano Zampini 5334df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 53356816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 53369448b7f1SJunchao Zhang ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 53376816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 53386816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 53396816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 53409448b7f1SJunchao Zhang ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 53416816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 5342d62866d3SStefano Zampini } 5343d62866d3SStefano Zampini } 5344a64d13efSStefano Zampini PetscFunctionReturn(0); 5345a64d13efSStefano Zampini } 5346a64d13efSStefano Zampini 53476d9e27e4SStefano Zampini static PetscErrorCode MatNullSpacePropagateAny_Private(Mat A, IS is, Mat B) 534892cccca0SStefano Zampini { 534992cccca0SStefano Zampini MatNullSpace NullSpace; 535092cccca0SStefano Zampini Mat dmat; 535192cccca0SStefano Zampini const Vec *nullvecs; 535292cccca0SStefano Zampini Vec v,v2,*nullvecs2; 53536d9e27e4SStefano Zampini VecScatter sct = NULL; 5354eb06acf8SStefano Zampini PetscContainer c; 5355eb06acf8SStefano Zampini PetscScalar *ddata; 5356295df10fSStefano Zampini PetscInt k,nnsp_size,bsiz,bsiz2,n,N,bs; 535792cccca0SStefano Zampini PetscBool nnsp_has_cnst; 535892cccca0SStefano Zampini PetscErrorCode ierr; 535992cccca0SStefano Zampini 536092cccca0SStefano Zampini PetscFunctionBegin; 53616d9e27e4SStefano Zampini if (!is && !B) { /* MATIS */ 53626d9e27e4SStefano Zampini Mat_IS* matis = (Mat_IS*)A->data; 53636d9e27e4SStefano Zampini 53646d9e27e4SStefano Zampini if (!B) { 53656d9e27e4SStefano Zampini ierr = MatISGetLocalMat(A,&B);CHKERRQ(ierr); 53666d9e27e4SStefano Zampini } 53676d9e27e4SStefano Zampini sct = matis->cctx; 53686d9e27e4SStefano Zampini ierr = PetscObjectReference((PetscObject)sct);CHKERRQ(ierr); 53696d9e27e4SStefano Zampini } else { 537092cccca0SStefano Zampini ierr = MatGetNullSpace(B,&NullSpace);CHKERRQ(ierr); 537192cccca0SStefano Zampini if (!NullSpace) { 537292cccca0SStefano Zampini ierr = MatGetNearNullSpace(B,&NullSpace);CHKERRQ(ierr); 537392cccca0SStefano Zampini } 537492cccca0SStefano Zampini if (NullSpace) PetscFunctionReturn(0); 53756d9e27e4SStefano Zampini } 537692cccca0SStefano Zampini ierr = MatGetNullSpace(A,&NullSpace);CHKERRQ(ierr); 537792cccca0SStefano Zampini if (!NullSpace) { 537892cccca0SStefano Zampini ierr = MatGetNearNullSpace(A,&NullSpace);CHKERRQ(ierr); 537992cccca0SStefano Zampini } 538092cccca0SStefano Zampini if (!NullSpace) PetscFunctionReturn(0); 53816d9e27e4SStefano Zampini 538292cccca0SStefano Zampini ierr = MatCreateVecs(A,&v,NULL);CHKERRQ(ierr); 538392cccca0SStefano Zampini ierr = MatCreateVecs(B,&v2,NULL);CHKERRQ(ierr); 53846d9e27e4SStefano Zampini if (!sct) { 53859448b7f1SJunchao Zhang ierr = VecScatterCreate(v,is,v2,NULL,&sct);CHKERRQ(ierr); 53866d9e27e4SStefano Zampini } 538792cccca0SStefano Zampini ierr = MatNullSpaceGetVecs(NullSpace,&nnsp_has_cnst,&nnsp_size,(const Vec**)&nullvecs);CHKERRQ(ierr); 5388295df10fSStefano Zampini bsiz = bsiz2 = nnsp_size+!!nnsp_has_cnst; 538992cccca0SStefano Zampini ierr = PetscMalloc1(bsiz,&nullvecs2);CHKERRQ(ierr); 539092cccca0SStefano Zampini ierr = VecGetBlockSize(v2,&bs);CHKERRQ(ierr); 539192cccca0SStefano Zampini ierr = VecGetSize(v2,&N);CHKERRQ(ierr); 539292cccca0SStefano Zampini ierr = VecGetLocalSize(v2,&n);CHKERRQ(ierr); 5393eb06acf8SStefano Zampini ierr = PetscMalloc1(n*bsiz,&ddata);CHKERRQ(ierr); 539492cccca0SStefano Zampini for (k=0;k<nnsp_size;k++) { 5395eb06acf8SStefano Zampini ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),bs,n,N,ddata + n*k,&nullvecs2[k]);CHKERRQ(ierr); 539692cccca0SStefano Zampini ierr = VecScatterBegin(sct,nullvecs[k],nullvecs2[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 539792cccca0SStefano Zampini ierr = VecScatterEnd(sct,nullvecs[k],nullvecs2[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 539892cccca0SStefano Zampini } 539992cccca0SStefano Zampini if (nnsp_has_cnst) { 5400eb06acf8SStefano Zampini ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),bs,n,N,ddata + n*nnsp_size,&nullvecs2[nnsp_size]);CHKERRQ(ierr); 540192cccca0SStefano Zampini ierr = VecSet(nullvecs2[nnsp_size],1.0);CHKERRQ(ierr); 540292cccca0SStefano Zampini } 5403295df10fSStefano Zampini ierr = PCBDDCOrthonormalizeVecs(&bsiz2,nullvecs2);CHKERRQ(ierr); 5404295df10fSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)B),PETSC_FALSE,bsiz2,nullvecs2,&NullSpace);CHKERRQ(ierr); 5405295df10fSStefano Zampini 5406eb06acf8SStefano Zampini ierr = MatCreateDense(PetscObjectComm((PetscObject)B),n,PETSC_DECIDE,N,bsiz2,ddata,&dmat);CHKERRQ(ierr); 5407eb06acf8SStefano Zampini ierr = PetscContainerCreate(PetscObjectComm((PetscObject)B),&c);CHKERRQ(ierr); 5408eb06acf8SStefano Zampini ierr = PetscContainerSetPointer(c,ddata);CHKERRQ(ierr); 5409eb06acf8SStefano Zampini ierr = PetscContainerSetUserDestroy(c,PetscContainerUserDestroyDefault);CHKERRQ(ierr); 5410eb06acf8SStefano Zampini ierr = PetscObjectCompose((PetscObject)dmat,"_PBDDC_Null_dmat_arr",(PetscObject)c);CHKERRQ(ierr); 5411eb06acf8SStefano Zampini ierr = PetscContainerDestroy(&c);CHKERRQ(ierr); 541292cccca0SStefano Zampini ierr = PetscObjectCompose((PetscObject)NullSpace,"_PBDDC_Null_dmat",(PetscObject)dmat);CHKERRQ(ierr); 541392cccca0SStefano Zampini ierr = MatDestroy(&dmat);CHKERRQ(ierr); 5414eb06acf8SStefano Zampini 541592cccca0SStefano Zampini for (k=0;k<bsiz;k++) { 541692cccca0SStefano Zampini ierr = VecDestroy(&nullvecs2[k]);CHKERRQ(ierr); 541792cccca0SStefano Zampini } 541892cccca0SStefano Zampini ierr = PetscFree(nullvecs2);CHKERRQ(ierr); 541992cccca0SStefano Zampini ierr = MatSetNearNullSpace(B,NullSpace);CHKERRQ(ierr); 542092cccca0SStefano Zampini ierr = MatNullSpaceDestroy(&NullSpace);CHKERRQ(ierr); 542192cccca0SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 542292cccca0SStefano Zampini ierr = VecDestroy(&v2);CHKERRQ(ierr); 542392cccca0SStefano Zampini ierr = VecScatterDestroy(&sct);CHKERRQ(ierr); 542492cccca0SStefano Zampini PetscFunctionReturn(0); 542592cccca0SStefano Zampini } 5426304d26faSStefano Zampini 5427684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 5428304d26faSStefano Zampini { 5429304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5430304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 5431304d26faSStefano Zampini PC pc_temp; 5432304d26faSStefano Zampini Mat A_RR; 543392cccca0SStefano Zampini MatNullSpace nnsp; 5434f4ddd8eeSStefano Zampini MatReuse reuse; 5435304d26faSStefano Zampini PetscScalar m_one = -1.0; 5436304d26faSStefano Zampini PetscReal value; 543704708bb6SStefano Zampini PetscInt n_D,n_R; 543892cccca0SStefano Zampini PetscBool issbaij,opts; 5439304d26faSStefano Zampini PetscErrorCode ierr; 54400a545947SLisandro Dalcin void (*f)(void) = NULL; 5441312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 5442e604994aSStefano Zampini size_t len; 5443304d26faSStefano Zampini 5444304d26faSStefano Zampini PetscFunctionBegin; 544543371fb9SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_LocalSolvers[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 54466d9e27e4SStefano Zampini /* approximate solver, propagate NearNullSpace if needed */ 54476d9e27e4SStefano Zampini if (!pc->setupcalled && (pcbddc->NullSpace_corr[0] || pcbddc->NullSpace_corr[2])) { 54486d9e27e4SStefano Zampini MatNullSpace gnnsp1,gnnsp2; 54496d9e27e4SStefano Zampini PetscBool lhas,ghas; 54506d9e27e4SStefano Zampini 54516d9e27e4SStefano Zampini ierr = MatGetNearNullSpace(pcbddc->local_mat,&nnsp);CHKERRQ(ierr); 54526d9e27e4SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&gnnsp1);CHKERRQ(ierr); 54536d9e27e4SStefano Zampini ierr = MatGetNullSpace(pc->pmat,&gnnsp2);CHKERRQ(ierr); 54546d9e27e4SStefano Zampini lhas = nnsp ? PETSC_TRUE : PETSC_FALSE; 5455820f2d46SBarry Smith ierr = MPIU_Allreduce(&lhas,&ghas,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 54566d9e27e4SStefano Zampini if (!ghas && (gnnsp1 || gnnsp2)) { 54576d9e27e4SStefano Zampini ierr = MatNullSpacePropagateAny_Private(pc->pmat,NULL,NULL);CHKERRQ(ierr); 54586d9e27e4SStefano Zampini } 54596d9e27e4SStefano Zampini } 54606d9e27e4SStefano Zampini 5461e604994aSStefano Zampini /* compute prefixes */ 5462e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 5463e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 5464e604994aSStefano Zampini if (!pcbddc->current_level) { 5465a126751eSBarry Smith ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,sizeof(dir_prefix));CHKERRQ(ierr); 5466a126751eSBarry Smith ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,sizeof(neu_prefix));CHKERRQ(ierr); 5467a126751eSBarry Smith ierr = PetscStrlcat(dir_prefix,"pc_bddc_dirichlet_",sizeof(dir_prefix));CHKERRQ(ierr); 5468a126751eSBarry Smith ierr = PetscStrlcat(neu_prefix,"pc_bddc_neumann_",sizeof(neu_prefix));CHKERRQ(ierr); 5469e604994aSStefano Zampini } else { 547035529e7bSStefano Zampini ierr = PetscSNPrintf(str_level,sizeof(str_level),"l%d_",(int)(pcbddc->current_level));CHKERRQ(ierr); 5471e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5472e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 5473312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5474312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 5475a126751eSBarry Smith /* Nonstandard use of PetscStrncpy() to only copy a portion of the input string */ 547634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 547734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5478a126751eSBarry Smith ierr = PetscStrlcat(dir_prefix,"pc_bddc_dirichlet_",sizeof(dir_prefix));CHKERRQ(ierr); 5479a126751eSBarry Smith ierr = PetscStrlcat(neu_prefix,"pc_bddc_neumann_",sizeof(neu_prefix));CHKERRQ(ierr); 5480a126751eSBarry Smith ierr = PetscStrlcat(dir_prefix,str_level,sizeof(dir_prefix));CHKERRQ(ierr); 5481a126751eSBarry Smith ierr = PetscStrlcat(neu_prefix,str_level,sizeof(neu_prefix));CHKERRQ(ierr); 5482e604994aSStefano Zampini } 5483e604994aSStefano Zampini 5484304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 5485684f6988SStefano Zampini if (dirichlet) { 5486d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5487450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 54886080607fSStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented"); 5489450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 5490a3df083aSStefano Zampini Mat A_IIn; 5491a3df083aSStefano Zampini 5492a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 5493a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 5494a3df083aSStefano Zampini pcis->A_II = A_IIn; 5495a3df083aSStefano Zampini } 5496450f8f5eSStefano Zampini } 54973301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 5498d1e098c7SStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric);CHKERRQ(ierr); 5499964fefecSStefano Zampini } 5500ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 5501964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 550292cccca0SStefano Zampini opts = PETSC_FALSE; 5503304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 550492cccca0SStefano Zampini opts = PETSC_TRUE; 5505304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 5506304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 5507304d26faSStefano Zampini /* default */ 5508304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 5509e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 55102f37b69bSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->pA_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 5511304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 55129577ea80SStefano Zampini if (issbaij) { 55139577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 55149577ea80SStefano Zampini } else { 5515304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 55169577ea80SStefano Zampini } 5517399ffe99SStefano Zampini ierr = KSPSetErrorIfNotConverged(pcbddc->ksp_D,pc->erroriffailure);CHKERRQ(ierr); 551892cccca0SStefano Zampini } 551992cccca0SStefano Zampini ierr = MatSetOptionsPrefix(pcis->pA_II,((PetscObject)pcbddc->ksp_D)->prefix);CHKERRQ(ierr); 55202f37b69bSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->pA_II);CHKERRQ(ierr); 5521304d26faSStefano Zampini /* Allow user's customization */ 552292cccca0SStefano Zampini if (opts) { 5523304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 552492cccca0SStefano Zampini } 55256d9e27e4SStefano Zampini ierr = MatGetNearNullSpace(pcis->pA_II,&nnsp);CHKERRQ(ierr); 55266d9e27e4SStefano Zampini if (pcbddc->NullSpace_corr[0] && !nnsp) { /* approximate solver, propagate NearNullSpace */ 55276d9e27e4SStefano Zampini ierr = MatNullSpacePropagateAny_Private(pcbddc->local_mat,pcis->is_I_local,pcis->pA_II);CHKERRQ(ierr); 552892cccca0SStefano Zampini } 552992cccca0SStefano Zampini ierr = MatGetNearNullSpace(pcis->pA_II,&nnsp);CHKERRQ(ierr); 553092cccca0SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 5531cd18cfedSStefano Zampini ierr = PetscObjectQueryFunction((PetscObject)pc_temp,"PCSetCoordinates_C",&f);CHKERRQ(ierr); 553292cccca0SStefano Zampini if (f && pcbddc->mat_graph->cloc && !nnsp) { 5533cd18cfedSStefano Zampini PetscReal *coords = pcbddc->mat_graph->coords,*scoords; 5534cd18cfedSStefano Zampini const PetscInt *idxs; 5535cd18cfedSStefano Zampini PetscInt cdim = pcbddc->mat_graph->cdim,nl,i,d; 5536cd18cfedSStefano Zampini 5537cd18cfedSStefano Zampini ierr = ISGetLocalSize(pcis->is_I_local,&nl);CHKERRQ(ierr); 5538cd18cfedSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 5539cd18cfedSStefano Zampini ierr = PetscMalloc1(nl*cdim,&scoords);CHKERRQ(ierr); 5540cd18cfedSStefano Zampini for (i=0;i<nl;i++) { 5541cd18cfedSStefano Zampini for (d=0;d<cdim;d++) { 5542cd18cfedSStefano Zampini scoords[i*cdim+d] = coords[idxs[i]*cdim+d]; 5543cd18cfedSStefano Zampini } 5544cd18cfedSStefano Zampini } 5545cd18cfedSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 5546cd18cfedSStefano Zampini ierr = PCSetCoordinates(pc_temp,cdim,nl,scoords);CHKERRQ(ierr); 5547cd18cfedSStefano Zampini ierr = PetscFree(scoords);CHKERRQ(ierr); 5548cd18cfedSStefano Zampini } 5549b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 5550df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5551d62866d3SStefano Zampini 5552df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 5553d5574798SStefano Zampini } 555492cccca0SStefano Zampini 5555304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 5556304d26faSStefano Zampini if (!n_D) { 5557304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 5558304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 5559304d26faSStefano Zampini } 556093120301SStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 5561304d26faSStefano Zampini /* set ksp_D into pcis data */ 5562304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 556392cccca0SStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 5564304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 5565684f6988SStefano Zampini } 5566304d26faSStefano Zampini 5567304d26faSStefano Zampini /* NEUMANN PROBLEM */ 55680a545947SLisandro Dalcin A_RR = NULL; 5569684f6988SStefano Zampini if (neumann) { 5570d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 557104708bb6SStefano Zampini PetscInt ibs,mbs; 55720aa714b2SStefano Zampini PetscBool issbaij, reuse_neumann_solver; 557304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 55740aa714b2SStefano Zampini 55750aa714b2SStefano Zampini reuse_neumann_solver = PETSC_FALSE; 55760aa714b2SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 55770aa714b2SStefano Zampini IS iP; 55780aa714b2SStefano Zampini 55790aa714b2SStefano Zampini reuse_neumann_solver = PETSC_TRUE; 55800aa714b2SStefano Zampini ierr = PetscObjectQuery((PetscObject)sub_schurs->A,"__KSPFETIDP_iP",(PetscObject*)&iP);CHKERRQ(ierr); 55810aa714b2SStefano Zampini if (iP) reuse_neumann_solver = PETSC_FALSE; 55820aa714b2SStefano Zampini } 5583f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 55848ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 5585f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 5586f4ddd8eeSStefano Zampini PetscInt nn_R; 558781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 5588f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 5589f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 5590f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 5591f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 5592f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5593f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5594f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 5595727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 5596f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5597f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5598f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 5599f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 5600f4ddd8eeSStefano Zampini } 5601f4ddd8eeSStefano Zampini } 5602f4ddd8eeSStefano Zampini /* last check */ 5603d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 5604f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5605f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5606f4ddd8eeSStefano Zampini } 5607f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 5608f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5609f4ddd8eeSStefano Zampini } 5610365a3a41SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection 5611365a3a41SStefano Zampini TODO: Get Rid of these conversions */ 5612af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 5613af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 561404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 561504708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 561604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 561704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 561804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 5619af732b37SStefano Zampini } else { 5620511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 56216816873aSStefano Zampini } 562204708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 562304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 562404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 5625365a3a41SStefano Zampini ierr = MatConvert(matis->A,mbs > 1 ? MATSEQBAIJ : MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 562604708bb6SStefano Zampini } else { 5627365a3a41SStefano Zampini ierr = MatConvert(pcbddc->local_mat,mbs > 1 ? MATSEQBAIJ : MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 562804708bb6SStefano Zampini } 562904708bb6SStefano Zampini } 5630a00504b5SStefano Zampini /* extract A_RR */ 56310aa714b2SStefano Zampini if (reuse_neumann_solver) { 5632a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5633a00504b5SStefano Zampini 5634a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 563516e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5636a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 563716e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 563816e386b8SStefano Zampini } else { 56397dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 5640a00504b5SStefano Zampini } 5641a00504b5SStefano Zampini } else { 5642a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5643a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 5644a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 5645a00504b5SStefano Zampini } 5646a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 56477dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 564816e386b8SStefano Zampini } 56493301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 5650d1e098c7SStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric);CHKERRQ(ierr); 56516816873aSStefano Zampini } 565292cccca0SStefano Zampini opts = PETSC_FALSE; 5653f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 565492cccca0SStefano Zampini opts = PETSC_TRUE; 5655304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 5656304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 5657304d26faSStefano Zampini /* default */ 5658304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 5659e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 5660304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 56619577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 56629577ea80SStefano Zampini if (issbaij) { 56639577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 56649577ea80SStefano Zampini } else { 5665304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 56669577ea80SStefano Zampini } 5667399ffe99SStefano Zampini ierr = KSPSetErrorIfNotConverged(pcbddc->ksp_R,pc->erroriffailure);CHKERRQ(ierr); 566892cccca0SStefano Zampini } 56692f37b69bSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 567092cccca0SStefano Zampini ierr = MatSetOptionsPrefix(A_RR,((PetscObject)pcbddc->ksp_R)->prefix);CHKERRQ(ierr); 567192cccca0SStefano Zampini if (opts) { /* Allow user's customization once */ 5672304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 567392cccca0SStefano Zampini } 56746d9e27e4SStefano Zampini ierr = MatGetNearNullSpace(A_RR,&nnsp);CHKERRQ(ierr); 56756d9e27e4SStefano Zampini if (pcbddc->NullSpace_corr[2] && !nnsp) { /* approximate solver, propagate NearNullSpace */ 56766d9e27e4SStefano Zampini ierr = MatNullSpacePropagateAny_Private(pcbddc->local_mat,pcbddc->is_R_local,A_RR);CHKERRQ(ierr); 567792cccca0SStefano Zampini } 567892cccca0SStefano Zampini ierr = MatGetNearNullSpace(A_RR,&nnsp);CHKERRQ(ierr); 567992cccca0SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 5680cd18cfedSStefano Zampini ierr = PetscObjectQueryFunction((PetscObject)pc_temp,"PCSetCoordinates_C",&f);CHKERRQ(ierr); 568192cccca0SStefano Zampini if (f && pcbddc->mat_graph->cloc && !nnsp) { 5682cd18cfedSStefano Zampini PetscReal *coords = pcbddc->mat_graph->coords,*scoords; 5683cd18cfedSStefano Zampini const PetscInt *idxs; 5684cd18cfedSStefano Zampini PetscInt cdim = pcbddc->mat_graph->cdim,nl,i,d; 5685cd18cfedSStefano Zampini 5686cd18cfedSStefano Zampini ierr = ISGetLocalSize(pcbddc->is_R_local,&nl);CHKERRQ(ierr); 5687cd18cfedSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,&idxs);CHKERRQ(ierr); 5688cd18cfedSStefano Zampini ierr = PetscMalloc1(nl*cdim,&scoords);CHKERRQ(ierr); 5689cd18cfedSStefano Zampini for (i=0;i<nl;i++) { 5690cd18cfedSStefano Zampini for (d=0;d<cdim;d++) { 5691cd18cfedSStefano Zampini scoords[i*cdim+d] = coords[idxs[i]*cdim+d]; 5692cd18cfedSStefano Zampini } 5693cd18cfedSStefano Zampini } 5694cd18cfedSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,&idxs);CHKERRQ(ierr); 5695cd18cfedSStefano Zampini ierr = PCSetCoordinates(pc_temp,cdim,nl,scoords);CHKERRQ(ierr); 5696cd18cfedSStefano Zampini ierr = PetscFree(scoords);CHKERRQ(ierr); 5697cd18cfedSStefano Zampini } 569892cccca0SStefano Zampini 5699304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 5700304d26faSStefano Zampini if (!n_R) { 5701304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 5702304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 5703304d26faSStefano Zampini } 5704df4d28bfSStefano Zampini /* Reuse solver if it is present */ 57050aa714b2SStefano Zampini if (reuse_neumann_solver) { 5706df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5707d62866d3SStefano Zampini 5708df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 5709d62866d3SStefano Zampini } 571093120301SStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 5711684f6988SStefano Zampini } 5712304d26faSStefano Zampini 5713684f6988SStefano Zampini if (pcbddc->dbg_flag) { 5714684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 57151575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5716684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5717684f6988SStefano Zampini } 57188ead10e4SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_LocalSolvers[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 5719c7017625SStefano Zampini 5720c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 5721c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 5722c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 5723c7017625SStefano Zampini } 5724c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 5725c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 5726c7017625SStefano Zampini } 5727c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 5728c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 5729c7017625SStefano Zampini } 5730c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 5731c7017625SStefano Zampini if (pcbddc->dbg_flag) { 5732684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 57330fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 57340fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 57350fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 5736c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_D,pc,pcis->vec2_D);CHKERRQ(ierr); 57370fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 57380fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 5739e604994aSStefano 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); 5740304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5741304d26faSStefano Zampini } 5742684f6988SStefano Zampini if (neumann) { /* Neumann */ 57430fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 57440fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 57450fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 5746c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);CHKERRQ(ierr); 57470fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 57480fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 5749e604994aSStefano 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); 5750304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5751304d26faSStefano Zampini } 5752684f6988SStefano Zampini } 57535cbda25cSStefano Zampini /* free Neumann problem's matrix */ 57545cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5755304d26faSStefano Zampini PetscFunctionReturn(0); 5756304d26faSStefano Zampini } 5757304d26faSStefano Zampini 575880677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 5759674ae819SStefano Zampini { 5760674ae819SStefano Zampini PetscErrorCode ierr; 5761674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 5762be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5763b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE) : PETSC_FALSE; 5764674ae819SStefano Zampini 5765674ae819SStefano Zampini PetscFunctionBegin; 5766b334f244SStefano Zampini if (!reuse_solver) { 576780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 576820c7b377SStefano Zampini } 576980677318SStefano Zampini if (!pcbddc->switch_static) { 577080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 577180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 577280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 577320c7b377SStefano Zampini } 5774b334f244SStefano Zampini if (!reuse_solver) { 577580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 577680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 577720c7b377SStefano Zampini } else { 5778df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5779be83ff47SStefano Zampini 5780df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5781df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 578220c7b377SStefano Zampini } 5783be83ff47SStefano Zampini } else { 578480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 578580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 578680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 578780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 578880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 578980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 579080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 579180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 579280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5793674ae819SStefano Zampini } 5794674ae819SStefano Zampini } 579555c176c0SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][1],pc,0,0,0);CHKERRQ(ierr); 5796b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 579780677318SStefano Zampini if (applytranspose) { 579880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 579980677318SStefano Zampini } else { 580080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 580180677318SStefano Zampini } 5802c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec1_R);CHKERRQ(ierr); 5803be83ff47SStefano Zampini } else { 5804df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5805be83ff47SStefano Zampini 5806be83ff47SStefano Zampini if (applytranspose) { 5807df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 5808be83ff47SStefano Zampini } else { 5809df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 5810be83ff47SStefano Zampini } 5811be83ff47SStefano Zampini } 581255c176c0SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][1],pc,0,0,0);CHKERRQ(ierr); 581380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 581480677318SStefano Zampini if (!pcbddc->switch_static) { 5815b334f244SStefano Zampini if (!reuse_solver) { 581680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 581780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5818be83ff47SStefano Zampini } else { 5819df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5820be83ff47SStefano Zampini 5821df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5822df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5823be83ff47SStefano Zampini } 582480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 582580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 582680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 582780677318SStefano Zampini } 582880677318SStefano Zampini } else { 582980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 583080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 583180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 583280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 583380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 583480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 583580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 583680677318SStefano Zampini } 583780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 583880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 583980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 584080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5841674ae819SStefano Zampini } 5842674ae819SStefano Zampini PetscFunctionReturn(0); 5843674ae819SStefano Zampini } 5844674ae819SStefano Zampini 5845dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 5846dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 5847674ae819SStefano Zampini { 5848674ae819SStefano Zampini PetscErrorCode ierr; 5849674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 5850674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 5851674ae819SStefano Zampini const PetscScalar zero = 0.0; 5852674ae819SStefano Zampini 5853674ae819SStefano Zampini PetscFunctionBegin; 5854dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 58554fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5856dc359a40SStefano Zampini if (applytranspose) { 5857674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 58588eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 5859dc359a40SStefano Zampini } else { 5860674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 5861674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 586215aaf578SStefano Zampini } 58634fee134fSStefano Zampini } else { 58644fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 58654fee134fSStefano Zampini } 5866efc2fbd9SStefano Zampini 5867efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 58684f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5869efc2fbd9SStefano Zampini PetscScalar *array; 58704f1b2e48SStefano Zampini PetscInt j; 5871efc2fbd9SStefano Zampini 5872efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 58734f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 5874efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5875efc2fbd9SStefano Zampini } 5876efc2fbd9SStefano Zampini 587712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 587812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 587912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 588012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 588112edc857SStefano Zampini 58829f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 588312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 588451694757SStefano Zampini Mat coarse_mat; 5885964fefecSStefano Zampini Vec rhs,sol; 588651694757SStefano Zampini MatNullSpace nullsp; 588727b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 5888964fefecSStefano Zampini 588927b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 589027b6a85dSStefano Zampini PC coarse_pc; 589127b6a85dSStefano Zampini 589227b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 589327b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 589427b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 589527b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 589627b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 589727b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 58983bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 589927b6a85dSStefano Zampini } 590027b6a85dSStefano Zampini } 5901964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 5902964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 590351694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 590412edc857SStefano Zampini if (applytranspose) { 59059a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 590655c176c0SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][2],pc,0,0,0);CHKERRQ(ierr); 5907964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 590855c176c0SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][2],pc,0,0,0);CHKERRQ(ierr); 5909c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->coarse_ksp,pc,sol);CHKERRQ(ierr); 59109bfcb8eaSStefano Zampini ierr = MatGetTransposeNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 59119bfcb8eaSStefano Zampini if (nullsp) { 59129bfcb8eaSStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 59139bfcb8eaSStefano Zampini } 59142701bc32SStefano Zampini } else { 59159bfcb8eaSStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 59161f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 59172701bc32SStefano Zampini PC coarse_pc; 59182701bc32SStefano Zampini 59199bfcb8eaSStefano Zampini if (nullsp) { 59209bfcb8eaSStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 59219bfcb8eaSStefano Zampini } 59222701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 59232701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 59243e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 59252701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 592612edc857SStefano Zampini } else { 592755c176c0SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][2],pc,0,0,0);CHKERRQ(ierr); 5928964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 592955c176c0SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][2],pc,0,0,0);CHKERRQ(ierr); 5930c0decd05SBarry Smith ierr = KSPCheckSolve(pcbddc->coarse_ksp,pc,sol);CHKERRQ(ierr); 59319bfcb8eaSStefano Zampini if (nullsp) { 59329bfcb8eaSStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 59339bfcb8eaSStefano Zampini } 593412edc857SStefano Zampini } 59352701bc32SStefano Zampini } 59361d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 593727b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 593827b6a85dSStefano Zampini PC coarse_pc; 593927b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 594027b6a85dSStefano Zampini 594127b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 594227b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 594327b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 59443bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 594527b6a85dSStefano Zampini } 594612edc857SStefano Zampini } 5947674ae819SStefano Zampini 5948674ae819SStefano Zampini /* Local solution on R nodes */ 59494fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 595080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 59519f00e9b4SStefano Zampini } 59529f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 59539f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 595412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5955674ae819SStefano Zampini 59564fee134fSStefano Zampini /* Sum contributions from the two levels */ 59574fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5958dc359a40SStefano Zampini if (applytranspose) { 5959dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5960dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5961dc359a40SStefano Zampini } else { 5962674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 59638eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5964dc359a40SStefano Zampini } 5965efc2fbd9SStefano Zampini /* store p0 */ 59664f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5967efc2fbd9SStefano Zampini PetscScalar *array; 59684f1b2e48SStefano Zampini PetscInt j; 5969efc2fbd9SStefano Zampini 5970efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 59714f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 5972efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5973efc2fbd9SStefano Zampini } 59744fee134fSStefano Zampini } else { /* expand the coarse solution */ 59754fee134fSStefano Zampini if (applytranspose) { 59764fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 59774fee134fSStefano Zampini } else { 59784fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 59794fee134fSStefano Zampini } 59804fee134fSStefano Zampini } 5981674ae819SStefano Zampini PetscFunctionReturn(0); 5982674ae819SStefano Zampini } 5983674ae819SStefano Zampini 598412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 5985674ae819SStefano Zampini { 5986674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 598712edc857SStefano Zampini Vec from,to; 59887ebab0bbSStefano Zampini const PetscScalar *array; 59897ebab0bbSStefano Zampini PetscErrorCode ierr; 5990674ae819SStefano Zampini 5991674ae819SStefano Zampini PetscFunctionBegin; 599212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 599312edc857SStefano Zampini from = pcbddc->coarse_vec; 599412edc857SStefano Zampini to = pcbddc->vec1_P; 599512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 599612edc857SStefano Zampini Vec tvec; 599758da7f69SStefano Zampini 599858da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 599958da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 600012edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 60017ebab0bbSStefano Zampini ierr = VecGetArrayRead(tvec,&array);CHKERRQ(ierr); 600258da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 60037ebab0bbSStefano Zampini ierr = VecRestoreArrayRead(tvec,&array);CHKERRQ(ierr); 600412edc857SStefano Zampini } 600512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 600612edc857SStefano Zampini from = pcbddc->vec1_P; 600712edc857SStefano Zampini to = pcbddc->coarse_vec; 600812edc857SStefano Zampini } 600912edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 6010674ae819SStefano Zampini PetscFunctionReturn(0); 6011674ae819SStefano Zampini } 6012674ae819SStefano Zampini 601312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 6014674ae819SStefano Zampini { 6015674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 601612edc857SStefano Zampini Vec from,to; 60177ebab0bbSStefano Zampini const PetscScalar *array; 60187ebab0bbSStefano Zampini PetscErrorCode ierr; 6019674ae819SStefano Zampini 6020674ae819SStefano Zampini PetscFunctionBegin; 602112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 602212edc857SStefano Zampini from = pcbddc->coarse_vec; 602312edc857SStefano Zampini to = pcbddc->vec1_P; 602412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 602512edc857SStefano Zampini from = pcbddc->vec1_P; 602612edc857SStefano Zampini to = pcbddc->coarse_vec; 602712edc857SStefano Zampini } 602812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 602912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 603012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 603112edc857SStefano Zampini Vec tvec; 603258da7f69SStefano Zampini 603312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 60347ebab0bbSStefano Zampini ierr = VecGetArrayRead(to,&array);CHKERRQ(ierr); 603558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 60367ebab0bbSStefano Zampini ierr = VecRestoreArrayRead(to,&array);CHKERRQ(ierr); 603758da7f69SStefano Zampini } 603858da7f69SStefano Zampini } else { 603958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 604058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 604112edc857SStefano Zampini } 604212edc857SStefano Zampini } 6043674ae819SStefano Zampini PetscFunctionReturn(0); 6044674ae819SStefano Zampini } 6045674ae819SStefano Zampini 6046674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 6047674ae819SStefano Zampini { 6048674ae819SStefano Zampini PetscErrorCode ierr; 6049674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 6050674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6051674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6052984c4197SStefano Zampini /* one and zero */ 6053984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 6054984c4197SStefano Zampini /* space to store constraints and their local indices */ 60559162d606SStefano Zampini PetscScalar *constraints_data; 60569162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 60579162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 60589162d606SStefano Zampini PetscInt *constraints_n; 6059984c4197SStefano Zampini /* iterators */ 6060b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 6061984c4197SStefano Zampini /* BLAS integers */ 6062e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 6063e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 6064c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 6065727cdba6SStefano Zampini /* reuse */ 60660e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 60670e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 6068984c4197SStefano Zampini /* change of basis */ 6069b3d85658SStefano Zampini PetscBool qr_needed; 60709162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 6071984c4197SStefano Zampini /* auxiliary stuff */ 607264efe560SStefano Zampini PetscInt *nnz,*is_indices; 60738a0068c3SStefano Zampini PetscInt ncc; 6074984c4197SStefano Zampini /* some quantities */ 607545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 6076a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 607757715f18SStefano Zampini PetscReal tol; /* tolerance for retaining eigenmodes */ 6078984c4197SStefano Zampini 6079674ae819SStefano Zampini PetscFunctionBegin; 608057715f18SStefano Zampini tol = PetscSqrtReal(PETSC_SMALL); 60818e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 60828e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 60838e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 608416909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 6085088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 6086088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 60870e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 60880e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 6089580bdb30SBarry Smith ierr = PetscArraycpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc);CHKERRQ(ierr); 6090580bdb30SBarry Smith ierr = PetscArraycpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc);CHKERRQ(ierr); 60910e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 6092088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 6093cf5a6209SStefano Zampini 6094cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 60959162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 6096cf5a6209SStefano Zampini MatNullSpace nearnullsp; 6097cf5a6209SStefano Zampini const Vec *nearnullvecs; 6098cf5a6209SStefano Zampini Vec *localnearnullsp; 6099cf5a6209SStefano Zampini PetscScalar *array; 6100cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 6101cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 6102674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 6103b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 6104674ae819SStefano Zampini PetscScalar *work; 6105674ae819SStefano Zampini PetscReal *singular_vals; 6106674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 6107674ae819SStefano Zampini PetscReal *rwork; 6108674ae819SStefano Zampini #endif 610955080a34SStefano Zampini PetscScalar *temp_basis = NULL,*correlation_mat = NULL; 6110964fefecSStefano Zampini PetscBLASInt dummy_int=1; 6111964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 611255080a34SStefano Zampini PetscBool use_pod = PETSC_FALSE; 6113674ae819SStefano Zampini 611455080a34SStefano Zampini /* MKL SVD with same input gives different results on different processes! */ 611555080a34SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) || defined(PETSC_HAVE_MKL) 611655080a34SStefano Zampini use_pod = PETSC_TRUE; 611755080a34SStefano Zampini #endif 6118674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 6119d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 6120e4d548c7SStefano Zampini /* print some info */ 61215c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 6122e4d548c7SStefano Zampini PetscInt nv; 6123e4d548c7SStefano Zampini 6124c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6125e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 6126e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6127e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 61286080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%D)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 61296080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%D)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 61306080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%D)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 6131e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6132e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6133e4d548c7SStefano Zampini } 6134e4d548c7SStefano Zampini 6135d06fc5fdSStefano Zampini /* free unneeded index sets */ 6136d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 6137d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 6138674ae819SStefano Zampini } 6139d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 6140d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 6141d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 6142d06fc5fdSStefano Zampini } 6143d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 6144d06fc5fdSStefano Zampini n_ISForEdges = 0; 6145d06fc5fdSStefano Zampini } 6146d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 6147d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 6148d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 6149d06fc5fdSStefano Zampini } 6150d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 6151d06fc5fdSStefano Zampini n_ISForFaces = 0; 6152d06fc5fdSStefano Zampini } 615370022509SStefano Zampini 6154674ae819SStefano Zampini /* check if near null space is attached to global mat */ 61556d9e27e4SStefano Zampini if (pcbddc->use_nnsp) { 6156674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 61576d9e27e4SStefano Zampini } else nearnullsp = NULL; 61586d9e27e4SStefano Zampini 6159674ae819SStefano Zampini if (nearnullsp) { 6160674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 6161f4ddd8eeSStefano Zampini /* remove any stored info */ 6162f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 6163f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 6164f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 6165f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 6166f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 6167473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 6168f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 6169f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 6170f4ddd8eeSStefano Zampini } 6171984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 6172984c4197SStefano Zampini nnsp_size = 0; 6173674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 6174674ae819SStefano Zampini } 6175984c4197SStefano Zampini /* get max number of constraints on a single cc */ 6176984c4197SStefano Zampini max_constraints = nnsp_size; 6177984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 6178984c4197SStefano Zampini 6179674ae819SStefano Zampini /* 6180674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 61819162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 61829162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 61839162d606SStefano Zampini There can be multiple constraints per connected component 6184674ae819SStefano Zampini */ 6185674ae819SStefano Zampini n_vertices = 0; 6186674ae819SStefano Zampini if (ISForVertices) { 6187674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 6188674ae819SStefano Zampini } 61899162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 61909162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 61919162d606SStefano Zampini 61929162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 61939162d606SStefano Zampini total_counts *= max_constraints; 6194674ae819SStefano Zampini total_counts += n_vertices; 61954641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 61969162d606SStefano Zampini 6197674ae819SStefano Zampini total_counts = 0; 6198674ae819SStefano Zampini max_size_of_constraint = 0; 6199674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 62009162d606SStefano Zampini IS used_is; 6201674ae819SStefano Zampini if (i<n_ISForEdges) { 62029162d606SStefano Zampini used_is = ISForEdges[i]; 6203674ae819SStefano Zampini } else { 62049162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 6205674ae819SStefano Zampini } 62069162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 6207674ae819SStefano Zampini total_counts += j; 6208674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 6209674ae819SStefano Zampini } 62109162d606SStefano 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); 62119162d606SStefano Zampini 6212984c4197SStefano Zampini /* get local part of global near null space vectors */ 6213785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 6214984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 6215984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 6216e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6217e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6218984c4197SStefano Zampini } 6219674ae819SStefano Zampini 6220242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 6221242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 6222a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 6223242a89d7SStefano Zampini 6224984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 6225a773dcb8SStefano Zampini if (!skip_lapack) { 6226674ae819SStefano Zampini PetscScalar temp_work; 6227911cabfeSStefano Zampini 622855080a34SStefano Zampini if (use_pod) { 6229984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 6230785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 6231785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 6232785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 6233674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 6234785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 6235674ae819SStefano Zampini #endif 6236674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 6237c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 6238c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 6239674ae819SStefano Zampini lwork = -1; 6240674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6241674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 6242c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 6243674ae819SStefano Zampini #else 6244c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 6245674ae819SStefano Zampini #endif 6246674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6247984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 624855080a34SStefano Zampini } else { 624955080a34SStefano Zampini #if !defined(PETSC_MISSING_LAPACK_GESVD) 6250674ae819SStefano Zampini /* SVD */ 6251674ae819SStefano Zampini PetscInt max_n,min_n; 6252674ae819SStefano Zampini max_n = max_size_of_constraint; 6253984c4197SStefano Zampini min_n = max_constraints; 6254984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 6255674ae819SStefano Zampini min_n = max_size_of_constraint; 6256984c4197SStefano Zampini max_n = max_constraints; 6257674ae819SStefano Zampini } 6258785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 6259674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 6260785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 6261674ae819SStefano Zampini #endif 6262674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 6263674ae819SStefano Zampini lwork = -1; 6264e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 6265e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 6266b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 6267674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6268674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 62699162d606SStefano 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)); 6270674ae819SStefano Zampini #else 62719162d606SStefano 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)); 6272674ae819SStefano Zampini #endif 6273674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6274984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 627555080a34SStefano Zampini #else 627655080a34SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"This should not happen"); 6277984c4197SStefano Zampini #endif /* on missing GESVD */ 627855080a34SStefano Zampini } 6279674ae819SStefano Zampini /* Allocate optimal workspace */ 6280674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 6281854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 6282674ae819SStefano Zampini } 6283674ae819SStefano Zampini /* Now we can loop on constraining sets */ 6284674ae819SStefano Zampini total_counts = 0; 62859162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 62869162d606SStefano Zampini constraints_data_ptr[0] = 0; 6287674ae819SStefano Zampini /* vertices */ 62889162d606SStefano Zampini if (n_vertices) { 6289674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6290580bdb30SBarry Smith ierr = PetscArraycpy(constraints_idxs,is_indices,n_vertices);CHKERRQ(ierr); 6291674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 62929162d606SStefano Zampini constraints_n[total_counts] = 1; 62939162d606SStefano Zampini constraints_data[total_counts] = 1.0; 62949162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 62959162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 6296674ae819SStefano Zampini total_counts++; 6297674ae819SStefano Zampini } 6298674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6299674ae819SStefano Zampini n_vertices = total_counts; 6300674ae819SStefano Zampini } 6301984c4197SStefano Zampini 6302674ae819SStefano Zampini /* edges and faces */ 63039162d606SStefano Zampini total_counts_cc = total_counts; 6304911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 63059162d606SStefano Zampini IS used_is; 63069162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 63079162d606SStefano Zampini 6308911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 63099162d606SStefano Zampini used_is = ISForEdges[ncc]; 6310984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 6311674ae819SStefano Zampini } else { 63129162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 6313984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 6314674ae819SStefano Zampini } 6315674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 63169162d606SStefano Zampini 63179162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 63189162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6319984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 6320984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 6321674ae819SStefano Zampini if (nnsp_has_cnst) { 63225b08dc53SStefano Zampini PetscScalar quad_value; 63239162d606SStefano Zampini 6324580bdb30SBarry Smith ierr = PetscArraycpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint);CHKERRQ(ierr); 63259162d606SStefano Zampini idxs_copied = PETSC_TRUE; 63269162d606SStefano Zampini 6327a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 6328674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 6329a773dcb8SStefano Zampini } else { 6330a773dcb8SStefano Zampini quad_value = 1.0; 6331a773dcb8SStefano Zampini } 6332674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 63339162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 6334674ae819SStefano Zampini } 63359162d606SStefano Zampini temp_constraints++; 6336674ae819SStefano Zampini total_counts++; 6337674ae819SStefano Zampini } 6338674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 6339984c4197SStefano Zampini PetscReal real_value; 63409162d606SStefano Zampini PetscScalar *ptr_to_data; 63419162d606SStefano Zampini 6342984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 63439162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 6344674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 63459162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 6346674ae819SStefano Zampini } 6347984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 6348984c4197SStefano Zampini /* check if array is null on the connected component */ 6349e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 63509162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 635157715f18SStefano Zampini if (real_value > tol*size_of_constraint) { /* keep indices and values */ 6352674ae819SStefano Zampini temp_constraints++; 6353674ae819SStefano Zampini total_counts++; 63549162d606SStefano Zampini if (!idxs_copied) { 6355580bdb30SBarry Smith ierr = PetscArraycpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint);CHKERRQ(ierr); 63569162d606SStefano Zampini idxs_copied = PETSC_TRUE; 6357674ae819SStefano Zampini } 6358674ae819SStefano Zampini } 63599162d606SStefano Zampini } 63609162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 636145a1bb75SStefano Zampini valid_constraints = temp_constraints; 6362eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 6363a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 63649162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 63659162d606SStefano Zampini 63669162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 6367a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 63689162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 6369a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 63709162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 6371a773dcb8SStefano Zampini } else { /* perform SVD */ 63729162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 6373674ae819SStefano Zampini 637455080a34SStefano Zampini if (use_pod) { 6375984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 6376984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 6377984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 6378984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 6379984c4197SStefano Zampini from that computed using LAPACKgesvd 6380984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 6381984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 6382580bdb30SBarry Smith ierr = PetscArrayzero(correlation_mat,temp_constraints*temp_constraints);CHKERRQ(ierr); 6383674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 6384e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 6385984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6386674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 6387674ae819SStefano Zampini for (k=0;k<j+1;k++) { 63889162d606SStefano 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)); 6389674ae819SStefano Zampini } 6390674ae819SStefano Zampini } 6391e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 6392e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 6393e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 6394674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 6395c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 6396674ae819SStefano Zampini #else 6397c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 6398674ae819SStefano Zampini #endif 6399674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6400984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 6401984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 6402674ae819SStefano Zampini j = 0; 640387b3baaaSStefano Zampini while (j < temp_constraints && singular_vals[j]/singular_vals[temp_constraints-1] < tol) j++; 6404674ae819SStefano Zampini total_counts = total_counts-j; 640545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 6406e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 6407c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 6408c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 6409c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 6410c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6411c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 6412c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 6413674ae819SStefano Zampini if (j<temp_constraints) { 6414984c4197SStefano Zampini PetscInt ii; 6415984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 6416674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 64179162d606SStefano 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)); 6418674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6419984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 6420674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 64219162d606SStefano 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]; 6422674ae819SStefano Zampini } 6423674ae819SStefano Zampini } 6424674ae819SStefano Zampini } 642555080a34SStefano Zampini } else { 642655080a34SStefano Zampini #if !defined(PETSC_MISSING_LAPACK_GESVD) 6427e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 6428e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 6429b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6430674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6431674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 64329162d606SStefano 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)); 6433674ae819SStefano Zampini #else 64349162d606SStefano 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)); 6435674ae819SStefano Zampini #endif 6436984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 6437674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6438984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 6439e310c8b4SStefano Zampini k = temp_constraints; 6440e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 6441674ae819SStefano Zampini j = 0; 644287b3baaaSStefano Zampini while (j < k && singular_vals[k-j-1]/singular_vals[0] < tol) j++; 644345a1bb75SStefano Zampini valid_constraints = k-j; 6444911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 644555080a34SStefano Zampini #else 644655080a34SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"This should not happen"); 6447984c4197SStefano Zampini #endif /* on missing GESVD */ 6448674ae819SStefano Zampini } 6449a773dcb8SStefano Zampini } 645055080a34SStefano Zampini } 64519162d606SStefano Zampini /* update pointers information */ 64529162d606SStefano Zampini if (valid_constraints) { 64539162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 64549162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 64559162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 64569162d606SStefano Zampini /* set change_of_basis flag */ 645745a1bb75SStefano Zampini if (boolforchange) { 6458b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 64599162d606SStefano Zampini } 6460b3d85658SStefano Zampini total_counts_cc++; 646145a1bb75SStefano Zampini } 646245a1bb75SStefano Zampini } 6463984c4197SStefano Zampini /* free workspace */ 64648f1c130eSStefano Zampini if (!skip_lapack) { 6465984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 6466984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 6467984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 6468984c4197SStefano Zampini #endif 6469984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 6470984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 6471984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 6472984c4197SStefano Zampini } 6473984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 6474984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 6475984c4197SStefano Zampini } 6476984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 6477cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 6478cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 6479cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 6480cf5a6209SStefano Zampini } 6481cf5a6209SStefano Zampini if (n_ISForFaces) { 6482cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 6483cf5a6209SStefano Zampini } 6484cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 6485cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 6486cf5a6209SStefano Zampini } 6487cf5a6209SStefano Zampini if (n_ISForEdges) { 6488cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 6489cf5a6209SStefano Zampini } 6490cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 649108122e43SStefano Zampini } else { 649208122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 6493984c4197SStefano Zampini 649408122e43SStefano Zampini total_counts = 0; 649508122e43SStefano Zampini n_vertices = 0; 6496d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 6497d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 649808122e43SStefano Zampini } 649908122e43SStefano Zampini max_constraints = 0; 65009162d606SStefano Zampini total_counts_cc = 0; 650108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 650208122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 65039162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 650408122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 650508122e43SStefano Zampini } 65069162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 65079162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 65089162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 65099162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 651074d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 65119162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 65129162d606SStefano Zampini total_counts_cc = 0; 65139162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 65149162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 65159162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 651608122e43SStefano Zampini } 651708122e43SStefano Zampini } 651808122e43SStefano Zampini 65198bec7fa6SStefano Zampini max_size_of_constraint = 0; 65209162d606SStefano 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]); 65219162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 652208122e43SStefano Zampini /* Change of basis */ 6523b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 652408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 652508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 652608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 6527b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 652808122e43SStefano Zampini } 652908122e43SStefano Zampini } 653008122e43SStefano Zampini } 653108122e43SStefano Zampini } 6532984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 65334f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 653408122e43SStefano Zampini 65359162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 65369162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 65376080607fSStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for constraints indices %D != %D",constraints_idxs_ptr[total_counts_cc],i); 6538674ae819SStefano Zampini 6539674ae819SStefano Zampini /* Create constraint matrix */ 6540674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 654116f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 6542984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 6543984c4197SStefano Zampini 6544984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 6545a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 65465a52fde0SStefano Zampini qr_needed = pcbddc->use_qr_single; 654774d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 6548984c4197SStefano Zampini total_primal_vertices=0; 6549b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 65509162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 65519162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 655272b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 65539162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 6554b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 655564efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 65569162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 65579162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 6558a717540cSStefano Zampini } 6559b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 656091af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 6561a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 6562a717540cSStefano Zampini qr_needed = PETSC_TRUE; 6563a717540cSStefano Zampini } 6564fa434743SStefano Zampini } else { 6565b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 6566fa434743SStefano Zampini } 6567a717540cSStefano Zampini } 6568b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 6569b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 6570674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 657170022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 65724f1b2e48SStefano 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); 6573580bdb30SBarry Smith ierr = PetscArraycpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices);CHKERRQ(ierr); 65740e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 6575984c4197SStefano Zampini 6576984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 657774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 6578785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 6579984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 658074d5cdf7SStefano Zampini 6581984c4197SStefano Zampini j = total_primal_vertices; 658274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 6583b3d85658SStefano Zampini cum = total_primal_vertices; 65849162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 65854641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 6586b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 6587b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 6588b3d85658SStefano Zampini cum++; 65899162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 659074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 659174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 659274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 659374d5cdf7SStefano Zampini } 65949162d606SStefano Zampini j += constraints_n[i]; 6595674ae819SStefano Zampini } 6596674ae819SStefano Zampini } 6597674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 6598e1b21442SStefano Zampini ierr = MatSetOption(pcbddc->ConstraintMatrix,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 6599674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 6600088faed8SStefano Zampini 6601674ae819SStefano Zampini /* set values in constraint matrix */ 6602984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 66030e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6604674ae819SStefano Zampini } 6605984c4197SStefano Zampini total_counts = total_primal_vertices; 66069162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 66074641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 66089162d606SStefano Zampini PetscInt *cols; 66099162d606SStefano Zampini 66109162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 66119162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 66129162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 66139162d606SStefano Zampini PetscInt row = total_counts+k; 66149162d606SStefano Zampini PetscScalar *vals; 66159162d606SStefano Zampini 66169162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 66179162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 66189162d606SStefano Zampini } 66199162d606SStefano Zampini total_counts += constraints_n[i]; 6620674ae819SStefano Zampini } 6621674ae819SStefano Zampini } 6622674ae819SStefano Zampini /* assembling */ 6623674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6624674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 66257ebab0bbSStefano Zampini ierr = MatViewFromOptions(pcbddc->ConstraintMatrix,(PetscObject)pc,"-pc_bddc_constraint_mat_view");CHKERRQ(ierr); 6626088faed8SStefano Zampini 6627674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 6628674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 6629026de310SStefano Zampini /* dual and primal dofs on a single cc */ 6630984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 6631984c4197SStefano Zampini /* working stuff for GEQRF */ 66325a52fde0SStefano Zampini PetscScalar *qr_basis = NULL,*qr_tau = NULL,*qr_work = NULL,lqr_work_t; 6633984c4197SStefano Zampini PetscBLASInt lqr_work; 6634984c4197SStefano Zampini /* working stuff for UNGQR */ 66353c377650SSatish Balay PetscScalar *gqr_work = NULL,lgqr_work_t=0.0; 6636984c4197SStefano Zampini PetscBLASInt lgqr_work; 6637984c4197SStefano Zampini /* working stuff for TRTRS */ 66385a52fde0SStefano Zampini PetscScalar *trs_rhs = NULL; 66393f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 6640984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 6641984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 6642984c4197SStefano Zampini PetscScalar *start_vals; 6643984c4197SStefano Zampini /* working stuff for values insertion */ 66444641a718SStefano Zampini PetscBT is_primal; 664564efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 6646906d46d4SStefano Zampini /* matrix sizes */ 6647906d46d4SStefano Zampini PetscInt global_size,local_size; 6648906d46d4SStefano Zampini /* temporary change of basis */ 6649906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 6650cf5a6209SStefano Zampini /* extra space for debugging */ 66515a52fde0SStefano Zampini PetscScalar *dbg_work = NULL; 6652984c4197SStefano Zampini 6653906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 6654906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 665516f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 6656bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 6657906d46d4SStefano Zampini /* nonzeros for local mat */ 6658bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 66591dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 6660bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 66611dd7afcfSStefano Zampini } else { 66621dd7afcfSStefano Zampini const PetscInt *ii; 66631dd7afcfSStefano Zampini PetscInt n; 66641dd7afcfSStefano Zampini PetscBool flg_row; 66651dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 66661dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 66671dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 66681dd7afcfSStefano Zampini } 66699162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 6670a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 66719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 6672a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 66739162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 6674a717540cSStefano Zampini } else { 66759162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 66769162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 6677a717540cSStefano Zampini } 6678a717540cSStefano Zampini } 6679a717540cSStefano Zampini } 6680906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 6681e1b21442SStefano Zampini ierr = MatSetOption(localChangeOfBasisMatrix,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 6682bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 66831dd7afcfSStefano Zampini /* Set interior change in the matrix */ 66841dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 6685bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 6686906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 6687a717540cSStefano Zampini } 66881dd7afcfSStefano Zampini } else { 66891dd7afcfSStefano Zampini const PetscInt *ii,*jj; 66901dd7afcfSStefano Zampini PetscScalar *aa; 66911dd7afcfSStefano Zampini PetscInt n; 66921dd7afcfSStefano Zampini PetscBool flg_row; 66931dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 66941dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 66951dd7afcfSStefano Zampini for (i=0;i<n;i++) { 66961dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 66971dd7afcfSStefano Zampini } 66981dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 66991dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 67001dd7afcfSStefano Zampini } 6701a717540cSStefano Zampini 6702a717540cSStefano Zampini if (pcbddc->dbg_flag) { 6703a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6704a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6705a717540cSStefano Zampini } 6706a717540cSStefano Zampini 6707a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 6708a717540cSStefano Zampini /* 6709a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 6710a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 6711a717540cSStefano Zampini 6712a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 6713a717540cSStefano Zampini 6714a6b551f4SStefano 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) 6715a6b551f4SStefano Zampini 6716a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 6717a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 6718a717540cSStefano Zampini | ... | 6719a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 6720a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 6721a717540cSStefano Zampini 6722a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 6723a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 6724a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 6725a6b551f4SStefano Zampini 6726a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 6727a717540cSStefano Zampini */ 67285a52fde0SStefano Zampini if (qr_needed && max_size_of_constraint) { 6729984c4197SStefano Zampini /* space to store Q */ 6730854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 67314e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 67324e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 6733984c4197SStefano Zampini /* first we issue queries for optimal work */ 67343f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 67353f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 67363f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6737984c4197SStefano Zampini lqr_work = -1; 67383f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 6739984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 6740984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 6741785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 6742984c4197SStefano Zampini lgqr_work = -1; 67433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 67443f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 67453f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 67463f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 67473f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 6748c964aadfSJose E. Roman PetscStackCallBLAS("LAPACKorgqr",LAPACKorgqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 6749c964aadfSJose E. Roman if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to ORGQR/UNGQR Lapack routine %d",(int)lierr); 6750984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 6751785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 6752984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 6753785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 6754a717540cSStefano Zampini /* allocating workspace for check */ 6755a717540cSStefano Zampini if (pcbddc->dbg_flag) { 6756cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 6757a717540cSStefano Zampini } 6758a717540cSStefano Zampini } 6759984c4197SStefano Zampini /* array to store whether a node is primal or not */ 67604641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 6761473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 67620e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 67636080607fSStefano Zampini if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D",total_primal_vertices,i); 676439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 676539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 676639e2fb2aSStefano Zampini } 676739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 6768984c4197SStefano Zampini 6769a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 67709162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 67719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 67724641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 6773984c4197SStefano Zampini /* get constraint info */ 67749162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 6775984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 6776984c4197SStefano Zampini 6777984c4197SStefano Zampini if (pcbddc->dbg_flag) { 67786080607fSStefano 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); 6779674ae819SStefano Zampini } 6780984c4197SStefano Zampini 6781fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 6782a717540cSStefano Zampini 6783a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 6784a717540cSStefano Zampini if (pcbddc->dbg_flag) { 6785580bdb30SBarry Smith ierr = PetscArraycpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs);CHKERRQ(ierr); 6786a717540cSStefano Zampini } 6787984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 6788580bdb30SBarry Smith ierr = PetscArraycpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs);CHKERRQ(ierr); 6789984c4197SStefano Zampini 6790984c4197SStefano Zampini /* compute QR decomposition of constraints */ 67913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 67923f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 67933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6794674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 67953f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 6796984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 6797674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6798984c4197SStefano Zampini 6799984c4197SStefano Zampini /* explictly compute R^-T */ 6800580bdb30SBarry Smith ierr = PetscArrayzero(trs_rhs,primal_dofs*primal_dofs);CHKERRQ(ierr); 6801984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 68023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 68033f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 68043f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 68053f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 6806984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 68073f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 6808984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 6809984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6810984c4197SStefano Zampini 6811a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2]) overwriting QR factorization in qr_basis */ 68123f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 68133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 68143f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 68153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6816984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6817c964aadfSJose E. Roman PetscStackCallBLAS("LAPACKorgqr",LAPACKorgqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 6818c964aadfSJose E. Roman if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in ORGQR/UNGQR Lapack routine %d",(int)lierr); 6819984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6820984c4197SStefano Zampini 6821984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 6822984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 6823984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 68243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 68253f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 68263f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 68273f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 68283f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 68293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 6830984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 68319162d606SStefano 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)); 6832984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6833580bdb30SBarry Smith ierr = PetscArraycpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs);CHKERRQ(ierr); 6834984c4197SStefano Zampini 6835984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 68369162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 6837984c4197SStefano Zampini /* insert cols for primal dofs */ 6838984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 6839984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 68409162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6841906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6842984c4197SStefano Zampini } 6843984c4197SStefano Zampini /* insert cols for dual dofs */ 6844984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 68459162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 6846984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 68479162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6848906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6849984c4197SStefano Zampini j++; 6850674ae819SStefano Zampini } 6851674ae819SStefano Zampini } 6852984c4197SStefano Zampini 6853984c4197SStefano Zampini /* check change of basis */ 6854984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6855984c4197SStefano Zampini PetscInt ii,jj; 6856984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 6857c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 6858c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 6859c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 6860c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6861c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 6862c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 6863984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6864cf5a6209SStefano 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)); 6865984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6866984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6867984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6868cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 6869c068d9bbSLisandro Dalcin if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-(PetscReal)1) > 1.e-12) valid_qr = PETSC_FALSE; 6870674ae819SStefano Zampini } 6871674ae819SStefano Zampini } 6872984c4197SStefano Zampini if (!valid_qr) { 687322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 6874984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6875984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6876cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 68776080607fSStefano Zampini ierr = 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]));CHKERRQ(ierr); 6878674ae819SStefano Zampini } 6879c068d9bbSLisandro Dalcin if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-(PetscReal)1) > 1.e-12) { 68806080607fSStefano Zampini ierr = 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]));CHKERRQ(ierr); 6881984c4197SStefano Zampini } 6882984c4197SStefano Zampini } 6883984c4197SStefano Zampini } 6884674ae819SStefano Zampini } else { 688522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 6886674ae819SStefano Zampini } 6887674ae819SStefano Zampini } 6888a717540cSStefano Zampini } else { /* simple transformation block */ 6889a717540cSStefano Zampini PetscInt row,col; 6890a6b551f4SStefano Zampini PetscScalar val,norm; 6891a6b551f4SStefano Zampini 6892a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 68939162d606SStefano 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)); 6894a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 68959162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 68969162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6897bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 68989162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 6899906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 69009162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 6901a717540cSStefano Zampini } else { 6902a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 69039162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6904a717540cSStefano Zampini if (row != col) { 69059162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 6906a717540cSStefano Zampini } else { 69079162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 6908a717540cSStefano Zampini } 6909906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 6910a717540cSStefano Zampini } 6911a717540cSStefano Zampini } 6912a717540cSStefano Zampini } 691398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 691422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 6915a717540cSStefano Zampini } 6916674ae819SStefano Zampini } 6917984c4197SStefano Zampini } else { 6918984c4197SStefano Zampini if (pcbddc->dbg_flag) { 69196080607fSStefano 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); 6920674ae819SStefano Zampini } 6921674ae819SStefano Zampini } 6922674ae819SStefano Zampini } 6923a717540cSStefano Zampini 6924a717540cSStefano Zampini /* free workspace */ 6925a717540cSStefano Zampini if (qr_needed) { 6926984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6927cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 6928984c4197SStefano Zampini } 6929984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 6930984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 6931984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 6932984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 6933984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 6934674ae819SStefano Zampini } 6935a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 6936906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6937906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6938906d46d4SStefano Zampini 6939906d46d4SStefano Zampini /* assembling of global change of variable */ 694088c03ad3SStefano Zampini if (!pcbddc->fake_change) { 6941bbb9e6c6SStefano Zampini Mat tmat; 694216f15bc4SStefano Zampini PetscInt bs; 694316f15bc4SStefano Zampini 6944906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 6945906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 6946bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 6947bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 6948487b449aSStefano Zampini ierr = MatAssemblyBegin(tmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6949487b449aSStefano Zampini ierr = MatAssemblyEnd(tmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6950bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6951bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 695216f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 695316f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 6954906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 6955bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 6956487b449aSStefano Zampini ierr = MatConvert(tmat,MATAIJ,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6957bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6958bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6959bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6960e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6961e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6962bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 6963bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 696488c03ad3SStefano Zampini 6965906d46d4SStefano Zampini /* check */ 6966906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 6967906d46d4SStefano Zampini PetscReal error; 6968906d46d4SStefano Zampini Vec x,x_change; 6969906d46d4SStefano Zampini 6970906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 6971906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 6972906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 6973906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 6974e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6975e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6976bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 6977e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6978e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6979906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 6980906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 6981906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 6982637e8532SStefano Zampini if (error > PETSC_SMALL) { 69836080607fSStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e",error); 6984637e8532SStefano Zampini } 6985906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 6986906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 6987906d46d4SStefano Zampini } 6988b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 6989b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 6990b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6991bf3a8328SStefano Zampini 699213903a91SSatish Balay 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"); 6993b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 6994ac632422SStefano Zampini Mat S_new,tmat; 6995bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 6996bbb9e6c6SStefano Zampini 6997bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 69987dae84e0SHong Zhang ierr = MatCreateSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 6999bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 7000bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 7001bf3a8328SStefano Zampini IS is_V; 7002b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 7003b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 7004b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 7005b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 7006b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 7007bf3a8328SStefano Zampini } 7008bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 7009ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 7010b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 7011ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 7012bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 7013bf3a8328SStefano Zampini const PetscScalar *array; 7014bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 7015bf3a8328SStefano Zampini PetscInt i,n_V; 7016bf3a8328SStefano Zampini 7017b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 7018b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 7019b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 7020b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 7021b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 7022b087196eSStefano Zampini for (i=0;i<n_V;i++) { 7023b087196eSStefano Zampini PetscScalar val; 7024b087196eSStefano Zampini PetscInt idx; 7025b087196eSStefano Zampini 7026b087196eSStefano Zampini idx = idxs_V[i]; 7027b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 7028b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 7029b087196eSStefano Zampini } 7030b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7031b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7032bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 7033bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 7034bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 7035bf3a8328SStefano Zampini } 7036ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 7037ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 7038ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 7039ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 7040b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 7041ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 7042bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 7043b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 7044bf3a8328SStefano Zampini } 7045ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 7046ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 7047ac632422SStefano Zampini } 7048b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 704988c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 7050b96c3477SStefano Zampini } 7051c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 7052b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 7053c9db6a07SStefano Zampini PetscInt i; 7054c9db6a07SStefano Zampini 7055c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7056c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 7057c9db6a07SStefano Zampini } 7058c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 7059c9db6a07SStefano Zampini } 7060b96c3477SStefano Zampini } 706116909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 706216909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 706316909a7fSStefano Zampini } else { 7064906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 706516909a7fSStefano Zampini } 70661dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 706727b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 706872b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 706972b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 707072b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 707172b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 707272b8c272SStefano Zampini } 70731dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 707427b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 7075b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 7076b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 7077906d46d4SStefano Zampini } else { 70781dd7afcfSStefano Zampini Mat benign_global = NULL; 707927b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 70801dd7afcfSStefano Zampini Mat M; 70811dd7afcfSStefano Zampini 70829e9b7b1fSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 70839e9b7b1fSStefano Zampini ierr = VecCopy(matis->counter,pcis->vec1_N);CHKERRQ(ierr); 70849e9b7b1fSStefano Zampini ierr = VecReciprocal(pcis->vec1_N);CHKERRQ(ierr); 70859e9b7b1fSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&benign_global);CHKERRQ(ierr); 70869e9b7b1fSStefano Zampini if (pcbddc->benign_change) { 70871dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 70881dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 7089906d46d4SStefano Zampini } else { 70909e9b7b1fSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&M);CHKERRQ(ierr); 70919e9b7b1fSStefano Zampini ierr = MatDiagonalSet(M,pcis->vec1_N,INSERT_VALUES);CHKERRQ(ierr); 7092906d46d4SStefano Zampini } 70939e9b7b1fSStefano Zampini ierr = MatISSetLocalMat(benign_global,M);CHKERRQ(ierr); 70949e9b7b1fSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 70959e9b7b1fSStefano Zampini ierr = MatAssemblyBegin(benign_global,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 70969e9b7b1fSStefano Zampini ierr = MatAssemblyEnd(benign_global,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 70971dd7afcfSStefano Zampini } 70981dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 70991dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 71001dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 710127b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 71021dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 71031dd7afcfSStefano Zampini } 71041dd7afcfSStefano Zampini } 710516909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 710616909a7fSStefano Zampini IS is_global; 710716909a7fSStefano Zampini const PetscInt *gidxs; 710816909a7fSStefano Zampini 710916909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 711016909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 711116909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 71127dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 711316909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 711416909a7fSStefano Zampini } 71151dd7afcfSStefano Zampini } 71161dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 71171dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 7118b9b85e73SStefano Zampini } 7119a717540cSStefano Zampini 712072b8c272SStefano Zampini if (!pcbddc->fake_change) { 71214f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 71224f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 71234f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 71244f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 7125019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 7126019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 7127019a44ceSStefano Zampini pcbddc->local_primal_size++; 7128019a44ceSStefano Zampini } 7129019a44ceSStefano Zampini 7130019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 7131727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 7132727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 7133580bdb30SBarry Smith ierr = PetscArraycmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 7134c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 71350e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 7136580bdb30SBarry Smith ierr = PetscArraycmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 7137727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 7138727cdba6SStefano Zampini } 71390e6343abSStefano Zampini } 7140727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 7141820f2d46SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 714272b8c272SStefano Zampini } 714372b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 7144727cdba6SStefano Zampini 7145a717540cSStefano Zampini /* flush dbg viewer */ 7146b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 7147b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7148b8ffe317SStefano Zampini } 7149a717540cSStefano Zampini 7150e310c8b4SStefano Zampini /* free workspace */ 7151a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 71524641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 715308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 71549162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 71559162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 715608122e43SStefano Zampini } else { 71579162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 71589162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 71599162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 716008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 716108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 71629162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 71639162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 716408122e43SStefano Zampini } 7165674ae819SStefano Zampini PetscFunctionReturn(0); 7166674ae819SStefano Zampini } 7167674ae819SStefano Zampini 7168674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 7169674ae819SStefano Zampini { 717071582508SStefano Zampini ISLocalToGlobalMapping map; 7171674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 7172674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 717366da6bd7Sstefano_zampini PetscInt i,N; 717466da6bd7Sstefano_zampini PetscBool rcsr = PETSC_FALSE; 717566da6bd7Sstefano_zampini PetscErrorCode ierr; 7176674ae819SStefano Zampini 7177674ae819SStefano Zampini PetscFunctionBegin; 71788af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 7179b03ebc13SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 71808e61c736SStefano Zampini /* Reset previously computed graph */ 71818e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 7182674ae819SStefano Zampini /* Init local Graph struct */ 71837fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 718471582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 7185be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 7186674ae819SStefano Zampini 71877a0e7b2cSstefano_zampini if (pcbddc->user_primal_vertices_local && !pcbddc->user_primal_vertices) { 71887a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 71897a0e7b2cSstefano_zampini } 7190575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 71916080607fSStefano 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",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs); 71929577ea80SStefano Zampini 7193674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 719466da6bd7Sstefano_zampini if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) { 71954d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 71964d379d7bSStefano Zampini PetscInt nvtxs; 7197e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 7198674ae819SStefano Zampini 71992fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 72002fffb893SStefano Zampini if (flg_row) { 72014d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 7202b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 72032fffb893SStefano Zampini } 72042fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 720566da6bd7Sstefano_zampini rcsr = PETSC_TRUE; 7206674ae819SStefano Zampini } 72079b28b941SStefano Zampini if (pcbddc->dbg_flag) { 72089b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7209674ae819SStefano Zampini } 7210674ae819SStefano Zampini 7211ab8c8b98SStefano Zampini if (pcbddc->mat_graph->cdim && !pcbddc->mat_graph->cloc) { 7212ab8c8b98SStefano Zampini PetscReal *lcoords; 7213ab8c8b98SStefano Zampini PetscInt n; 7214ab8c8b98SStefano Zampini MPI_Datatype dimrealtype; 7215ab8c8b98SStefano Zampini 72164f819b78SStefano Zampini /* TODO: support for blocked */ 7217ab8c8b98SStefano Zampini if (pcbddc->mat_graph->cnloc != pc->pmat->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid number of local coordinates! Got %D, expected %D",pcbddc->mat_graph->cnloc,pc->pmat->rmap->n); 7218ab8c8b98SStefano Zampini ierr = MatGetLocalSize(matis->A,&n,NULL);CHKERRQ(ierr); 7219ab8c8b98SStefano Zampini ierr = PetscMalloc1(pcbddc->mat_graph->cdim*n,&lcoords);CHKERRQ(ierr); 7220ffc4695bSBarry Smith ierr = MPI_Type_contiguous(pcbddc->mat_graph->cdim,MPIU_REAL,&dimrealtype);CHKERRMPI(ierr); 7221ffc4695bSBarry Smith ierr = MPI_Type_commit(&dimrealtype);CHKERRMPI(ierr); 7222ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(matis->sf,dimrealtype,pcbddc->mat_graph->coords,lcoords,MPI_REPLACE);CHKERRQ(ierr); 7223ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(matis->sf,dimrealtype,pcbddc->mat_graph->coords,lcoords,MPI_REPLACE);CHKERRQ(ierr); 7224ffc4695bSBarry Smith ierr = MPI_Type_free(&dimrealtype);CHKERRMPI(ierr); 7225ab8c8b98SStefano Zampini ierr = PetscFree(pcbddc->mat_graph->coords);CHKERRQ(ierr); 7226ab8c8b98SStefano Zampini 7227ab8c8b98SStefano Zampini pcbddc->mat_graph->coords = lcoords; 7228ab8c8b98SStefano Zampini pcbddc->mat_graph->cloc = PETSC_TRUE; 7229ab8c8b98SStefano Zampini pcbddc->mat_graph->cnloc = n; 7230ab8c8b98SStefano Zampini } 7231ab8c8b98SStefano Zampini if (pcbddc->mat_graph->cnloc && pcbddc->mat_graph->cnloc != pcbddc->mat_graph->nvtxs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid number of local subdomain coordinates! Got %D, expected %D",pcbddc->mat_graph->cnloc,pcbddc->mat_graph->nvtxs); 7232625961bdSStefano Zampini pcbddc->mat_graph->active_coords = (PetscBool)(pcbddc->corner_selection && pcbddc->mat_graph->cdim && !pcbddc->corner_selected); 7233ab8c8b98SStefano Zampini 7234674ae819SStefano Zampini /* Setup of Graph */ 72354b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 723614f95afaSStefano 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); 7237674ae819SStefano Zampini 72384f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 72394f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 724020c3699dSStefano Zampini PetscInt *local_subs,n,totn; 72414f1b2e48SStefano Zampini 724220c3699dSStefano Zampini ierr = MatGetLocalSize(matis->A,&n,NULL);CHKERRQ(ierr); 724320c3699dSStefano Zampini ierr = PetscMalloc1(n,&local_subs);CHKERRQ(ierr); 724420c3699dSStefano Zampini for (i=0;i<n;i++) local_subs[i] = pcbddc->n_local_subs; 72454f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 72464f1b2e48SStefano Zampini const PetscInt *idxs; 72474f1b2e48SStefano Zampini PetscInt nl,j; 72484f1b2e48SStefano Zampini 72494f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 72504f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 725171582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 72524f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 72534f1b2e48SStefano Zampini } 725420c3699dSStefano Zampini for (i=0,totn=0;i<n;i++) totn = PetscMax(totn,local_subs[i]); 725520c3699dSStefano Zampini pcbddc->mat_graph->n_local_subs = totn + 1; 72564f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 72574f1b2e48SStefano Zampini } 72588af8fcf9SStefano Zampini } 72594f1b2e48SStefano Zampini 7260cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 7261674ae819SStefano Zampini /* Graph's connected components analysis */ 7262674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 726371582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 72644f819b78SStefano Zampini pcbddc->corner_selected = pcbddc->corner_selection; 72658af8fcf9SStefano Zampini } 726666da6bd7Sstefano_zampini if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0; 7267674ae819SStefano Zampini PetscFunctionReturn(0); 7268674ae819SStefano Zampini } 7269674ae819SStefano Zampini 7270295df10fSStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt *nio, Vec vecs[]) 72719a7d3425SStefano Zampini { 7272295df10fSStefano Zampini PetscInt i,j,n; 72739a7d3425SStefano Zampini PetscScalar *alphas; 7274295df10fSStefano Zampini PetscReal norm,*onorms; 72759a7d3425SStefano Zampini PetscErrorCode ierr; 72769a7d3425SStefano Zampini 72779a7d3425SStefano Zampini PetscFunctionBegin; 7278295df10fSStefano Zampini n = *nio; 72798c0031efSStefano Zampini if (!n) PetscFunctionReturn(0); 7280295df10fSStefano Zampini ierr = PetscMalloc2(n,&alphas,n,&onorms);CHKERRQ(ierr); 728192cccca0SStefano Zampini ierr = VecNormalize(vecs[0],&norm);CHKERRQ(ierr); 728292cccca0SStefano Zampini if (norm < PETSC_SMALL) { 7283295df10fSStefano Zampini onorms[0] = 0.0; 728492cccca0SStefano Zampini ierr = VecSet(vecs[0],0.0);CHKERRQ(ierr); 7285295df10fSStefano Zampini } else { 7286295df10fSStefano Zampini onorms[0] = norm; 728792cccca0SStefano Zampini } 7288295df10fSStefano Zampini 72898c0031efSStefano Zampini for (i=1;i<n;i++) { 72908c0031efSStefano Zampini ierr = VecMDot(vecs[i],i,vecs,alphas);CHKERRQ(ierr); 72918c0031efSStefano Zampini for (j=0;j<i;j++) alphas[j] = PetscConj(-alphas[j]); 72928c0031efSStefano Zampini ierr = VecMAXPY(vecs[i],i,alphas,vecs);CHKERRQ(ierr); 729392cccca0SStefano Zampini ierr = VecNormalize(vecs[i],&norm);CHKERRQ(ierr); 729492cccca0SStefano Zampini if (norm < PETSC_SMALL) { 7295295df10fSStefano Zampini onorms[i] = 0.0; 729692cccca0SStefano Zampini ierr = VecSet(vecs[i],0.0);CHKERRQ(ierr); 7297295df10fSStefano Zampini } else { 7298295df10fSStefano Zampini onorms[i] = norm; 729992cccca0SStefano Zampini } 73009a7d3425SStefano Zampini } 7301295df10fSStefano Zampini /* push nonzero vectors at the beginning */ 7302295df10fSStefano Zampini for (i=0;i<n;i++) { 7303295df10fSStefano Zampini if (onorms[i] == 0.0) { 7304295df10fSStefano Zampini for (j=i+1;j<n;j++) { 7305295df10fSStefano Zampini if (onorms[j] != 0.0) { 7306295df10fSStefano Zampini ierr = VecCopy(vecs[j],vecs[i]);CHKERRQ(ierr); 7307295df10fSStefano Zampini onorms[j] = 0.0; 7308295df10fSStefano Zampini } 7309295df10fSStefano Zampini } 7310295df10fSStefano Zampini } 7311295df10fSStefano Zampini } 7312295df10fSStefano Zampini for (i=0,*nio=0;i<n;i++) *nio += onorms[i] != 0.0 ? 1 : 0; 7313295df10fSStefano Zampini ierr = PetscFree2(alphas,onorms);CHKERRQ(ierr); 73149a7d3425SStefano Zampini PetscFunctionReturn(0); 73159a7d3425SStefano Zampini } 73169a7d3425SStefano Zampini 7317bb360cb4SStefano Zampini PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 7318e7931f94SStefano Zampini { 731957de7509SStefano Zampini Mat A; 7320e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 7321e7931f94SStefano Zampini PetscMPIInt size,rank,color; 732252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 732352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 7324bb360cb4SStefano Zampini PetscInt im_active,active_procs,N,n,i,j,threshold = 2; 732557de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 732627b6a85dSStefano Zampini PetscInt xadj_count,*count; 732727b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 732827b6a85dSStefano Zampini PetscSubcomm psubcomm; 732927b6a85dSStefano Zampini MPI_Comm subcomm; 733052e5ac9dSStefano Zampini PetscErrorCode ierr; 7331a57a6d2fSStefano Zampini 7332e7931f94SStefano Zampini PetscFunctionBegin; 733357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 733457de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 7335fbfcfee5SBarry Smith if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME); 733657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 733757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 73386080607fSStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %D",*n_subdomains); 733957de7509SStefano Zampini 734057de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 7341ffc4695bSBarry Smith ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRMPI(ierr); 7342ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRMPI(ierr); 734357de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 734457de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 7345bb360cb4SStefano Zampini im_active = !!n; 7346820f2d46SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRMPI(ierr); 734757de7509SStefano Zampini void_procs = size - active_procs; 734857de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 734957de7509SStefano Zampini if (void_procs) { 735057de7509SStefano Zampini PetscInt ncand; 735157de7509SStefano Zampini 735257de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 735357de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 7354ffc4695bSBarry Smith ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRMPI(ierr); 735557de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 735657de7509SStefano Zampini if (!procs_candidates[i]) { 735757de7509SStefano Zampini procs_candidates[ncand++] = i; 735857de7509SStefano Zampini } 735957de7509SStefano Zampini } 736057de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 736157de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 736257de7509SStefano Zampini } 736357de7509SStefano Zampini 7364bb360cb4SStefano Zampini /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix 73659dddd249SSatish Balay number of subdomains requested 1 -> send to rank-0 or first candidate in voids */ 7366bb360cb4SStefano Zampini ierr = MatGetSize(mat,&N,NULL);CHKERRQ(ierr); 7367bb360cb4SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) { 736814f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 736914f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 737014f0bfb9SStefano Zampini else dest = rank; 737157de7509SStefano Zampini if (im_active) { 737257de7509SStefano Zampini issize = 1; 737357de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 737414f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 737557de7509SStefano Zampini } else { 737614f0bfb9SStefano Zampini isidx = dest; 737757de7509SStefano Zampini } 737857de7509SStefano Zampini } else { 737957de7509SStefano Zampini issize = 0; 738057de7509SStefano Zampini isidx = -1; 738157de7509SStefano Zampini } 7382bb360cb4SStefano Zampini if (*n_subdomains != 1) *n_subdomains = active_procs; 738357de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 7384daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 738557de7509SStefano Zampini PetscFunctionReturn(0); 738657de7509SStefano Zampini } 7387c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 7388c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 738927b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 7390e7931f94SStefano Zampini 7391e7931f94SStefano Zampini /* Get info on mapping */ 73923bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 7393e7931f94SStefano Zampini 7394e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 7395785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 7396e7931f94SStefano Zampini xadj[0] = 0; 7397e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 7398785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 7399785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 7400bb360cb4SStefano Zampini ierr = PetscCalloc1(n,&count);CHKERRQ(ierr); 740127b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 740227b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 740327b6a85dSStefano Zampini count[shared[i][j]] += 1; 7404e7931f94SStefano Zampini 740527b6a85dSStefano Zampini xadj_count = 0; 74062b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 740727b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 740827b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 7409d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 7410d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 7411d023bfaeSStefano Zampini xadj_count++; 741227b6a85dSStefano Zampini break; 741327b6a85dSStefano Zampini } 7414e7931f94SStefano Zampini } 7415e7931f94SStefano Zampini } 7416d023bfaeSStefano Zampini xadj[1] = xadj_count; 741727b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 74183bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 7419e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 7420e7931f94SStefano Zampini 74213837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 7422e7931f94SStefano Zampini 742327b6a85dSStefano Zampini /* Restrict work on active processes only */ 742427b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 742527b6a85dSStefano Zampini if (void_procs) { 742627b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 742727b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 742827b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 742927b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 743027b6a85dSStefano Zampini } else { 743127b6a85dSStefano Zampini psubcomm = NULL; 743227b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 743327b6a85dSStefano Zampini } 743427b6a85dSStefano Zampini 743527b6a85dSStefano Zampini v_wgt = NULL; 743627b6a85dSStefano Zampini if (!color) { 7437e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 7438e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 7439e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 7440c8587f34SStefano Zampini } else { 744152e5ac9dSStefano Zampini Mat subdomain_adj; 744252e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 744352e5ac9dSStefano Zampini MatPartitioning partitioner; 744427b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 744552e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 744657de7509SStefano Zampini PetscMPIInt size; 7447b0c7d250SStefano Zampini PetscBool aggregate; 7448b0c7d250SStefano Zampini 7449ffc4695bSBarry Smith ierr = MPI_Comm_size(subcomm,&size);CHKERRMPI(ierr); 745027b6a85dSStefano Zampini if (void_procs) { 745127b6a85dSStefano Zampini PetscInt prank = rank; 7452785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 7453ffc4695bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRMPI(ierr); 7454e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 7455e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 7456c8587f34SStefano Zampini } 7457e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 745827b6a85dSStefano Zampini } else { 745927b6a85dSStefano Zampini oldranks = NULL; 746027b6a85dSStefano Zampini } 7461b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 746227b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 7463b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 7464b0c7d250SStefano Zampini PetscMPIInt nrank; 7465b0c7d250SStefano Zampini PetscScalar *vals; 7466b0c7d250SStefano Zampini 7467ffc4695bSBarry Smith ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRMPI(ierr); 7468b0c7d250SStefano Zampini lrows = 0; 7469b0c7d250SStefano Zampini if (nrank<redprocs) { 7470b0c7d250SStefano Zampini lrows = size/redprocs; 7471b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 7472b0c7d250SStefano Zampini } 747327b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 7474b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 7475b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7476b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7477b0c7d250SStefano Zampini row = nrank; 7478b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 7479b0c7d250SStefano Zampini cols = adjncy; 7480b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 7481b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 7482b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 7483b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7484b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 748552e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 748652e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 748752e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 7488b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 748927b6a85dSStefano Zampini if (use_vwgt) { 749027b6a85dSStefano Zampini Vec v; 749127b6a85dSStefano Zampini const PetscScalar *array; 749227b6a85dSStefano Zampini PetscInt nl; 749327b6a85dSStefano Zampini 749427b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 7495bb360cb4SStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)n,INSERT_VALUES);CHKERRQ(ierr); 749627b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 749727b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 749827b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 749927b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 750027b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 750122db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 750227b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 750327b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 750427b6a85dSStefano Zampini } 7505b0c7d250SStefano Zampini } else { 750627b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 750727b6a85dSStefano Zampini if (use_vwgt) { 750827b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 7509bb360cb4SStefano Zampini v_wgt[0] = n; 751027b6a85dSStefano Zampini } 7511b0c7d250SStefano Zampini } 751222b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 7513e7931f94SStefano Zampini 7514e7931f94SStefano Zampini /* Partition */ 751527b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 7516ce64c636SStefano Zampini #if defined(PETSC_HAVE_PTSCOTCH) 7517ce64c636SStefano Zampini ierr = MatPartitioningSetType(partitioner,MATPARTITIONINGPTSCOTCH);CHKERRQ(ierr); 7518ce64c636SStefano Zampini #elif defined(PETSC_HAVE_PARMETIS) 7519ce64c636SStefano Zampini ierr = MatPartitioningSetType(partitioner,MATPARTITIONINGPARMETIS);CHKERRQ(ierr); 7520ce64c636SStefano Zampini #else 7521ce64c636SStefano Zampini ierr = MatPartitioningSetType(partitioner,MATPARTITIONINGAVERAGE);CHKERRQ(ierr); 7522ce64c636SStefano Zampini #endif 7523e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 752427b6a85dSStefano Zampini if (v_wgt) { 7525e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 7526c8587f34SStefano Zampini } 752757de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 752857de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 7529e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 7530e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 753122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 7532e7931f94SStefano Zampini 753352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 75346583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 753552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 753652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 753757de7509SStefano Zampini if (!aggregate) { 753857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 7539cf9c20a2SJed Brown if (PetscUnlikelyDebug(!oldranks)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 754057de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 754127b6a85dSStefano Zampini } else if (oldranks) { 7542b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 754327b6a85dSStefano Zampini } else { 754427b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 754557de7509SStefano Zampini } 754628143c3dSStefano Zampini } else { 75477fb8a5e4SKarl Rupp PetscInt idx = 0; 7548b0c7d250SStefano Zampini PetscMPIInt tag; 7549b0c7d250SStefano Zampini MPI_Request *reqs; 7550b0c7d250SStefano Zampini 7551b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 7552b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 7553b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 7554ffc4695bSBarry Smith ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRMPI(ierr); 755528143c3dSStefano Zampini } 7556ffc4695bSBarry Smith ierr = MPI_Recv(&idx,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRMPI(ierr); 7557ffc4695bSBarry Smith ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 7558b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 755957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 7560cf9c20a2SJed Brown if (PetscUnlikelyDebug(!oldranks)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 75617fb8a5e4SKarl Rupp ranks_send_to_idx[0] = procs_candidates[oldranks[idx]]; 756227b6a85dSStefano Zampini } else if (oldranks) { 75637fb8a5e4SKarl Rupp ranks_send_to_idx[0] = oldranks[idx]; 756427b6a85dSStefano Zampini } else { 75657fb8a5e4SKarl Rupp ranks_send_to_idx[0] = idx; 7566e7931f94SStefano Zampini } 756757de7509SStefano Zampini } 756852e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 7569e7931f94SStefano Zampini /* clean up */ 7570e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 757152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 7572e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 7573e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 7574e7931f94SStefano Zampini } 757527b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 757657de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 7577e7931f94SStefano Zampini 7578e7931f94SStefano Zampini /* assemble parallel IS for sends */ 7579e7931f94SStefano Zampini i = 1; 758027b6a85dSStefano Zampini if (!color) i=0; 758157de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 7582e7931f94SStefano Zampini PetscFunctionReturn(0); 7583e7931f94SStefano Zampini } 7584e7931f94SStefano Zampini 7585e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 7586e7931f94SStefano Zampini 75871e0482f5SStefano Zampini PetscErrorCode PCBDDCMatISSubassemble(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[]) 7588e7931f94SStefano Zampini { 758970cf5478SStefano Zampini Mat local_mat; 7590e7931f94SStefano Zampini IS is_sends_internal; 75919d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 75921ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 75939d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 7594e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 7595e7931f94SStefano Zampini PetscInt* l2gmap_indices; 7596e7931f94SStefano Zampini const PetscInt* is_indices; 7597e7931f94SStefano Zampini MatType new_local_type; 7598e7931f94SStefano Zampini /* buffers */ 7599e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 760028143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 76019d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 76021683a169SBarry Smith PetscScalar *ptr_vals,*recv_buffer_vals; 76031683a169SBarry Smith const PetscScalar *send_buffer_vals; 76041ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 7605e7931f94SStefano Zampini /* MPI */ 760628143c3dSStefano Zampini MPI_Comm comm,comm_n; 760728143c3dSStefano Zampini PetscSubcomm subcomm; 7608e569e4e1SStefano Zampini PetscMPIInt n_sends,n_recvs,size; 760928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 761028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 76111ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 76121ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 76131ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 7614e7931f94SStefano Zampini PetscErrorCode ierr; 7615e7931f94SStefano Zampini 7616e7931f94SStefano Zampini PetscFunctionBegin; 761757de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 7618e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 7619fbfcfee5SBarry Smith if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME); 762057de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 762157de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 762257de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 762357de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 762457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 76251ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 76261ae86dd6SStefano Zampini if (nvecs) { 76271ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 76281ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 76291ae86dd6SStefano Zampini } 763057de7509SStefano Zampini /* further checks */ 7631e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 7632e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 7633e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 7634e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 7635e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 763657de7509SStefano Zampini if (reuse && *mat_n) { 763770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 763857de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 763970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 764028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 764170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 764270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 764370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 764470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 764570cf5478SStefano Zampini } 7646e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 7647064a246eSJacob Faibussowitsch PetscValidLogicalCollectiveInt(mat,bs,1); 764857de7509SStefano Zampini 7649e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 7650e7931f94SStefano Zampini if (!is_sends) { 765128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 7652bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 7653c8587f34SStefano Zampini } else { 7654e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 7655e7931f94SStefano Zampini is_sends_internal = is_sends; 7656c8587f34SStefano Zampini } 7657e7931f94SStefano Zampini 7658e7931f94SStefano Zampini /* get comm */ 7659a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 7660e7931f94SStefano Zampini 7661e7931f94SStefano Zampini /* compute number of sends */ 7662e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 7663e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 7664e7931f94SStefano Zampini 7665e7931f94SStefano Zampini /* compute number of receives */ 7666ffc4695bSBarry Smith ierr = MPI_Comm_size(comm,&size);CHKERRMPI(ierr); 7667e569e4e1SStefano Zampini ierr = PetscMalloc1(size,&iflags);CHKERRQ(ierr); 7668580bdb30SBarry Smith ierr = PetscArrayzero(iflags,size);CHKERRQ(ierr); 7669e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 7670e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 7671e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 7672e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 7673e7931f94SStefano Zampini 767428143c3dSStefano Zampini /* restrict comm if requested */ 76750a545947SLisandro Dalcin subcomm = NULL; 767628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 767728143c3dSStefano Zampini if (restrict_comm) { 7678779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 7679779c1cceSStefano Zampini 768028143c3dSStefano Zampini color = 0; 768153a05cb3SStefano Zampini if (restrict_full) { 768253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 768353a05cb3SStefano Zampini } else { 768453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 768553a05cb3SStefano Zampini } 7686820f2d46SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRMPI(ierr); 7687e569e4e1SStefano Zampini subcommsize = size - subcommsize; 768828143c3dSStefano Zampini /* check if reuse has been requested */ 768957de7509SStefano Zampini if (reuse) { 769028143c3dSStefano Zampini if (*mat_n) { 769128143c3dSStefano Zampini PetscMPIInt subcommsize2; 7692ffc4695bSBarry Smith ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRMPI(ierr); 769328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 769428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 769528143c3dSStefano Zampini } else { 769628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 769728143c3dSStefano Zampini } 769828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 7699779c1cceSStefano Zampini PetscMPIInt rank; 7700779c1cceSStefano Zampini 7701ffc4695bSBarry Smith ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 770228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 770328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 770428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 7705306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 770628143c3dSStefano Zampini } 770728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 770828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 770928143c3dSStefano Zampini } else { 771028143c3dSStefano Zampini comm_n = comm; 771128143c3dSStefano Zampini } 771228143c3dSStefano Zampini 7713e7931f94SStefano Zampini /* prepare send/receive buffers */ 7714e569e4e1SStefano Zampini ierr = PetscMalloc1(size,&ilengths_idxs);CHKERRQ(ierr); 7715580bdb30SBarry Smith ierr = PetscArrayzero(ilengths_idxs,size);CHKERRQ(ierr); 7716e569e4e1SStefano Zampini ierr = PetscMalloc1(size,&ilengths_vals);CHKERRQ(ierr); 7717580bdb30SBarry Smith ierr = PetscArrayzero(ilengths_vals,size);CHKERRQ(ierr); 771828143c3dSStefano Zampini if (nis) { 7719e569e4e1SStefano Zampini ierr = PetscCalloc1(size,&ilengths_idxs_is);CHKERRQ(ierr); 772028143c3dSStefano Zampini } 7721e7931f94SStefano Zampini 772228143c3dSStefano Zampini /* Get data from local matrices */ 77236c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 7724e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 7725e7931f94SStefano Zampini /* 7726e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 7727e7931f94SStefano Zampini send_buffer_idxs should contain: 7728e7931f94SStefano Zampini - MatType_PRIVATE type 7729e7931f94SStefano Zampini - PetscInt size_of_l2gmap 7730e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 7731e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 7732e7931f94SStefano Zampini */ 77336c4ed002SBarry Smith else { 77341683a169SBarry Smith ierr = MatDenseGetArrayRead(local_mat,&send_buffer_vals);CHKERRQ(ierr); 77353bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 7736854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 7737e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 7738e7931f94SStefano Zampini send_buffer_idxs[1] = i; 77393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 7740580bdb30SBarry Smith ierr = PetscArraycpy(&send_buffer_idxs[2],ptr_idxs,i);CHKERRQ(ierr); 77413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 7742e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 7743e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 7744e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 7745e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 7746c8587f34SStefano Zampini } 7747c8587f34SStefano Zampini } 7748e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 774928143c3dSStefano Zampini /* additional is (if any) */ 775028143c3dSStefano Zampini if (nis) { 775128143c3dSStefano Zampini PetscMPIInt psum; 775228143c3dSStefano Zampini PetscInt j; 775328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 775428143c3dSStefano Zampini PetscInt plen; 775528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 775628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 775728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 775828143c3dSStefano Zampini } 7759854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 776028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 776128143c3dSStefano Zampini PetscInt plen; 776228143c3dSStefano Zampini const PetscInt *is_array_idxs; 776328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 776428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 776528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 7766580bdb30SBarry Smith ierr = PetscArraycpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen);CHKERRQ(ierr); 776728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 776828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 776928143c3dSStefano Zampini } 777028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 777128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 777228143c3dSStefano Zampini } 777328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 777428143c3dSStefano Zampini } 77753b3b1effSJed Brown ierr = MatISRestoreLocalMat(mat,&local_mat);CHKERRQ(ierr); 777628143c3dSStefano Zampini 7777e7931f94SStefano Zampini buf_size_idxs = 0; 7778e7931f94SStefano Zampini buf_size_vals = 0; 777928143c3dSStefano Zampini buf_size_idxs_is = 0; 77801ae86dd6SStefano Zampini buf_size_vecs = 0; 7781e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7782e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 7783e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 778428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 77851ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 7786e7931f94SStefano Zampini } 7787785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 7788785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 778995ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 77901ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 7791e7931f94SStefano Zampini 7792e7931f94SStefano Zampini /* get new tags for clean communications */ 7793e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 7794e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 779528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 77961ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 7797e7931f94SStefano Zampini 7798e7931f94SStefano Zampini /* allocate for requests */ 7799785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 7800785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 780195ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 78021ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 7803785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 7804785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 780595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 78061ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 7807e7931f94SStefano Zampini 7808e7931f94SStefano Zampini /* communications */ 7809e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 7810e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 781128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 78121ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 7813e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7814e7931f94SStefano Zampini source_dest = onodes[i]; 7815ffc4695bSBarry Smith ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRMPI(ierr); 7816ffc4695bSBarry Smith ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRMPI(ierr); 7817e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7818e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 781928143c3dSStefano Zampini if (nis) { 782057de7509SStefano Zampini source_dest = onodes_is[i]; 7821ffc4695bSBarry Smith ierr = MPI_Irecv(ptr_idxs_is,olengths_idxs_is[i],MPIU_INT,source_dest,tag_idxs_is,comm,&recv_req_idxs_is[i]);CHKERRMPI(ierr); 782228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 782328143c3dSStefano Zampini } 78241ae86dd6SStefano Zampini if (nvecs) { 78251ae86dd6SStefano Zampini source_dest = onodes[i]; 7826ffc4695bSBarry Smith ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRMPI(ierr); 78271ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 78281ae86dd6SStefano Zampini } 7829e7931f94SStefano Zampini } 7830e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 7831e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 7832ffc4695bSBarry Smith ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRMPI(ierr); 7833ffc4695bSBarry Smith ierr = MPI_Isend((PetscScalar*)send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRMPI(ierr); 783428143c3dSStefano Zampini if (nis) { 7835ffc4695bSBarry Smith 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]);CHKERRMPI(ierr); 783628143c3dSStefano Zampini } 78371ae86dd6SStefano Zampini if (nvecs) { 78381ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 7839ffc4695bSBarry Smith ierr = MPI_Isend(send_buffer_vecs,ilengths_idxs[source_dest]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&send_req_vecs[i]);CHKERRMPI(ierr); 78401ae86dd6SStefano Zampini } 7841e7931f94SStefano Zampini } 7842e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 7843e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 7844e7931f94SStefano Zampini 7845e7931f94SStefano Zampini /* assemble new l2g map */ 7846ffc4695bSBarry Smith ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 7847e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 78489d30be91SStefano Zampini new_local_rows = 0; 7849e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 78509d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 7851e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7852e7931f94SStefano Zampini } 78539d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 7854e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 78559d30be91SStefano Zampini new_local_rows = 0; 7856e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7857580bdb30SBarry Smith ierr = PetscArraycpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,*(ptr_idxs+1));CHKERRQ(ierr); 78589d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 7859e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7860e7931f94SStefano Zampini } 78619d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 78629d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 7863e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 7864e7931f94SStefano Zampini 7865e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 7866e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 7867e7931f94SStefano 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) */ 7868e7931f94SStefano Zampini if (n_recvs) { 786928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 7870e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 7871e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7872e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 7873e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 7874e7931f94SStefano Zampini break; 7875e7931f94SStefano Zampini } 7876e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7877e7931f94SStefano Zampini } 7878e7931f94SStefano Zampini switch (new_local_type_private) { 787928143c3dSStefano Zampini case MATDENSE_PRIVATE: 7880e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7881e7931f94SStefano Zampini bs = 1; 7882e7931f94SStefano Zampini break; 7883e7931f94SStefano Zampini case MATAIJ_PRIVATE: 7884e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7885e7931f94SStefano Zampini bs = 1; 7886e7931f94SStefano Zampini break; 7887e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 7888e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 7889e7931f94SStefano Zampini break; 7890e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 7891e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 7892e7931f94SStefano Zampini break; 7893e7931f94SStefano Zampini default: 7894fbfcfee5SBarry Smith SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,PETSC_FUNCTION_NAME); 7895e7931f94SStefano Zampini } 7896ed8ed4edSstefano_zampini } else { /* by default, new_local_type is seqaij */ 7897ed8ed4edSstefano_zampini new_local_type = MATSEQAIJ; 789828143c3dSStefano Zampini bs = 1; 7899e7931f94SStefano Zampini } 7900e7931f94SStefano Zampini 790170cf5478SStefano Zampini /* create MATIS object if needed */ 790257de7509SStefano Zampini if (!reuse) { 7903e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 7904e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 790570cf5478SStefano Zampini } else { 790670cf5478SStefano Zampini /* it also destroys the local matrices */ 790757de7509SStefano Zampini if (*mat_n) { 790870cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 790957de7509SStefano Zampini } else { /* this is a fake object */ 791057de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 791157de7509SStefano Zampini } 791270cf5478SStefano Zampini } 791370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 7914e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 79159d30be91SStefano Zampini 7916ffc4695bSBarry Smith ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 79179d30be91SStefano Zampini 79189d30be91SStefano Zampini /* Global to local map of received indices */ 79199d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 79209d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 79219d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 79229d30be91SStefano Zampini 79239d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 79249d30be91SStefano Zampini buf_size_idxs = 0; 79259d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 79269d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 79279d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 79289d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 79299d30be91SStefano Zampini } 79309d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 79319d30be91SStefano Zampini 79329d30be91SStefano Zampini /* set preallocation */ 79339d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 79349d30be91SStefano Zampini if (!newisdense) { 79350a545947SLisandro Dalcin PetscInt *new_local_nnz=NULL; 79369d30be91SStefano Zampini 79379d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 79389d30be91SStefano Zampini if (n_recvs) { 79399d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 79409d30be91SStefano Zampini } 79419d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 79429d30be91SStefano Zampini PetscInt j; 79439d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 79449d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 79459d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 79469d30be91SStefano Zampini } 79479d30be91SStefano Zampini } else { 79489d30be91SStefano Zampini /* TODO */ 79499d30be91SStefano Zampini } 79509d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 79519d30be91SStefano Zampini } 79529d30be91SStefano Zampini if (new_local_nnz) { 79539d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 79549d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 79559d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 79569d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 79579d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 79589d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 79599d30be91SStefano Zampini } else { 79609d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 79619d30be91SStefano Zampini } 79629d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 79639d30be91SStefano Zampini } else { 79649d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 79659d30be91SStefano Zampini } 7966e7931f94SStefano Zampini 7967e7931f94SStefano Zampini /* set values */ 7968e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 79699d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 7970e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7971e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 7972e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 79739d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 7974e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7975e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7976e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 797728143c3dSStefano Zampini } else { 797828143c3dSStefano Zampini /* TODO */ 7979e7931f94SStefano Zampini } 7980e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7981e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 7982e7931f94SStefano Zampini } 7983e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7984e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 79853b3b1effSJed Brown ierr = MatISRestoreLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 798670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 798770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 79889d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 7989e7931f94SStefano Zampini 7990dfd14d43SStefano Zampini #if 0 799128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 7992e7931f94SStefano Zampini Vec lvec,rvec; 7993e7931f94SStefano Zampini PetscReal infty_error; 7994e7931f94SStefano Zampini 79952a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 7996e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 7997e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 7998e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 799970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 8000e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 8001*1e1ea65dSPierre Jolivet ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error);CHKERRQ(ierr); 8002e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 8003e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 8004e7931f94SStefano Zampini } 800528143c3dSStefano Zampini #endif 8006e7931f94SStefano Zampini 800728143c3dSStefano Zampini /* assemble new additional is (if any) */ 800828143c3dSStefano Zampini if (nis) { 800928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 801028143c3dSStefano Zampini 8011ffc4695bSBarry Smith ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 8012854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 801328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 801428143c3dSStefano Zampini psum = 0; 801528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 801628143c3dSStefano Zampini for (j=0;j<nis;j++) { 801728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 801828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 801928143c3dSStefano Zampini psum += plen; 802028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 802128143c3dSStefano Zampini } 802228143c3dSStefano Zampini } 8023854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 8024854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 802528143c3dSStefano Zampini for (i=1;i<nis;i++) { 802628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 802728143c3dSStefano Zampini } 8028580bdb30SBarry Smith ierr = PetscArrayzero(count_is,nis);CHKERRQ(ierr); 802928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 803028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 803128143c3dSStefano Zampini for (j=0;j<nis;j++) { 803228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 8033580bdb30SBarry Smith ierr = PetscArraycpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen);CHKERRQ(ierr); 803428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 803528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 803628143c3dSStefano Zampini } 803728143c3dSStefano Zampini } 803828143c3dSStefano Zampini for (i=0;i<nis;i++) { 803928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 8040c3b366b1Sprj- ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr); 804128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 804228143c3dSStefano Zampini } 804328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 804428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 804528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 804628143c3dSStefano Zampini } 8047e7931f94SStefano Zampini /* free workspace */ 804828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 8049ffc4695bSBarry Smith ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 8050e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 8051ffc4695bSBarry Smith ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 8052e7931f94SStefano Zampini if (isdense) { 8053e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 80541683a169SBarry Smith ierr = MatDenseRestoreArrayRead(local_mat,&send_buffer_vals);CHKERRQ(ierr); 80553b3b1effSJed Brown ierr = MatISRestoreLocalMat(mat,&local_mat);CHKERRQ(ierr); 8056e7931f94SStefano Zampini } else { 8057e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 8058e7931f94SStefano Zampini } 805928143c3dSStefano Zampini if (nis) { 8060ffc4695bSBarry Smith ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 806128143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 806228143c3dSStefano Zampini } 80631ae86dd6SStefano Zampini 80641ae86dd6SStefano Zampini if (nvecs) { 8065ffc4695bSBarry Smith ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 8066ffc4695bSBarry Smith ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 80671ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 80681ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 80691ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 80701ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 80711ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 80721ae86dd6SStefano Zampini /* set values */ 80731ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 80741ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 80751ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 80761ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 80771ae86dd6SStefano Zampini PetscInt j; 80781ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 80791ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 80801ae86dd6SStefano Zampini } 80811ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 80821ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 80831ae86dd6SStefano Zampini } 80841ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 80851ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 80861ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 80871ae86dd6SStefano Zampini } 80881ae86dd6SStefano Zampini 80891ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 80901ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 8091e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 8092e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 80931ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 809428143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 8095e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 8096e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 80971ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 809828143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 8099e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 8100e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 8101e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 8102e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 8103e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 810428143c3dSStefano Zampini if (nis) { 810528143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 810628143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 810728143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 810828143c3dSStefano Zampini } 810928143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 811028143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 811128143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 811228143c3dSStefano Zampini for (i=0;i<nis;i++) { 811328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 811428143c3dSStefano Zampini } 81151ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 81161ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 81171ae86dd6SStefano Zampini } 811853a05cb3SStefano Zampini *mat_n = NULL; 811928143c3dSStefano Zampini } 8120e7931f94SStefano Zampini PetscFunctionReturn(0); 8121e7931f94SStefano Zampini } 8122a57a6d2fSStefano Zampini 812312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 8124af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 812512edc857SStefano Zampini 8126c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 8127c8587f34SStefano Zampini { 8128c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 8129c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 813020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 81311ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 81321e0482f5SStefano Zampini Mat coarseG,t_coarse_mat_is; 81339881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 813420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 81354f819b78SStefano Zampini IS coarse_is,*isarray,corners; 81366e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 813730368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 8138e569e4e1SStefano Zampini PetscInt coarse_eqs_per_proc; 8139f9eb5b7dSStefano Zampini PC pc_temp; 8140c8587f34SStefano Zampini PCType coarse_pc_type; 8141c8587f34SStefano Zampini KSPType coarse_ksp_type; 8142f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 81437274672aSStefano Zampini PetscBool coarse_reuse; 81441e0482f5SStefano Zampini PetscInt ncoarse,nedcfield; 814568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 814622bc73bbSStefano Zampini PetscScalar *array; 814757de7509SStefano Zampini MatReuse coarse_mat_reuse; 814857de7509SStefano Zampini PetscBool restr, full_restr, have_void; 8149e569e4e1SStefano Zampini PetscMPIInt size; 81509881197aSStefano Zampini PetscErrorCode ierr; 8151fdc09c96SStefano Zampini 8152c8587f34SStefano Zampini PetscFunctionBegin; 815343371fb9SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_CoarseSetUp[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 8154c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 815568457ee5SStefano 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 */ 8156fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 81575a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 81587de4f681Sstefano_zampini 81597de4f681Sstefano_zampini pcbddc->new_primal_space = PETSC_TRUE; 8160fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 8161f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 8162f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 8163f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 8164fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 816551bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 816651bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 8167727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 8168fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 8169fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 8170fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 8171f4ddd8eeSStefano Zampini } 8172fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 8173fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 8174f4ddd8eeSStefano Zampini } 817570cf5478SStefano Zampini /* reset any subassembling information */ 817657de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 817770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 817857de7509SStefano Zampini } 81796e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 8180fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 8181f4ddd8eeSStefano Zampini } 818257de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 818357de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 818457de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 818557de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 818618a45a71SStefano Zampini } else { 818757de7509SStefano Zampini coarse_mat = NULL; 818857de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 81896e683305SStefano Zampini } 8190e7931f94SStefano Zampini 8191abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 8192abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 8193abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 8194abbbba34SStefano Zampini 8195abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 81964f819b78SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_submat_dense);CHKERRQ(ierr); 8197e176bc59SStefano 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); 81986e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 81996e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 82006e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 8201abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 8202abbbba34SStefano Zampini 820357de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 820457de7509SStefano Zampini im_active = !!(pcis->n); 8205820f2d46SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 820657de7509SStefano Zampini 820714f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 820828d58a37SPierre Jolivet /* restr : whether we want to exclude senders (which are not receivers) from the subassembling pattern */ 820957de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 8210ffc4695bSBarry Smith ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRMPI(ierr); 821157de7509SStefano Zampini coarse_mat_is = NULL; 821257de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 821357de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 8214e569e4e1SStefano Zampini coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 8215ce64c636SStefano Zampini if (coarse_eqs_per_proc < 0) coarse_eqs_per_proc = pcbddc->coarse_size; 821657de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 8217e569e4e1SStefano Zampini if (pcbddc->coarse_size <= pcbddc->coarse_eqs_limit) multilevel_requested = PETSC_FALSE; 821857de7509SStefano Zampini if (multilevel_requested) { 821957de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 822057de7509SStefano Zampini restr = PETSC_FALSE; 822157de7509SStefano Zampini full_restr = PETSC_FALSE; 822257de7509SStefano Zampini } else { 8223e569e4e1SStefano Zampini ncoarse = pcbddc->coarse_size/coarse_eqs_per_proc + !!(pcbddc->coarse_size%coarse_eqs_per_proc); 822457de7509SStefano Zampini restr = PETSC_TRUE; 822557de7509SStefano Zampini full_restr = PETSC_TRUE; 822657de7509SStefano Zampini } 8227e569e4e1SStefano Zampini if (!pcbddc->coarse_size || size == 1) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 822857de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 822957de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 8230a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 8231bb360cb4SStefano Zampini if (multilevel_requested) { 8232bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 8233bb360cb4SStefano Zampini } else { 8234bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 8235bb360cb4SStefano Zampini } 8236a198735bSStefano Zampini } else { 82377de4f681Sstefano_zampini PetscMPIInt rank; 823828d58a37SPierre Jolivet 8239ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRMPI(ierr); 8240e569e4e1SStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 8241a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 8242a198735bSStefano Zampini } 824357de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 824457de7509SStefano Zampini PetscInt psum; 824557de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 824657de7509SStefano Zampini else psum = 0; 8247820f2d46SBarry Smith ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 8248075e25bcSStefano Zampini have_void = ncoarse < size ? PETSC_TRUE : PETSC_FALSE; 824957de7509SStefano Zampini } 825057de7509SStefano Zampini /* determine if we can go multilevel */ 825157de7509SStefano Zampini if (multilevel_requested) { 825257de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 825357de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 825457de7509SStefano Zampini } 825557de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 825657de7509SStefano Zampini 8257e4d548c7SStefano Zampini /* dump subassembling pattern */ 8258e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 8259e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 8260e4d548c7SStefano Zampini } 82616e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 82621e0482f5SStefano Zampini nedcfield = -1; 82634f819b78SStefano Zampini corners = NULL; 82648966356dSPierre Jolivet if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal || pcbddc->corner_selected)) { /* protects from unneeded computations */ 82656e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 82666e683305SStefano Zampini const PetscInt *idxs; 82676e683305SStefano Zampini ISLocalToGlobalMapping tmap; 82686e683305SStefano Zampini 82696e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 82700be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 82716e683305SStefano Zampini /* allocate space for temporary storage */ 8272854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 8273854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 82746e683305SStefano Zampini /* allocate for IS array */ 82756e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 82761e0482f5SStefano Zampini if (pcbddc->nedclocal) { 82771e0482f5SStefano Zampini if (pcbddc->nedfield > -1) { 82781e0482f5SStefano Zampini nedcfield = pcbddc->nedfield; 82791e0482f5SStefano Zampini } else { 82801e0482f5SStefano Zampini nedcfield = 0; 82816080607fSStefano Zampini if (nisdofs) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen (%D)",nisdofs); 82821e0482f5SStefano Zampini nisdofs = 1; 82831e0482f5SStefano Zampini } 82841e0482f5SStefano Zampini } 82856e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 828627b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 828730368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 8288854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 82896e683305SStefano Zampini /* dofs splitting */ 82906e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 82916e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 82921e0482f5SStefano Zampini if (nedcfield != i) { 82936e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 82946e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 82956e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 82966e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 82971e0482f5SStefano Zampini } else { 82981e0482f5SStefano Zampini ierr = ISGetLocalSize(pcbddc->nedclocal,&tsize);CHKERRQ(ierr); 82991e0482f5SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 83001e0482f5SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 83016080607fSStefano Zampini if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping coarse nedelec field! %D != %D",tsize,nout); 83021e0482f5SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 83031e0482f5SStefano Zampini } 83046e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 830530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 83066e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 83076e683305SStefano Zampini } 83086e683305SStefano Zampini /* neumann boundaries */ 83096e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 83106e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 83116e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 83126e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 83136e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 83146e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 83156e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 831630368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 83176e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 83186e683305SStefano Zampini } 83194f819b78SStefano Zampini /* coordinates */ 83204f819b78SStefano Zampini if (pcbddc->corner_selected) { 83214f819b78SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&corners);CHKERRQ(ierr); 83224f819b78SStefano Zampini ierr = ISGetLocalSize(corners,&tsize);CHKERRQ(ierr); 83234f819b78SStefano Zampini ierr = ISGetIndices(corners,&idxs);CHKERRQ(ierr); 83244f819b78SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 83254f819b78SStefano Zampini if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping corners! %D != %D",tsize,nout); 83264f819b78SStefano Zampini ierr = ISRestoreIndices(corners,&idxs);CHKERRQ(ierr); 83274f819b78SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&corners);CHKERRQ(ierr); 83284f819b78SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 83294f819b78SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&corners);CHKERRQ(ierr); 83304f819b78SStefano Zampini } 83316e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 83326e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 83336e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 83346e683305SStefano Zampini } else { 83356e683305SStefano Zampini nis = 0; 83366e683305SStefano Zampini nisdofs = 0; 83376e683305SStefano Zampini nisneu = 0; 833830368db7SStefano Zampini nisvert = 0; 83396e683305SStefano Zampini isarray = NULL; 83406e683305SStefano Zampini } 83416e683305SStefano Zampini /* destroy no longer needed map */ 83426e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 83436e683305SStefano Zampini 834457de7509SStefano Zampini /* subassemble */ 834557de7509SStefano Zampini if (multilevel_allowed) { 83461ae86dd6SStefano Zampini Vec vp[1]; 83471ae86dd6SStefano Zampini PetscInt nvecs = 0; 834857de7509SStefano Zampini PetscBool reuse,reuser; 83491ae86dd6SStefano Zampini 835057de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 835157de7509SStefano Zampini else reuse = PETSC_FALSE; 8352820f2d46SBarry Smith ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 83531ae86dd6SStefano Zampini vp[0] = NULL; 83541ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 83551ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 83561ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 83571ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 83581ae86dd6SStefano Zampini nvecs = 1; 83591ae86dd6SStefano Zampini 83601ae86dd6SStefano Zampini if (pcbddc->divudotp) { 8361a198735bSStefano Zampini Mat B,loc_divudotp; 83621ae86dd6SStefano Zampini Vec v,p; 83631ae86dd6SStefano Zampini IS dummy; 83641ae86dd6SStefano Zampini PetscInt np; 83651ae86dd6SStefano Zampini 8366a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 8367a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 83681ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 83697dae84e0SHong Zhang ierr = MatCreateSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 83701ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 83711ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 83721ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 83731ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 83741ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83751ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 83761ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 83771ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 83781ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 83791ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 83801ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 83811ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 838274e2c79eSStefano Zampini } 83831ae86dd6SStefano Zampini } 83841ae86dd6SStefano Zampini if (reuser) { 83851e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 838674e2c79eSStefano Zampini } else { 83871e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray,nvecs,vp);CHKERRQ(ierr); 83881ae86dd6SStefano Zampini } 83891ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 83901683a169SBarry Smith PetscScalar *arraym; 83911683a169SBarry Smith const PetscScalar *arrayv; 83921ae86dd6SStefano Zampini PetscInt nl; 83931ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 83941ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 83951ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 83961683a169SBarry Smith ierr = VecGetArrayRead(vp[0],&arrayv);CHKERRQ(ierr); 8397580bdb30SBarry Smith ierr = PetscArraycpy(arraym,arrayv,nl);CHKERRQ(ierr); 83981683a169SBarry Smith ierr = VecRestoreArrayRead(vp[0],&arrayv);CHKERRQ(ierr); 83991ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 84001ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 8401a198735bSStefano Zampini } else { 8402a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 84031ae86dd6SStefano Zampini } 84041ae86dd6SStefano Zampini } else { 84051e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,0,NULL,0,NULL);CHKERRQ(ierr); 84066e683305SStefano Zampini } 840757de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 840857de7509SStefano Zampini if (!multilevel_allowed) { 8409487b449aSStefano Zampini ierr = MatConvert(coarse_mat_is,MATAIJ,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 84106e683305SStefano Zampini } else { 841157de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 841257de7509SStefano Zampini if (coarse_mat_is) { 841357de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 841457de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 841557de7509SStefano Zampini coarse_mat = coarse_mat_is; 841657de7509SStefano Zampini } 8417779c1cceSStefano Zampini } 8418779c1cceSStefano Zampini } 841957de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 842057de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 84216e683305SStefano Zampini 84226e683305SStefano Zampini /* create local to global scatters for coarse problem */ 842368457ee5SStefano Zampini if (compute_vecs) { 84246e683305SStefano Zampini PetscInt lrows; 84256e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 842657de7509SStefano Zampini if (coarse_mat) { 842757de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 84286e683305SStefano Zampini } else { 84296e683305SStefano Zampini lrows = 0; 84306e683305SStefano Zampini } 84316e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 84326e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 843389535278SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,coarse_mat ? coarse_mat->defaultvectype : VECSTANDARD);CHKERRQ(ierr); 84346e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 84359448b7f1SJunchao Zhang ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 84366e683305SStefano Zampini } 84376e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 8438c8587f34SStefano Zampini 8439f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 8440f9eb5b7dSStefano Zampini if (multilevel_allowed) { 8441f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 8442f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 8443f9eb5b7dSStefano Zampini } else { 8444f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 8445f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 8446c8587f34SStefano Zampini } 8447c8587f34SStefano Zampini 84486e683305SStefano Zampini /* print some info if requested */ 84496e683305SStefano Zampini if (pcbddc->dbg_flag) { 84506e683305SStefano Zampini if (!multilevel_allowed) { 84516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 84526e683305SStefano Zampini if (multilevel_requested) { 84536080607fSStefano 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); 84546e683305SStefano Zampini } else if (pcbddc->max_levels) { 84556080607fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%D)\n",pcbddc->max_levels);CHKERRQ(ierr); 84566e683305SStefano Zampini } 84576e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 84586e683305SStefano Zampini } 84596e683305SStefano Zampini } 84606e683305SStefano Zampini 84611e0482f5SStefano Zampini /* communicate coarse discrete gradient */ 84621e0482f5SStefano Zampini coarseG = NULL; 84631e0482f5SStefano Zampini if (pcbddc->nedcG && multilevel_allowed) { 84641e0482f5SStefano Zampini MPI_Comm ccomm; 84651e0482f5SStefano Zampini if (coarse_mat) { 84661e0482f5SStefano Zampini ccomm = PetscObjectComm((PetscObject)coarse_mat); 84671e0482f5SStefano Zampini } else { 84681e0482f5SStefano Zampini ccomm = MPI_COMM_NULL; 84691e0482f5SStefano Zampini } 84701e0482f5SStefano Zampini ierr = MatMPIAIJRestrict(pcbddc->nedcG,ccomm,&coarseG);CHKERRQ(ierr); 84711e0482f5SStefano Zampini } 84721e0482f5SStefano Zampini 8473f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 847457de7509SStefano Zampini if (coarse_mat) { 847528d58a37SPierre Jolivet PetscBool isredundant,isbddc,force,valid; 84766a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 84777274672aSStefano Zampini 84786e683305SStefano Zampini if (pcbddc->dbg_flag) { 847957de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 84806e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 84816e683305SStefano Zampini } 8482f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 8483312be037SStefano Zampini char prefix[256],str_level[16]; 8484e604994aSStefano Zampini size_t len; 84851e0482f5SStefano Zampini 848657de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 8487422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 8488c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 8489f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 849057de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 8491c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 84926e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 8493c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 84941e0482f5SStefano Zampini /* TODO is this logic correct? should check for coarse_mat type */ 8495c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 8496e604994aSStefano Zampini /* prefix */ 8497e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 8498e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 8499e604994aSStefano Zampini if (!pcbddc->current_level) { 8500a126751eSBarry Smith ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,sizeof(prefix));CHKERRQ(ierr); 8501a126751eSBarry Smith ierr = PetscStrlcat(prefix,"pc_bddc_coarse_",sizeof(prefix));CHKERRQ(ierr); 8502c8587f34SStefano Zampini } else { 8503e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 8504312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 8505312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 8506a126751eSBarry Smith /* Nonstandard use of PetscStrncpy() to copy only a portion of the string */ 850734d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 850835529e7bSStefano Zampini ierr = PetscSNPrintf(str_level,sizeof(str_level),"l%d_",(int)(pcbddc->current_level));CHKERRQ(ierr); 8509a126751eSBarry Smith ierr = PetscStrlcat(prefix,str_level,sizeof(prefix));CHKERRQ(ierr); 8510e604994aSStefano Zampini } 8511e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 85123e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 85133e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 85143e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 85153e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 8516f9eb5b7dSStefano Zampini /* allow user customization */ 8517f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 8518e569e4e1SStefano Zampini /* get some info after set from options */ 8519e569e4e1SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 852028d58a37SPierre Jolivet /* multilevel cannot be done with coarse PC different from BDDC, NN, HPDDM, unless forced to */ 852128d58a37SPierre Jolivet force = PETSC_FALSE; 852228d58a37SPierre Jolivet ierr = PetscOptionsGetBool(NULL,((PetscObject)pc_temp)->prefix,"-pc_type_forced",&force,NULL);CHKERRQ(ierr); 852328d58a37SPierre Jolivet ierr = PetscObjectTypeCompareAny((PetscObject)pc_temp,&valid,PCBDDC,PCNN,PCHPDDM,"");CHKERRQ(ierr); 8524e569e4e1SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 852528d58a37SPierre Jolivet if (multilevel_allowed && !force && !valid) { 8526e569e4e1SStefano Zampini isbddc = PETSC_TRUE; 8527e569e4e1SStefano Zampini ierr = PCSetType(pc_temp,PCBDDC);CHKERRQ(ierr); 8528e569e4e1SStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 8529e569e4e1SStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 8530e569e4e1SStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 85314f819b78SStefano Zampini if (pc_temp->ops->setfromoptions) { /* need to setfromoptions again, skipping the pc_type */ 85324f819b78SStefano Zampini ierr = PetscObjectOptionsBegin((PetscObject)pc_temp);CHKERRQ(ierr); 85334f819b78SStefano Zampini ierr = (*pc_temp->ops->setfromoptions)(PetscOptionsObject,pc_temp);CHKERRQ(ierr); 85344f819b78SStefano Zampini ierr = PetscObjectProcessOptionsHandlers(PetscOptionsObject,(PetscObject)pc_temp);CHKERRQ(ierr); 85354f819b78SStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 85364f819b78SStefano Zampini pc_temp->setfromoptionscalled++; 85374f819b78SStefano Zampini } 8538e569e4e1SStefano Zampini } 85393e3c6dadSStefano Zampini } 85403e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 854151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 85423e3c6dadSStefano Zampini if (nisdofs) { 85433e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 85443e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 85453e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 85463e3c6dadSStefano Zampini } 85473e3c6dadSStefano Zampini } 85483e3c6dadSStefano Zampini if (nisneu) { 85493e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 85503e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 8551312be037SStefano Zampini } 855230368db7SStefano Zampini if (nisvert) { 855330368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 855430368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 855530368db7SStefano Zampini } 85561e0482f5SStefano Zampini if (coarseG) { 85571e0482f5SStefano Zampini ierr = PCBDDCSetDiscreteGradient(pc_temp,coarseG,1,nedcfield,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 85581e0482f5SStefano Zampini } 8559f9eb5b7dSStefano Zampini 8560f9eb5b7dSStefano Zampini /* get some info after set from options */ 8561f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 85624f819b78SStefano Zampini 8563b76f3995Sstefano_zampini /* multilevel can only be requested via -pc_bddc_levels or PCBDDCSetLevels */ 8564b76f3995Sstefano_zampini if (isbddc && !multilevel_allowed) { 8565f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 8566f9eb5b7dSStefano Zampini } 856728d58a37SPierre Jolivet /* multilevel cannot be done with coarse PC different from BDDC, NN, HPDDM, unless forced to */ 856828d58a37SPierre Jolivet force = PETSC_FALSE; 856928d58a37SPierre Jolivet ierr = PetscOptionsGetBool(NULL,((PetscObject)pc_temp)->prefix,"-pc_type_forced",&force,NULL);CHKERRQ(ierr); 857028d58a37SPierre Jolivet ierr = PetscObjectTypeCompareAny((PetscObject)pc_temp,&valid,PCBDDC,PCNN,PCHPDDM,"");CHKERRQ(ierr); 857128d58a37SPierre Jolivet if (multilevel_requested && multilevel_allowed && !valid && !force) { 8572b76f3995Sstefano_zampini ierr = PCSetType(pc_temp,PCBDDC);CHKERRQ(ierr); 8573b76f3995Sstefano_zampini } 85747274672aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 85754f3a063dSStefano Zampini if (isredundant) { 85764f3a063dSStefano Zampini KSP inner_ksp; 85774f3a063dSStefano Zampini PC inner_pc; 85789326c5c6Sstefano_zampini 85794f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 85804f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 85814f3a063dSStefano Zampini } 8582f9eb5b7dSStefano Zampini 858357de7509SStefano Zampini /* parameters which miss an API */ 85847274672aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 858557de7509SStefano Zampini if (isbddc) { 8586720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 85877274672aSStefano Zampini 8588720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 858957de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 8590e569e4e1SStefano Zampini pcbddc_coarse->coarse_eqs_limit = pcbddc->coarse_eqs_limit; 859127b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 859227b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 8593a198735bSStefano Zampini Mat coarsedivudotp_is; 8594a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 8595a198735bSStefano Zampini IS row,col; 8596a198735bSStefano Zampini const PetscInt *gidxs; 8597a198735bSStefano Zampini PetscInt n,st,M,N; 8598a198735bSStefano Zampini 8599a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 8600ffc4695bSBarry Smith ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRMPI(ierr); 8601a198735bSStefano Zampini st = st-n; 8602a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 8603a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 8604a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 8605a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 8606a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 8607a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 8608a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 8609a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 8610a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 8611a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 8612a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 8613a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 8614a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 8615a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 8616a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 8617a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 8618a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 8619a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 8620a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 8621a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 86228ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 8623a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 8624720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 8625bd2a564bSStefano Zampini if (pcbddc->adaptive_threshold[0] == 0.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 8626720d30f9SStefano Zampini } 8627d4d8cf7bSStefano Zampini } 86289881197aSStefano Zampini 86293301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 86305a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 86313301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 86323301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 86333301b35fSStefano Zampini } 86343301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 86353301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 86363301b35fSStefano Zampini } 86373301b35fSStefano Zampini if (pc->pmat->spd_set) { 86383301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 86393301b35fSStefano Zampini } 864027b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 864127b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 864227b6a85dSStefano Zampini } 86436e683305SStefano Zampini /* set operators */ 864404fe1396SStefano Zampini ierr = MatViewFromOptions(coarse_mat,(PetscObject)pc,"-pc_bddc_coarse_mat_view");CHKERRQ(ierr); 86453007b4efSStefano Zampini ierr = MatSetOptionsPrefix(coarse_mat,((PetscObject)pcbddc->coarse_ksp)->prefix);CHKERRQ(ierr); 86465f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 86476e683305SStefano Zampini if (pcbddc->dbg_flag) { 86486e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 86496e683305SStefano Zampini } 86506e683305SStefano Zampini } 86511e0482f5SStefano Zampini ierr = MatDestroy(&coarseG);CHKERRQ(ierr); 86526e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 8653b1ecc7b1SStefano Zampini #if 0 8654b9b85e73SStefano Zampini { 8655b9b85e73SStefano Zampini PetscViewer viewer; 8656b9b85e73SStefano Zampini char filename[256]; 8657b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 8658b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 86596a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 8660b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 8661f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 8662b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 8663b9b85e73SStefano Zampini } 8664b9b85e73SStefano Zampini #endif 8665f9eb5b7dSStefano Zampini 86664f819b78SStefano Zampini if (corners) { 86674f819b78SStefano Zampini Vec gv; 86684f819b78SStefano Zampini IS is; 86694f819b78SStefano Zampini const PetscInt *idxs; 86704f819b78SStefano Zampini PetscInt i,d,N,n,cdim = pcbddc->mat_graph->cdim; 86714f819b78SStefano Zampini PetscScalar *coords; 86724f819b78SStefano Zampini 86734f819b78SStefano Zampini if (!pcbddc->mat_graph->cloc) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing local coordinates"); 86744f819b78SStefano Zampini ierr = VecGetSize(pcbddc->coarse_vec,&N);CHKERRQ(ierr); 86754f819b78SStefano Zampini ierr = VecGetLocalSize(pcbddc->coarse_vec,&n);CHKERRQ(ierr); 86764f819b78SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcbddc->coarse_vec),&gv);CHKERRQ(ierr); 86774f819b78SStefano Zampini ierr = VecSetBlockSize(gv,cdim);CHKERRQ(ierr); 86784f819b78SStefano Zampini ierr = VecSetSizes(gv,n*cdim,N*cdim);CHKERRQ(ierr); 86794f819b78SStefano Zampini ierr = VecSetType(gv,VECSTANDARD);CHKERRQ(ierr); 86804f819b78SStefano Zampini ierr = VecSetFromOptions(gv);CHKERRQ(ierr); 86814f819b78SStefano Zampini ierr = VecSet(gv,PETSC_MAX_REAL);CHKERRQ(ierr); /* we only propagate coordinates from vertices constraints */ 86824f819b78SStefano Zampini 86834f819b78SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&is);CHKERRQ(ierr); 86844f819b78SStefano Zampini ierr = ISGetLocalSize(is,&n);CHKERRQ(ierr); 86854f819b78SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 86864f819b78SStefano Zampini ierr = PetscMalloc1(n*cdim,&coords);CHKERRQ(ierr); 86874f819b78SStefano Zampini for (i=0;i<n;i++) { 86884f819b78SStefano Zampini for (d=0;d<cdim;d++) { 86894f819b78SStefano Zampini coords[cdim*i+d] = pcbddc->mat_graph->coords[cdim*idxs[i]+d]; 86904f819b78SStefano Zampini } 86914f819b78SStefano Zampini } 86924f819b78SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 86934f819b78SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&is);CHKERRQ(ierr); 86944f819b78SStefano Zampini 86954f819b78SStefano Zampini ierr = ISGetLocalSize(corners,&n);CHKERRQ(ierr); 86964f819b78SStefano Zampini ierr = ISGetIndices(corners,&idxs);CHKERRQ(ierr); 86974f819b78SStefano Zampini ierr = VecSetValuesBlocked(gv,n,idxs,coords,INSERT_VALUES);CHKERRQ(ierr); 86984f819b78SStefano Zampini ierr = ISRestoreIndices(corners,&idxs);CHKERRQ(ierr); 86994f819b78SStefano Zampini ierr = PetscFree(coords);CHKERRQ(ierr); 87004f819b78SStefano Zampini ierr = VecAssemblyBegin(gv);CHKERRQ(ierr); 87014f819b78SStefano Zampini ierr = VecAssemblyEnd(gv);CHKERRQ(ierr); 87024f819b78SStefano Zampini ierr = VecGetArray(gv,&coords);CHKERRQ(ierr); 87034f819b78SStefano Zampini if (pcbddc->coarse_ksp) { 87044f819b78SStefano Zampini PC coarse_pc; 87054f819b78SStefano Zampini PetscBool isbddc; 87064f819b78SStefano Zampini 87074f819b78SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 87084f819b78SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 87094f819b78SStefano Zampini if (isbddc) { /* coarse coordinates have PETSC_MAX_REAL, specific for BDDC */ 87104f819b78SStefano Zampini PetscReal *realcoords; 87114f819b78SStefano Zampini 87124f819b78SStefano Zampini ierr = VecGetLocalSize(gv,&n);CHKERRQ(ierr); 87134f819b78SStefano Zampini #if defined(PETSC_USE_COMPLEX) 87144f819b78SStefano Zampini ierr = PetscMalloc1(n,&realcoords);CHKERRQ(ierr); 87154f819b78SStefano Zampini for (i=0;i<n;i++) realcoords[i] = PetscRealPart(coords[i]); 87164f819b78SStefano Zampini #else 87174f819b78SStefano Zampini realcoords = coords; 87184f819b78SStefano Zampini #endif 87194f819b78SStefano Zampini ierr = PCSetCoordinates(coarse_pc,cdim,n/cdim,realcoords);CHKERRQ(ierr); 87204f819b78SStefano Zampini #if defined(PETSC_USE_COMPLEX) 87214f819b78SStefano Zampini ierr = PetscFree(realcoords);CHKERRQ(ierr); 87224f819b78SStefano Zampini #endif 87234f819b78SStefano Zampini } 87244f819b78SStefano Zampini } 87254f819b78SStefano Zampini ierr = VecRestoreArray(gv,&coords);CHKERRQ(ierr); 87264f819b78SStefano Zampini ierr = VecDestroy(&gv);CHKERRQ(ierr); 87274f819b78SStefano Zampini } 87284f819b78SStefano Zampini ierr = ISDestroy(&corners);CHKERRQ(ierr); 87294f819b78SStefano Zampini 873098a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 873198a51de6SStefano Zampini Vec crhs,csol; 873204708bb6SStefano Zampini 8733f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 8734f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 8735f347579bSStefano Zampini if (!csol) { 87362a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 8737f9eb5b7dSStefano Zampini } 8738f347579bSStefano Zampini if (!crhs) { 87392a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 8740f347579bSStefano Zampini } 8741b0f5fe93SStefano Zampini } 87421ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 8743b0f5fe93SStefano Zampini 8744b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 8745b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 8746b0f5fe93SStefano Zampini 8747b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 87484f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 87494f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 87504f1b2e48SStefano Zampini } 8751b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 8752b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 8753b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8754b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8755b0f5fe93SStefano Zampini if (coarse_mat) { 8756b0f5fe93SStefano Zampini Vec nullv; 8757b0f5fe93SStefano Zampini PetscScalar *array,*array2; 8758b0f5fe93SStefano Zampini PetscInt nl; 8759b0f5fe93SStefano Zampini 8760b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 8761b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 8762b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 8763b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 8764580bdb30SBarry Smith ierr = PetscArraycpy(array2,array,nl);CHKERRQ(ierr); 8765b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 8766b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 8767b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 8768b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 8769b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 8770b0f5fe93SStefano Zampini } 8771b0f5fe93SStefano Zampini } 877243371fb9SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_CoarseSetUp[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 8773b0f5fe93SStefano Zampini 877443371fb9SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_CoarseSolver[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 8775b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 8776b0f5fe93SStefano Zampini PetscBool ispreonly; 8777b0f5fe93SStefano Zampini 8778b0f5fe93SStefano Zampini if (CoarseNullSpace) { 8779b0f5fe93SStefano Zampini PetscBool isnull; 8780b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 8781bef83e63SStefano Zampini if (isnull) { 8782b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 8783b0f5fe93SStefano Zampini } 8784bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 8785b0f5fe93SStefano Zampini } 8786b0f5fe93SStefano Zampini /* setup coarse ksp */ 8787b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 8788cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 8789cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 87906e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates)) { 8791c8587f34SStefano Zampini KSP check_ksp; 87922b510759SStefano Zampini KSPType check_ksp_type; 8793c8587f34SStefano Zampini PC check_pc; 87946e683305SStefano Zampini Vec check_vec,coarse_vec; 87956a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 87962b510759SStefano Zampini PetscInt its; 87976e683305SStefano Zampini PetscBool compute_eigs; 87986e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 87996e683305SStefano Zampini PetscInt neigs; 88008e185a42SStefano Zampini const char *prefix; 8801c8587f34SStefano Zampini 88022b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 88036e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 880415579a77SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)check_ksp,(PetscObject)pcbddc->coarse_ksp,0);CHKERRQ(ierr); 8805399ffe99SStefano Zampini ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,PETSC_FALSE);CHKERRQ(ierr); 880623ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 8807f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 8808e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 8809e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 8810e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 88112b510759SStefano Zampini if (ispreonly) { 88122b510759SStefano Zampini check_ksp_type = KSPPREONLY; 88136e683305SStefano Zampini compute_eigs = PETSC_FALSE; 88142b510759SStefano Zampini } else { 8815cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 88166e683305SStefano Zampini compute_eigs = PETSC_TRUE; 8817c8587f34SStefano Zampini } 8818c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 88196e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 88206e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 88216e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 8822a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 8823a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 8824a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 8825a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 8826c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 8827c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 8828c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 8829c8587f34SStefano Zampini /* create random vec */ 88302701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 8831c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 88326e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 8833c8587f34SStefano Zampini /* solve coarse problem */ 88346e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 8835c0decd05SBarry Smith ierr = KSPCheckSolve(check_ksp,pc,coarse_vec);CHKERRQ(ierr); 8836cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 88376e683305SStefano Zampini if (compute_eigs) { 8838854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 8839854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 88406e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 88411ae86dd6SStefano Zampini if (neigs) { 88426e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 88436e683305SStefano Zampini lambda_min = eigs_r[0]; 88446e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 88452701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 88462701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 8847cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 8848cbcc2c2aSStefano Zampini } 8849c8587f34SStefano Zampini } 8850c8587f34SStefano Zampini } 88511ae86dd6SStefano Zampini } 8852cbcc2c2aSStefano Zampini 8853c8587f34SStefano Zampini /* check coarse problem residual error */ 88546e683305SStefano Zampini if (pcbddc->dbg_flag) { 88556e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 88566e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 88576e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 8858c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 88596e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 88606e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 8861779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 88626e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 88636e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 88646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 88656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 8866b0f5fe93SStefano Zampini if (CoarseNullSpace) { 8867b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 8868b0f5fe93SStefano Zampini } 88696e683305SStefano Zampini if (compute_eigs) { 88706e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 8871b03ebc13SStefano Zampini KSPConvergedReason reason; 8872deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 8873c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 8874b03ebc13SStefano Zampini ierr = KSPGetConvergedReason(check_ksp,&reason);CHKERRQ(ierr); 88756e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 8876b03ebc13SStefano 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); 88776e683305SStefano Zampini for (i=0;i<neigs;i++) { 88786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 8879c8587f34SStefano Zampini } 88806e683305SStefano Zampini } 88816e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 88826e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 88836e683305SStefano Zampini } 8884e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 88852701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 8886c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 88876e683305SStefano Zampini if (compute_eigs) { 88886e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 88896e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 8890c8587f34SStefano Zampini } 88916e683305SStefano Zampini } 88926e683305SStefano Zampini } 8893bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 8894cbcc2c2aSStefano Zampini /* print additional info */ 8895cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 88966e683305SStefano Zampini /* waits until all processes reaches this point */ 88976e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 88986080607fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %D\n",pcbddc->current_level);CHKERRQ(ierr); 8899cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8900cbcc2c2aSStefano Zampini } 8901cbcc2c2aSStefano Zampini 89022b510759SStefano Zampini /* free memory */ 8903fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 890443371fb9SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_CoarseSolver[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 8905c8587f34SStefano Zampini PetscFunctionReturn(0); 8906c8587f34SStefano Zampini } 8907674ae819SStefano Zampini 8908f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 8909f34684f1SStefano Zampini { 8910f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8911f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 8912f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 8913dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 8914dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 891573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 8916dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 8917f34684f1SStefano Zampini PetscErrorCode ierr; 8918f34684f1SStefano Zampini 8919f34684f1SStefano Zampini PetscFunctionBegin; 8920f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 89216c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 8922dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 89233bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 8924dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 8925dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 89266583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 8927dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 8928dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 8929dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 89306c4ed002SBarry 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); 8931dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 8932dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 8933580bdb30SBarry Smith ierr = PetscArraycpy(local_primal_indices,t_local_primal_indices,local_size);CHKERRQ(ierr); 8934dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 8935dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 8936f34684f1SStefano Zampini 8937f34684f1SStefano Zampini /* check numbering */ 8938f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 8939019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 8940dc456d91SStefano Zampini PetscInt i; 8941b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 8942f34684f1SStefano Zampini 8943f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8944f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 8945f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 89461575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8947019a44ceSStefano Zampini /* counter */ 8948019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8949019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 8950019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8951019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8952019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8953019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8954f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 8955f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 8956727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 8957f34684f1SStefano Zampini } 8958f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8959f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8960f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8961e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8962e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8963e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8964e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8965f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8966019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8967f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8968019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 89692c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 897075c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 8971b9b85e73SStefano Zampini set_error = PETSC_TRUE; 89722c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 89736080607fSStefano 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); 8974f34684f1SStefano Zampini } 8975f34684f1SStefano Zampini } 8976019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8977820f2d46SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 8978f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8979f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8980f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 8981f34684f1SStefano Zampini } 8982f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8983f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8984e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8985e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8986f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 89876080607fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %D (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 8988b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 8989ca8b9ea9SStefano Zampini PetscInt *gidxs; 8990ca8b9ea9SStefano Zampini 8991ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 89923bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 8993f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 8994f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8995f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 8996f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 89976080607fSStefano 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); 8998f34684f1SStefano Zampini } 8999f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 9000ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 9001f34684f1SStefano Zampini } 9002f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 90031575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 9004302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 9005f34684f1SStefano Zampini } 90066080607fSStefano Zampini 9007f34684f1SStefano Zampini /* get back data */ 9008f34684f1SStefano Zampini *coarse_size_n = coarse_size; 9009f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 9010674ae819SStefano Zampini PetscFunctionReturn(0); 9011674ae819SStefano Zampini } 9012674ae819SStefano Zampini 9013a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 9014e456f2a8SStefano Zampini { 9015e456f2a8SStefano Zampini IS localis_t; 9016a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 9017e456f2a8SStefano Zampini PetscScalar *vals; 9018e456f2a8SStefano Zampini PetscErrorCode ierr; 9019e456f2a8SStefano Zampini 9020e456f2a8SStefano Zampini PetscFunctionBegin; 9021a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 9022e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 9023854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 9024e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 9025e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 9026a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 9027a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 90281035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 90293151afb1SStefano Zampini ierr = VecSetOption(gwork,VEC_IGNORE_NEGATIVE_INDICES,PETSC_TRUE);CHKERRQ(ierr); 9030a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 90311035eff8SStefano Zampini } 9032a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 9033e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 9034e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 9035a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 9036a7dc3881SStefano Zampini /* now compute set in local ordering */ 9037a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9038a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9039a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 9040a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 9041a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 9042ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 9043e456f2a8SStefano Zampini lsize++; 9044e456f2a8SStefano Zampini } 9045e456f2a8SStefano Zampini } 9046854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 9047a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 9048ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 9049e456f2a8SStefano Zampini idxs[lsize++] = i; 9050e456f2a8SStefano Zampini } 9051e456f2a8SStefano Zampini } 9052a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 9053a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 9054e456f2a8SStefano Zampini *localis = localis_t; 9055e456f2a8SStefano Zampini PetscFunctionReturn(0); 9056e456f2a8SStefano Zampini } 9057906d46d4SStefano Zampini 905808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 9059b96c3477SStefano Zampini { 9060a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 9061b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 9062b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 9063a64f4aa4SStefano Zampini Mat S_j; 9064b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 9065b96c3477SStefano Zampini PetscBool free_used_adj; 9066b96c3477SStefano Zampini PetscErrorCode ierr; 9067b96c3477SStefano Zampini 9068b96c3477SStefano Zampini PetscFunctionBegin; 906943371fb9SStefano Zampini ierr = PetscLogEventBegin(PC_BDDC_Schurs[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 9070b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 9071b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 907208122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 9073b96c3477SStefano Zampini used_xadj = NULL; 9074b96c3477SStefano Zampini used_adjncy = NULL; 9075b96c3477SStefano Zampini } else { 907608122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 907708122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 907808122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 907908122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 9080b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 9081b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 9082b96c3477SStefano Zampini } else { 90832fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 9084b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 9085b96c3477SStefano Zampini PetscInt nvtxs; 9086b96c3477SStefano Zampini 90872fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 90882fffb893SStefano Zampini if (flg_row) { 9089b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 9090580bdb30SBarry Smith ierr = PetscArraycpy(used_xadj,xadj,nvtxs+1);CHKERRQ(ierr); 9091580bdb30SBarry Smith ierr = PetscArraycpy(used_adjncy,adjncy,xadj[nvtxs]);CHKERRQ(ierr); 9092b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 90932fffb893SStefano Zampini } else { 90942fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 90952fffb893SStefano Zampini used_xadj = NULL; 90962fffb893SStefano Zampini used_adjncy = NULL; 90972fffb893SStefano Zampini } 90982fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 9099b96c3477SStefano Zampini } 9100b96c3477SStefano Zampini } 9101d5574798SStefano Zampini 9102d5574798SStefano Zampini /* setup sub_schurs data */ 91032f37b69bSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->pA_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 9104df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 9105df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 9106a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 910791af6908SStefano 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); 9108a64f4aa4SStefano Zampini } else { 910972b8c272SStefano Zampini Mat change = NULL; 91109d54b7f4SStefano Zampini Vec scaling = NULL; 9111111315fdSstefano_zampini IS change_primal = NULL, iP; 9112111315fdSstefano_zampini PetscInt benign_n; 9113111315fdSstefano_zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 91147ebab0bbSStefano Zampini PetscBool need_change = PETSC_FALSE; 9115111315fdSstefano_zampini PetscBool discrete_harmonic = PETSC_FALSE; 9116a3df083aSStefano Zampini 91175feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 91185feab87aSStefano Zampini PetscInt n_vertices; 91195feab87aSStefano Zampini 91205feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 91212034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 91225feab87aSStefano Zampini } 9123a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 9124a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 9125ca92afb2SStefano Zampini } else { 9126a3df083aSStefano Zampini benign_n = 0; 9127ca92afb2SStefano Zampini } 9128b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 9129b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 9130b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 913172b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 913222db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 9133820f2d46SBarry Smith ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRMPI(ierr); 913422db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 9135b7ab4a40SStefano Zampini } 9136b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 9137b7ab4a40SStefano 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 */ 9138b7ab4a40SStefano Zampini if (need_change) { 913988c03ad3SStefano Zampini PC_IS *pcisf; 914088c03ad3SStefano Zampini PC_BDDC *pcbddcf; 914188c03ad3SStefano Zampini PC pcf; 914288c03ad3SStefano Zampini 9143e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 914488c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 914588c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 914688c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 9147b9be95fcSstefano_zampini 914888c03ad3SStefano Zampini /* hacks */ 914988c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 915072b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 915172b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 915272b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 915372b8c272SStefano Zampini pcisf->n = pcis->n; 915472b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 915588c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 915688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 915788c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 915888c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 915988c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 916088c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 916172b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 916288c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 9163b9be95fcSstefano_zampini 9164b9be95fcSstefano_zampini /* setup constraints so that we can get information on primal vertices and change of basis (in local numbering) */ 916588c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 916672b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 916772b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 916872b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 916972b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 9170b9be95fcSstefano_zampini 917188c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 917272b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 917388c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 917488c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 917588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 917688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 917788c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 917888c03ad3SStefano Zampini pcf->ops->destroy = NULL; 9179b9be95fcSstefano_zampini pcf->ops->reset = NULL; 918088c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 918188c03ad3SStefano Zampini } 91829d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 9183111315fdSstefano_zampini 9184111315fdSstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_iP",(PetscObject*)&iP);CHKERRQ(ierr); 9185111315fdSstefano_zampini if (iP) { 9186111315fdSstefano_zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)iP),sub_schurs->prefix,"BDDC sub_schurs options","PC");CHKERRQ(ierr); 9187111315fdSstefano_zampini ierr = PetscOptionsBool("-sub_schurs_discrete_harmonic",NULL,NULL,discrete_harmonic,&discrete_harmonic,NULL);CHKERRQ(ierr); 9188111315fdSstefano_zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 9189111315fdSstefano_zampini } 9190111315fdSstefano_zampini if (discrete_harmonic) { 9191111315fdSstefano_zampini Mat A; 9192111315fdSstefano_zampini ierr = MatDuplicate(pcbddc->local_mat,MAT_COPY_VALUES,&A);CHKERRQ(ierr); 9193111315fdSstefano_zampini ierr = MatZeroRowsColumnsIS(A,iP,1.0,NULL,NULL);CHKERRQ(ierr); 9194111315fdSstefano_zampini ierr = PetscObjectCompose((PetscObject)A,"__KSPFETIDP_iP",(PetscObject)iP);CHKERRQ(ierr); 9195111315fdSstefano_zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,A,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); 9196111315fdSstefano_zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 9197111315fdSstefano_zampini } else { 919891af6908SStefano 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); 9199111315fdSstefano_zampini } 920072b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 920172b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 9202ca92afb2SStefano Zampini } 9203d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 9204b96c3477SStefano Zampini 9205b96c3477SStefano Zampini /* free adjacency */ 9206b96c3477SStefano Zampini if (free_used_adj) { 9207b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 9208b96c3477SStefano Zampini } 920943371fb9SStefano Zampini ierr = PetscLogEventEnd(PC_BDDC_Schurs[pcbddc->current_level],pc,0,0,0);CHKERRQ(ierr); 9210b96c3477SStefano Zampini PetscFunctionReturn(0); 9211b96c3477SStefano Zampini } 9212b96c3477SStefano Zampini 921308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 9214b96c3477SStefano Zampini { 9215b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 9216b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 9217b96c3477SStefano Zampini PCBDDCGraph graph; 9218b96c3477SStefano Zampini PetscErrorCode ierr; 9219b96c3477SStefano Zampini 9220b96c3477SStefano Zampini PetscFunctionBegin; 9221b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 922208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 92233301b35fSStefano Zampini IS verticesIS,verticescomm; 92243301b35fSStefano Zampini PetscInt vsize,*idxs; 9225b96c3477SStefano Zampini 9226b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 92273301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 92283301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 92293301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 92303301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 9231c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 9232b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 9233be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 9234441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 92353301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 9236b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 9237b96c3477SStefano Zampini } else { 9238b96c3477SStefano Zampini graph = pcbddc->mat_graph; 9239b96c3477SStefano Zampini } 9240e4d548c7SStefano Zampini /* print some info */ 92415c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 9242e4d548c7SStefano Zampini IS vertices; 9243e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 9244c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 9245e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 9246e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 9247e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 9248e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 92496080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%D)\n",PetscGlobalRank,(int)nv,pcbddc->use_vertices);CHKERRQ(ierr); 92506080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%D)\n",PetscGlobalRank,(int)nedges,pcbddc->use_edges);CHKERRQ(ierr); 92516080607fSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%D)\n",PetscGlobalRank,(int)nfaces,pcbddc->use_faces);CHKERRQ(ierr); 9252e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 9253e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 9254c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 9255e4d548c7SStefano Zampini } 9256b96c3477SStefano Zampini 9257b96c3477SStefano Zampini /* sub_schurs init */ 9258b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 9259b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 9260b334f244SStefano Zampini } 926188113c35SStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,((PetscObject)pc)->prefix,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 9262a64f4aa4SStefano Zampini 9263b96c3477SStefano Zampini /* free graph struct */ 926408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 9265b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 9266b96c3477SStefano Zampini } 9267b96c3477SStefano Zampini PetscFunctionReturn(0); 9268b96c3477SStefano Zampini } 9269fa34dd3eSStefano Zampini 9270fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 9271fa34dd3eSStefano Zampini { 9272fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 9273fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 9274fa34dd3eSStefano Zampini PetscErrorCode ierr; 9275fa34dd3eSStefano Zampini 9276fa34dd3eSStefano Zampini PetscFunctionBegin; 9277fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 9278fa34dd3eSStefano Zampini IS zerodiag = NULL; 92794f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 9280fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 92814f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 928275c01103SStefano Zampini PetscReal norm; 9283fa34dd3eSStefano Zampini PetscInt i; 9284fa34dd3eSStefano Zampini 9285fa34dd3eSStefano Zampini /* B0 and B0_B */ 9286fa34dd3eSStefano Zampini if (zerodiag) { 9287fa34dd3eSStefano Zampini IS dummy; 9288fa34dd3eSStefano Zampini 92894f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 92907dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 9291fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 9292fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 9293fa34dd3eSStefano Zampini } 9294fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 9295fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 9296fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 9297fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9298fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9299fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9300fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9301fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 9302fa34dd3eSStefano Zampini /* S_j */ 93032f37b69bSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->pA_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 9304fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 9305fa34dd3eSStefano Zampini 9306fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 9307fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 9308fa34dd3eSStefano Zampini /* continuous in primal space */ 9309fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 9310fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9311fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9312fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 93134f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 93144f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 9315fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 9316fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 9317fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 9318fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 9319fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9320fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9321fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 9322fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 9323fa34dd3eSStefano Zampini 9324fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 9325fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 9326fa34dd3eSStefano Zampini /* local with Schur */ 9327fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 9328fa34dd3eSStefano Zampini if (zerodiag) { 9329fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 93304f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 9331fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 9332fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 9333fa34dd3eSStefano Zampini } 9334fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 9335fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9336fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9337fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 9338fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 9339fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 9340fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 9341fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 9342fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 9343fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9344fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9345fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9346fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 9347fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 9348fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 9349fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 9350fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 9351fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 9352fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 9353fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 9354fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9355fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 9356fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 9357fa34dd3eSStefano Zampini if (zerodiag) { 9358fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 9359fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 93604f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 9361fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 9362fa34dd3eSStefano Zampini } 9363fa34dd3eSStefano Zampini /* BDDC */ 9364fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 9365fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 9366fa34dd3eSStefano Zampini 9367fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 9368fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 9369fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 93706080607fSStefano Zampini ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm);CHKERRQ(ierr); 93714f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 93726080607fSStefano Zampini ierr = PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%D] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i]));CHKERRQ(ierr); 9373fa34dd3eSStefano Zampini } 93744f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 9375fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 9376fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 9377fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 9378fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 9379fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 9380fa34dd3eSStefano Zampini } 9381fa34dd3eSStefano Zampini PetscFunctionReturn(0); 9382fa34dd3eSStefano Zampini } 93831e0482f5SStefano Zampini 93841e0482f5SStefano Zampini #include <../src/mat/impls/aij/mpi/mpiaij.h> 93851e0482f5SStefano Zampini PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B) 93861e0482f5SStefano Zampini { 93871e0482f5SStefano Zampini Mat At; 93881e0482f5SStefano Zampini IS rows; 93891e0482f5SStefano Zampini PetscInt rst,ren; 93901e0482f5SStefano Zampini PetscErrorCode ierr; 93911e0482f5SStefano Zampini PetscLayout rmap; 93921e0482f5SStefano Zampini 93931e0482f5SStefano Zampini PetscFunctionBegin; 93941e0482f5SStefano Zampini rst = ren = 0; 93951e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 93961e0482f5SStefano Zampini ierr = PetscLayoutCreate(ccomm,&rmap);CHKERRQ(ierr); 93971e0482f5SStefano Zampini ierr = PetscLayoutSetSize(rmap,A->rmap->N);CHKERRQ(ierr); 93981e0482f5SStefano Zampini ierr = PetscLayoutSetBlockSize(rmap,1);CHKERRQ(ierr); 93991e0482f5SStefano Zampini ierr = PetscLayoutSetUp(rmap);CHKERRQ(ierr); 94001e0482f5SStefano Zampini ierr = PetscLayoutGetRange(rmap,&rst,&ren);CHKERRQ(ierr); 94011e0482f5SStefano Zampini } 9402e07686f2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)A),ren-rst,rst,1,&rows);CHKERRQ(ierr); 94037dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,rows,NULL,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 94041e0482f5SStefano Zampini ierr = ISDestroy(&rows);CHKERRQ(ierr); 94051e0482f5SStefano Zampini 94061e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 94071e0482f5SStefano Zampini Mat_MPIAIJ *a,*b; 94081e0482f5SStefano Zampini IS from,to; 94091e0482f5SStefano Zampini Vec gvec; 94101e0482f5SStefano Zampini PetscInt lsize; 94111e0482f5SStefano Zampini 94121e0482f5SStefano Zampini ierr = MatCreate(ccomm,B);CHKERRQ(ierr); 94131e0482f5SStefano Zampini ierr = MatSetSizes(*B,ren-rst,PETSC_DECIDE,PETSC_DECIDE,At->cmap->N);CHKERRQ(ierr); 94141e0482f5SStefano Zampini ierr = MatSetType(*B,MATAIJ);CHKERRQ(ierr); 94151e0482f5SStefano Zampini ierr = PetscLayoutDestroy(&((*B)->rmap));CHKERRQ(ierr); 94161e0482f5SStefano Zampini ierr = PetscLayoutSetUp((*B)->cmap);CHKERRQ(ierr); 94171e0482f5SStefano Zampini a = (Mat_MPIAIJ*)At->data; 94181e0482f5SStefano Zampini b = (Mat_MPIAIJ*)(*B)->data; 9419ffc4695bSBarry Smith ierr = MPI_Comm_size(ccomm,&b->size);CHKERRMPI(ierr); 9420ffc4695bSBarry Smith ierr = MPI_Comm_rank(ccomm,&b->rank);CHKERRMPI(ierr); 94211e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->A);CHKERRQ(ierr); 94221e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->B);CHKERRQ(ierr); 94231e0482f5SStefano Zampini b->A = a->A; 94241e0482f5SStefano Zampini b->B = a->B; 94251e0482f5SStefano Zampini 94261e0482f5SStefano Zampini b->donotstash = a->donotstash; 94271e0482f5SStefano Zampini b->roworiented = a->roworiented; 94280a545947SLisandro Dalcin b->rowindices = NULL; 94290a545947SLisandro Dalcin b->rowvalues = NULL; 94301e0482f5SStefano Zampini b->getrowactive = PETSC_FALSE; 94311e0482f5SStefano Zampini 94321e0482f5SStefano Zampini (*B)->rmap = rmap; 94331e0482f5SStefano Zampini (*B)->factortype = A->factortype; 94341e0482f5SStefano Zampini (*B)->assembled = PETSC_TRUE; 94351e0482f5SStefano Zampini (*B)->insertmode = NOT_SET_VALUES; 94361e0482f5SStefano Zampini (*B)->preallocated = PETSC_TRUE; 94371e0482f5SStefano Zampini 94381e0482f5SStefano Zampini if (a->colmap) { 94391e0482f5SStefano Zampini #if defined(PETSC_USE_CTABLE) 94401e0482f5SStefano Zampini ierr = PetscTableCreateCopy(a->colmap,&b->colmap);CHKERRQ(ierr); 94411e0482f5SStefano Zampini #else 94421e0482f5SStefano Zampini ierr = PetscMalloc1(At->cmap->N,&b->colmap);CHKERRQ(ierr); 94431e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)*B,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 9444580bdb30SBarry Smith ierr = PetscArraycpy(b->colmap,a->colmap,At->cmap->N);CHKERRQ(ierr); 94451e0482f5SStefano Zampini #endif 94460a545947SLisandro Dalcin } else b->colmap = NULL; 94471e0482f5SStefano Zampini if (a->garray) { 94481e0482f5SStefano Zampini PetscInt len; 94491e0482f5SStefano Zampini len = a->B->cmap->n; 94501e0482f5SStefano Zampini ierr = PetscMalloc1(len+1,&b->garray);CHKERRQ(ierr); 94511e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)(*B),len*sizeof(PetscInt));CHKERRQ(ierr); 9452580bdb30SBarry Smith if (len) { ierr = PetscArraycpy(b->garray,a->garray,len);CHKERRQ(ierr); } 94530a545947SLisandro Dalcin } else b->garray = NULL; 94541e0482f5SStefano Zampini 94551e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->lvec);CHKERRQ(ierr); 94561e0482f5SStefano Zampini b->lvec = a->lvec; 94571e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->lvec);CHKERRQ(ierr); 94581e0482f5SStefano Zampini 94591e0482f5SStefano Zampini /* cannot use VecScatterCopy */ 94601e0482f5SStefano Zampini ierr = VecGetLocalSize(b->lvec,&lsize);CHKERRQ(ierr); 94611e0482f5SStefano Zampini ierr = ISCreateGeneral(ccomm,lsize,b->garray,PETSC_USE_POINTER,&from);CHKERRQ(ierr); 94621e0482f5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,lsize,0,1,&to);CHKERRQ(ierr); 94631e0482f5SStefano Zampini ierr = MatCreateVecs(*B,&gvec,NULL);CHKERRQ(ierr); 94649448b7f1SJunchao Zhang ierr = VecScatterCreate(gvec,from,b->lvec,to,&b->Mvctx);CHKERRQ(ierr); 94651e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->Mvctx);CHKERRQ(ierr); 94661e0482f5SStefano Zampini ierr = ISDestroy(&from);CHKERRQ(ierr); 94671e0482f5SStefano Zampini ierr = ISDestroy(&to);CHKERRQ(ierr); 94681e0482f5SStefano Zampini ierr = VecDestroy(&gvec);CHKERRQ(ierr); 94691e0482f5SStefano Zampini } 94701e0482f5SStefano Zampini ierr = MatDestroy(&At);CHKERRQ(ierr); 94711e0482f5SStefano Zampini PetscFunctionReturn(0); 94721e0482f5SStefano Zampini } 9473