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> 4c80a6c00SStefano Zampini #include <petscdmplex.h> 5674ae819SStefano Zampini #include <petscblaslapack.h> 6daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 7c80a6c00SStefano Zampini #include <petsc/private/dmpleximpl.h> 87620a527SStefano Zampini #include <petscdmda.h> 9674ae819SStefano Zampini 101e0482f5SStefano Zampini static PetscErrorCode MatMPIAIJRestrict(Mat,MPI_Comm,Mat*); 111e0482f5SStefano Zampini 12f498cd09SStefano Zampini /* if range is true, it returns B s.t. span{B} = range(A) 13f498cd09SStefano Zampini if range is false, it returns B s.t. range(B) _|_ range(A) */ 14f498cd09SStefano Zampini PetscErrorCode MatDenseOrthogonalRangeOrComplement(Mat A, PetscBool range, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B) 15a13144ffSStefano Zampini { 16a13144ffSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 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; 22a13144ffSStefano Zampini 23a13144ffSStefano Zampini PetscFunctionBegin; 24a13144ffSStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 25a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"LAPACK _GESVD not available"); 26614dbb09SStefano Zampini #else 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); 52a13144ffSStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,&lierr)); 53a13144ffSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 54a13144ffSStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 55a13144ffSStefano Zampini ierr = MatDenseRestoreArray(A,&data);CHKERRQ(ierr); 56a13144ffSStefano Zampini for (i=0;i<n;i++) if (sing[i] < PETSC_SMALL) break; 57a13144ffSStefano Zampini if (!rwork) { 58a13144ffSStefano Zampini ierr = PetscFree(sing);CHKERRQ(ierr); 59a13144ffSStefano Zampini } 60a13144ffSStefano Zampini if (!work) { 61a13144ffSStefano Zampini ierr = PetscFree(uwork);CHKERRQ(ierr); 62a13144ffSStefano Zampini } 63a13144ffSStefano Zampini /* create B */ 64f498cd09SStefano Zampini if (!range) { 65a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,nr-i,NULL,B);CHKERRQ(ierr); 66a13144ffSStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 67a13144ffSStefano Zampini ierr = PetscMemcpy(data,U+nr*i,(nr-i)*nr*sizeof(PetscScalar));CHKERRQ(ierr); 68f498cd09SStefano Zampini } else { 69f498cd09SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,i,NULL,B);CHKERRQ(ierr); 70f498cd09SStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 71f498cd09SStefano Zampini ierr = PetscMemcpy(data,U,i*nr*sizeof(PetscScalar));CHKERRQ(ierr); 72f498cd09SStefano Zampini } 73a13144ffSStefano Zampini ierr = MatDenseRestoreArray(*B,&data);CHKERRQ(ierr); 74a13144ffSStefano Zampini ierr = PetscFree(U);CHKERRQ(ierr); 75614dbb09SStefano Zampini #endif 76614dbb09SStefano Zampini #else /* PETSC_USE_COMPLEX */ 77a13144ffSStefano Zampini PetscFunctionBegin; 78a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for complexes"); 79a13144ffSStefano Zampini #endif 80a13144ffSStefano Zampini PetscFunctionReturn(0); 81a13144ffSStefano Zampini } 82a13144ffSStefano Zampini 831e0482f5SStefano Zampini /* TODO REMOVE */ 841e0482f5SStefano Zampini #if defined(PRINT_GDET) 851e0482f5SStefano Zampini static int inc = 0; 861e0482f5SStefano Zampini static int lev = 0; 871e0482f5SStefano Zampini #endif 881e0482f5SStefano Zampini 891e0482f5SStefano Zampini PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, IS corners, Mat* Gins, Mat* GKins, PetscScalar cvals[2], PetscScalar *work, PetscReal *rwork) 90a13144ffSStefano Zampini { 91a13144ffSStefano Zampini PetscErrorCode ierr; 92a13144ffSStefano Zampini Mat GE,GEd; 93a13144ffSStefano Zampini PetscInt rsize,csize,esize; 94a13144ffSStefano Zampini PetscScalar *ptr; 95a13144ffSStefano Zampini 96a13144ffSStefano Zampini PetscFunctionBegin; 97a13144ffSStefano Zampini ierr = ISGetSize(edge,&esize);CHKERRQ(ierr); 98c3c0e390SStefano Zampini if (!esize) PetscFunctionReturn(0); 99a13144ffSStefano Zampini ierr = ISGetSize(extrow,&rsize);CHKERRQ(ierr); 100a13144ffSStefano Zampini ierr = ISGetSize(extcol,&csize);CHKERRQ(ierr); 101a13144ffSStefano Zampini 102a13144ffSStefano Zampini /* gradients */ 103a13144ffSStefano Zampini ptr = work + 5*esize; 1047dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,extrow,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 105a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,rsize,csize,ptr,Gins);CHKERRQ(ierr); 106a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,Gins);CHKERRQ(ierr); 107a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 108a13144ffSStefano Zampini 109a13144ffSStefano Zampini /* constants */ 110a13144ffSStefano Zampini ptr += rsize*csize; 111a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,esize,csize,ptr,&GEd);CHKERRQ(ierr); 1127dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,edge,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 113a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,&GEd);CHKERRQ(ierr); 114a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 115f498cd09SStefano Zampini ierr = MatDenseOrthogonalRangeOrComplement(GEd,PETSC_FALSE,5*esize,work,rwork,GKins);CHKERRQ(ierr); 116a13144ffSStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1171e0482f5SStefano Zampini 1181e0482f5SStefano Zampini if (corners) { 1191e0482f5SStefano Zampini Mat GEc; 1201e0482f5SStefano Zampini PetscScalar *vals,v; 1211e0482f5SStefano Zampini 1227dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,edge,corners,MAT_INITIAL_MATRIX,&GEc);CHKERRQ(ierr); 1231e0482f5SStefano Zampini ierr = MatTransposeMatMult(GEc,*GKins,MAT_INITIAL_MATRIX,1.0,&GEd);CHKERRQ(ierr); 1241e0482f5SStefano Zampini ierr = MatDenseGetArray(GEd,&vals);CHKERRQ(ierr); 125637e8532SStefano Zampini /* v = PetscAbsScalar(vals[0]) */; 126637e8532SStefano Zampini v = 1.; 1271e0482f5SStefano Zampini cvals[0] = vals[0]/v; 1281e0482f5SStefano Zampini cvals[1] = vals[1]/v; 1291e0482f5SStefano Zampini ierr = MatDenseRestoreArray(GEd,&vals);CHKERRQ(ierr); 1301e0482f5SStefano Zampini ierr = MatScale(*GKins,1./v);CHKERRQ(ierr); 1311e0482f5SStefano Zampini #if defined(PRINT_GDET) 1321e0482f5SStefano Zampini { 1331e0482f5SStefano Zampini PetscViewer viewer; 1341e0482f5SStefano Zampini char filename[256]; 1351e0482f5SStefano Zampini sprintf(filename,"Gdet_l%d_r%d_cc%d.m",lev,PetscGlobalRank,inc++); 1361e0482f5SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1371e0482f5SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1381e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEc,"GEc");CHKERRQ(ierr); 1391e0482f5SStefano Zampini ierr = MatView(GEc,viewer);CHKERRQ(ierr); 1401e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)(*GKins),"GK");CHKERRQ(ierr); 1411e0482f5SStefano Zampini ierr = MatView(*GKins,viewer);CHKERRQ(ierr); 1421e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEd,"Gproj");CHKERRQ(ierr); 1431e0482f5SStefano Zampini ierr = MatView(GEd,viewer);CHKERRQ(ierr); 1441e0482f5SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1451e0482f5SStefano Zampini } 1461e0482f5SStefano Zampini #endif 1471e0482f5SStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1481e0482f5SStefano Zampini ierr = MatDestroy(&GEc);CHKERRQ(ierr); 1491e0482f5SStefano Zampini } 1501e0482f5SStefano Zampini 151a13144ffSStefano Zampini PetscFunctionReturn(0); 152a13144ffSStefano Zampini } 153a13144ffSStefano Zampini 154a13144ffSStefano Zampini PetscErrorCode PCBDDCNedelecSupport(PC pc) 155a13144ffSStefano Zampini { 156a13144ffSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 157a13144ffSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1580569b399SStefano Zampini Mat G,T,conn,lG,lGt,lGis,lGall,lGe,lGinit; 159eee23b56SStefano Zampini Vec tvec; 160a13144ffSStefano Zampini PetscSF sfv; 1611e0482f5SStefano Zampini ISLocalToGlobalMapping el2g,vl2g,fl2g,al2g; 162a13144ffSStefano Zampini MPI_Comm comm; 163c2151214SStefano Zampini IS lned,primals,allprimals,nedfieldlocal; 164c2151214SStefano Zampini IS *eedges,*extrows,*extcols,*alleedges; 1657d871cd7SStefano Zampini PetscBT btv,bte,btvc,btb,btbd,btvcand,btvi,btee,bter; 166a13144ffSStefano Zampini PetscScalar *vals,*work; 167a13144ffSStefano Zampini PetscReal *rwork; 168a13144ffSStefano Zampini const PetscInt *idxs,*ii,*jj,*iit,*jjt; 1691e0482f5SStefano Zampini PetscInt ne,nv,Lv,order,n,field; 170a13144ffSStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 171eee23b56SStefano Zampini PetscInt i,j,extmem,cum,maxsize,nee; 172b03ebc13SStefano Zampini PetscInt *extrow,*extrowcum,*marks,*vmarks,*gidxs; 173a13144ffSStefano Zampini PetscInt *sfvleaves,*sfvroots; 174b03ebc13SStefano Zampini PetscInt *corners,*cedges; 175637e8532SStefano Zampini PetscInt *ecount,**eneighs,*vcount,**vneighs; 176b03ebc13SStefano Zampini #if defined(PETSC_USE_DEBUG) 177b03ebc13SStefano Zampini PetscInt *emarks; 178b03ebc13SStefano Zampini #endif 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); 217a13144ffSStefano Zampini ierr = MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 218a13144ffSStefano Zampini if (!lrc[1]) PetscFunctionReturn(0); 219213b8bfaSStefano Zampini } 220a13144ffSStefano Zampini 221213b8bfaSStefano Zampini /* Get Nedelec field */ 222213b8bfaSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 223c2151214SStefano 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); 224213b8bfaSStefano Zampini if (pcbddc->n_ISForDofsLocal && field >= 0) { 225c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);CHKERRQ(ierr); 226c2151214SStefano Zampini nedfieldlocal = pcbddc->ISForDofsLocal[field]; 227c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 228213b8bfaSStefano Zampini } else if (!pcbddc->n_ISForDofsLocal && field != PETSC_DECIDE) { 229213b8bfaSStefano Zampini ne = n; 230213b8bfaSStefano Zampini nedfieldlocal = NULL; 231213b8bfaSStefano Zampini global = PETSC_TRUE; 232213b8bfaSStefano Zampini } else if (field == PETSC_DECIDE) { 233213b8bfaSStefano Zampini PetscInt rst,ren,*idx; 234213b8bfaSStefano Zampini 235213b8bfaSStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 236213b8bfaSStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 237213b8bfaSStefano Zampini ierr = MatGetOwnershipRange(pcbddc->discretegradient,&rst,&ren);CHKERRQ(ierr); 238213b8bfaSStefano Zampini for (i=rst;i<ren;i++) { 239213b8bfaSStefano Zampini PetscInt nc; 240213b8bfaSStefano Zampini 241213b8bfaSStefano Zampini ierr = MatGetRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 242213b8bfaSStefano Zampini if (nc > 1) matis->sf_rootdata[i-rst] = 1; 243213b8bfaSStefano Zampini ierr = MatRestoreRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 244213b8bfaSStefano Zampini } 245213b8bfaSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 246213b8bfaSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 247213b8bfaSStefano Zampini ierr = PetscMalloc1(n,&idx);CHKERRQ(ierr); 248213b8bfaSStefano Zampini for (i=0,ne=0;i<n;i++) if (matis->sf_leafdata[i]) idx[ne++] = i; 249213b8bfaSStefano Zampini ierr = ISCreateGeneral(comm,ne,idx,PETSC_OWN_POINTER,&nedfieldlocal);CHKERRQ(ierr); 250213b8bfaSStefano Zampini } else { 251213b8bfaSStefano Zampini SETERRQ(comm,PETSC_ERR_USER,"When multiple fields are present, the Nedelec field has to be specified"); 252213b8bfaSStefano Zampini } 253213b8bfaSStefano Zampini 254213b8bfaSStefano Zampini /* Sanity checks */ 255213b8bfaSStefano Zampini if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time"); 256213b8bfaSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis"); 257213b8bfaSStefano 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); 258213b8bfaSStefano Zampini 259213b8bfaSStefano Zampini /* Just set primal dofs and return */ 2601e0482f5SStefano Zampini if (setprimal) { 261eee23b56SStefano Zampini IS enedfieldlocal; 262eee23b56SStefano Zampini PetscInt *eidxs; 263eee23b56SStefano Zampini 264eee23b56SStefano Zampini ierr = PetscMalloc1(ne,&eidxs);CHKERRQ(ierr); 265eee23b56SStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 266213b8bfaSStefano Zampini if (nedfieldlocal) { 267213b8bfaSStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 268eee23b56SStefano Zampini for (i=0,cum=0;i<ne;i++) { 269eee23b56SStefano Zampini if (PetscRealPart(vals[idxs[i]]) > 2.) { 270eee23b56SStefano Zampini eidxs[cum++] = idxs[i]; 271eee23b56SStefano Zampini } 272eee23b56SStefano Zampini } 273eee23b56SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 274213b8bfaSStefano Zampini } else { 275213b8bfaSStefano Zampini for (i=0,cum=0;i<ne;i++) { 276213b8bfaSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 277213b8bfaSStefano Zampini eidxs[cum++] = i; 278213b8bfaSStefano Zampini } 279213b8bfaSStefano Zampini } 280213b8bfaSStefano Zampini } 281213b8bfaSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 282eee23b56SStefano Zampini ierr = ISCreateGeneral(comm,cum,eidxs,PETSC_COPY_VALUES,&enedfieldlocal);CHKERRQ(ierr); 283eee23b56SStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,enedfieldlocal);CHKERRQ(ierr); 284eee23b56SStefano Zampini ierr = PetscFree(eidxs);CHKERRQ(ierr); 2851e0482f5SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 286eee23b56SStefano Zampini ierr = ISDestroy(&enedfieldlocal);CHKERRQ(ierr); 2871e0482f5SStefano Zampini PetscFunctionReturn(0); 2881e0482f5SStefano Zampini } 289a13144ffSStefano Zampini 290213b8bfaSStefano Zampini /* Compute some l2g maps */ 291213b8bfaSStefano Zampini if (nedfieldlocal) { 292c2151214SStefano Zampini IS is; 293c2151214SStefano Zampini 294c2151214SStefano Zampini /* need to map from the local Nedelec field to local numbering */ 295c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);CHKERRQ(ierr); 2961e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering for the whole dofs*/ 2971e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(al2g,nedfieldlocal,&is);CHKERRQ(ierr); 2981e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is,&al2g);CHKERRQ(ierr); 2991e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering (for Nedelec only) */ 3001e0482f5SStefano Zampini if (global) { 3011e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3021e0482f5SStefano Zampini el2g = al2g; 3031e0482f5SStefano Zampini } else { 3041e0482f5SStefano Zampini IS gis; 3051e0482f5SStefano Zampini 3061e0482f5SStefano Zampini ierr = ISRenumber(is,NULL,NULL,&gis);CHKERRQ(ierr); 3071e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(gis,&el2g);CHKERRQ(ierr); 3081e0482f5SStefano Zampini ierr = ISDestroy(&gis);CHKERRQ(ierr); 3091e0482f5SStefano Zampini } 310c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 311c2151214SStefano Zampini } else { 3121e0482f5SStefano Zampini /* restore default */ 3131e0482f5SStefano Zampini pcbddc->nedfield = -1; 3141e0482f5SStefano Zampini /* one ref for the destruction of al2g, one for el2g */ 3151e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3161e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3171e0482f5SStefano Zampini el2g = al2g; 318c2151214SStefano Zampini fl2g = NULL; 319c2151214SStefano Zampini } 320a13144ffSStefano Zampini 321213b8bfaSStefano Zampini /* Start communication to drop connections for interior edges (for cc analysis only) */ 322c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 323c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 324c2151214SStefano Zampini if (nedfieldlocal) { 325c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 326c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1; 327c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 328c2151214SStefano Zampini } else { 329c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1; 330c2151214SStefano Zampini } 331c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 332c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 333213b8bfaSStefano Zampini 334213b8bfaSStefano Zampini if (!singular) { /* drop connections with interior edges to avoid unneeded communications and memory movements */ 335213b8bfaSStefano Zampini ierr = MatDuplicate(pcbddc->discretegradient,MAT_COPY_VALUES,&G);CHKERRQ(ierr); 336213b8bfaSStefano Zampini ierr = MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 3371e0482f5SStefano Zampini if (global) { 3381e0482f5SStefano Zampini PetscInt rst; 3391e0482f5SStefano Zampini 340c2151214SStefano Zampini ierr = MatGetOwnershipRange(G,&rst,NULL);CHKERRQ(ierr); 341c2151214SStefano Zampini for (i=0,cum=0;i<pc->pmat->rmap->n;i++) { 342c2151214SStefano Zampini if (matis->sf_rootdata[i] < 2) { 343c2151214SStefano Zampini matis->sf_rootdata[cum++] = i + rst; 344c2151214SStefano Zampini } 345c2151214SStefano Zampini } 346a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);CHKERRQ(ierr); 347c2151214SStefano Zampini ierr = MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);CHKERRQ(ierr); 3481e0482f5SStefano Zampini } else { 3491e0482f5SStefano Zampini PetscInt *tbz; 3501e0482f5SStefano Zampini 3511e0482f5SStefano Zampini ierr = PetscMalloc1(ne,&tbz);CHKERRQ(ierr); 3521e0482f5SStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 3531e0482f5SStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 3541e0482f5SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3551e0482f5SStefano Zampini for (i=0,cum=0;i<ne;i++) 3561e0482f5SStefano Zampini if (matis->sf_leafdata[idxs[i]] == 1) 3571e0482f5SStefano Zampini tbz[cum++] = i; 3581e0482f5SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3591e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApply(el2g,cum,tbz,tbz);CHKERRQ(ierr); 3601e0482f5SStefano Zampini ierr = MatZeroRows(G,cum,tbz,0.,NULL,NULL);CHKERRQ(ierr); 3611e0482f5SStefano Zampini ierr = PetscFree(tbz);CHKERRQ(ierr); 3621e0482f5SStefano Zampini } 363213b8bfaSStefano Zampini } else { /* we need the entire G to infer the nullspace */ 364213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->discretegradient);CHKERRQ(ierr); 365213b8bfaSStefano Zampini G = pcbddc->discretegradient; 366213b8bfaSStefano Zampini } 367a13144ffSStefano Zampini 368a13144ffSStefano Zampini /* Extract subdomain relevant rows of G */ 369a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(el2g,&idxs);CHKERRQ(ierr); 370a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);CHKERRQ(ierr); 3717dae84e0SHong Zhang ierr = MatCreateSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);CHKERRQ(ierr); 372a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);CHKERRQ(ierr); 373a13144ffSStefano Zampini ierr = ISDestroy(&lned);CHKERRQ(ierr); 374a13144ffSStefano Zampini ierr = MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);CHKERRQ(ierr); 375a13144ffSStefano Zampini ierr = MatDestroy(&lGall);CHKERRQ(ierr); 376a13144ffSStefano Zampini ierr = MatISGetLocalMat(lGis,&lG);CHKERRQ(ierr); 377a13144ffSStefano Zampini 378213b8bfaSStefano Zampini /* SF for nodal dofs communications */ 379c2151214SStefano Zampini ierr = MatGetLocalSize(G,NULL,&Lv);CHKERRQ(ierr); 380a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);CHKERRQ(ierr); 381a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)vl2g);CHKERRQ(ierr); 382a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(vl2g,&nv);CHKERRQ(ierr); 383a13144ffSStefano Zampini ierr = PetscSFCreate(comm,&sfv);CHKERRQ(ierr); 384a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(vl2g,&idxs);CHKERRQ(ierr); 385a13144ffSStefano Zampini ierr = PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);CHKERRQ(ierr); 386a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);CHKERRQ(ierr); 387213b8bfaSStefano Zampini i = singular ? 2 : 1; 388213b8bfaSStefano Zampini ierr = PetscMalloc2(i*nv,&sfvleaves,i*Lv,&sfvroots);CHKERRQ(ierr); 389a13144ffSStefano Zampini 3901e0482f5SStefano Zampini /* Destroy temporary G created in MATIS format and modified G */ 391213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)lG);CHKERRQ(ierr); 392a13144ffSStefano Zampini ierr = MatDestroy(&lGis);CHKERRQ(ierr); 393213b8bfaSStefano Zampini ierr = MatDestroy(&G);CHKERRQ(ierr); 394a13144ffSStefano Zampini 395213b8bfaSStefano Zampini if (print) { 396213b8bfaSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"initial_lG");CHKERRQ(ierr); 397213b8bfaSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 398213b8bfaSStefano Zampini } 399213b8bfaSStefano Zampini 400213b8bfaSStefano Zampini /* Save lG for values insertion in change of basis */ 4010569b399SStefano Zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);CHKERRQ(ierr); 4020569b399SStefano Zampini 403a13144ffSStefano Zampini /* Analyze the edge-nodes connections (duplicate lG) */ 4044e64d54eSstefano_zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGe);CHKERRQ(ierr); 4054e64d54eSstefano_zampini ierr = MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 406a13144ffSStefano Zampini ierr = PetscBTCreate(nv,&btv);CHKERRQ(ierr); 407a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&bte);CHKERRQ(ierr); 408a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&btb);CHKERRQ(ierr); 4097d871cd7SStefano Zampini ierr = PetscBTCreate(ne,&btbd);CHKERRQ(ierr); 410c2151214SStefano Zampini ierr = PetscBTCreate(nv,&btvcand);CHKERRQ(ierr); 411a13144ffSStefano Zampini /* need to import the boundary specification to ensure the 412a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 413a13144ffSStefano Zampini if (pcbddc->DirichletBoundariesLocal) { 414c2151214SStefano Zampini IS is; 415c2151214SStefano Zampini 416c2151214SStefano Zampini if (fl2g) { 417c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);CHKERRQ(ierr); 418c2151214SStefano Zampini } else { 419c2151214SStefano Zampini is = pcbddc->DirichletBoundariesLocal; 420c2151214SStefano Zampini } 421c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 422c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 423a13144ffSStefano Zampini for (i=0;i<cum;i++) { 424a13144ffSStefano Zampini if (idxs[i] >= 0) { 425a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 4267d871cd7SStefano Zampini ierr = PetscBTSet(btbd,idxs[i]);CHKERRQ(ierr); 427a13144ffSStefano Zampini } 428a13144ffSStefano Zampini } 429c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 430c2151214SStefano Zampini if (fl2g) { 431c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 432c2151214SStefano Zampini } 433a13144ffSStefano Zampini } 434a13144ffSStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 435c2151214SStefano Zampini IS is; 436c2151214SStefano Zampini 437c2151214SStefano Zampini if (fl2g) { 438c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);CHKERRQ(ierr); 439c2151214SStefano Zampini } else { 440c2151214SStefano Zampini is = pcbddc->NeumannBoundariesLocal; 441c2151214SStefano Zampini } 442c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 443c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 444a13144ffSStefano Zampini for (i=0;i<cum;i++) { 445a13144ffSStefano Zampini if (idxs[i] >= 0) { 446a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 447a13144ffSStefano Zampini } 448a13144ffSStefano Zampini } 449c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 450c2151214SStefano Zampini if (fl2g) { 451c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 452a13144ffSStefano Zampini } 453c2151214SStefano Zampini } 454c2151214SStefano Zampini 455213b8bfaSStefano Zampini /* Count neighs per dof */ 456637e8532SStefano Zampini ierr = PetscCalloc1(ne,&ecount);CHKERRQ(ierr); 457637e8532SStefano Zampini ierr = PetscMalloc1(ne,&eneighs);CHKERRQ(ierr); 458637e8532SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 459637e8532SStefano Zampini for (i=1,cum=0;i<n_neigh;i++) { 460637e8532SStefano Zampini cum += n_shared[i]; 461637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 462637e8532SStefano Zampini ecount[shared[i][j]]++; 463637e8532SStefano Zampini } 464637e8532SStefano Zampini } 465637e8532SStefano Zampini if (ne) { 466637e8532SStefano Zampini ierr = PetscMalloc1(cum,&eneighs[0]);CHKERRQ(ierr); 467637e8532SStefano Zampini } 468637e8532SStefano Zampini for (i=1;i<ne;i++) eneighs[i] = eneighs[i-1] + ecount[i-1]; 469637e8532SStefano Zampini ierr = PetscMemzero(ecount,ne*sizeof(PetscInt));CHKERRQ(ierr); 470637e8532SStefano Zampini for (i=1;i<n_neigh;i++) { 471637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 472637e8532SStefano Zampini PetscInt k = shared[i][j]; 473637e8532SStefano Zampini eneighs[k][ecount[k]] = neigh[i]; 474637e8532SStefano Zampini ecount[k]++; 475637e8532SStefano Zampini } 476637e8532SStefano Zampini } 477637e8532SStefano Zampini for (i=0;i<ne;i++) { 478637e8532SStefano Zampini ierr = PetscSortRemoveDupsInt(&ecount[i],eneighs[i]);CHKERRQ(ierr); 479637e8532SStefano Zampini } 480637e8532SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 481637e8532SStefano Zampini ierr = PetscCalloc1(nv,&vcount);CHKERRQ(ierr); 482637e8532SStefano Zampini ierr = PetscMalloc1(nv,&vneighs);CHKERRQ(ierr); 483637e8532SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 484637e8532SStefano Zampini for (i=1,cum=0;i<n_neigh;i++) { 485637e8532SStefano Zampini cum += n_shared[i]; 486637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 487637e8532SStefano Zampini vcount[shared[i][j]]++; 488637e8532SStefano Zampini } 489637e8532SStefano Zampini } 490637e8532SStefano Zampini if (nv) { 491637e8532SStefano Zampini ierr = PetscMalloc1(cum,&vneighs[0]);CHKERRQ(ierr); 492637e8532SStefano Zampini } 493637e8532SStefano Zampini for (i=1;i<nv;i++) vneighs[i] = vneighs[i-1] + vcount[i-1]; 494637e8532SStefano Zampini ierr = PetscMemzero(vcount,nv*sizeof(PetscInt));CHKERRQ(ierr); 495637e8532SStefano Zampini for (i=1;i<n_neigh;i++) { 496637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 497637e8532SStefano Zampini PetscInt k = shared[i][j]; 498637e8532SStefano Zampini vneighs[k][vcount[k]] = neigh[i]; 499637e8532SStefano Zampini vcount[k]++; 500637e8532SStefano Zampini } 501637e8532SStefano Zampini } 502637e8532SStefano Zampini for (i=0;i<nv;i++) { 503637e8532SStefano Zampini ierr = PetscSortRemoveDupsInt(&vcount[i],vneighs[i]);CHKERRQ(ierr); 504637e8532SStefano Zampini } 505637e8532SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 506637e8532SStefano Zampini 5077d871cd7SStefano Zampini /* need to remove coarse faces' dofs and coarse edges' dirichlet dofs 5087d871cd7SStefano Zampini for proper detection of coarse edges' endpoints */ 50962b0c6f7SStefano Zampini ierr = PetscBTCreate(ne,&btee);CHKERRQ(ierr); 51062b0c6f7SStefano Zampini for (i=0;i<ne;i++) { 5117d871cd7SStefano Zampini if ((ecount[i] > 1 && !PetscBTLookup(btbd,i)) || (ecount[i] == 1 && PetscBTLookup(btb,i))) { 51262b0c6f7SStefano Zampini ierr = PetscBTSet(btee,i);CHKERRQ(ierr); 51362b0c6f7SStefano Zampini } 51462b0c6f7SStefano Zampini } 515637e8532SStefano Zampini ierr = PetscMalloc1(ne,&marks);CHKERRQ(ierr); 51662b0c6f7SStefano Zampini if (!conforming) { 51762b0c6f7SStefano Zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 51862b0c6f7SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 51962b0c6f7SStefano Zampini } 5204e64d54eSstefano_zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 521dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGe,&vals);CHKERRQ(ierr); 52262b0c6f7SStefano Zampini cum = 0; 523a13144ffSStefano Zampini for (i=0;i<ne;i++) { 524dec27d64SStefano Zampini /* eliminate rows corresponding to edge dofs belonging to coarse faces */ 52562b0c6f7SStefano Zampini if (!PetscBTLookup(btee,i)) { 526a13144ffSStefano Zampini marks[cum++] = i; 527dec27d64SStefano Zampini continue; 528dec27d64SStefano Zampini } 529dec27d64SStefano Zampini /* set badly connected edge dofs as primal */ 53062b0c6f7SStefano Zampini if (!conforming) { 53162b0c6f7SStefano Zampini if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */ 532a13144ffSStefano Zampini marks[cum++] = i; 533a13144ffSStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 534a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 535a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 536a13144ffSStefano Zampini } 53762b0c6f7SStefano Zampini } else { 53862b0c6f7SStefano Zampini /* every edge dofs should be connected trough a certain number of nodal dofs 53962b0c6f7SStefano Zampini to other edge dofs belonging to coarse edges 54062b0c6f7SStefano Zampini - at most 2 endpoints 54162b0c6f7SStefano Zampini - order-1 interior nodal dofs 54262b0c6f7SStefano Zampini - no undefined nodal dofs (nconn < order) 54362b0c6f7SStefano Zampini */ 54462b0c6f7SStefano Zampini PetscInt ends = 0,ints = 0, undef = 0; 54562b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 54662b0c6f7SStefano Zampini PetscInt v = jj[j],k; 54762b0c6f7SStefano Zampini PetscInt nconn = iit[v+1]-iit[v]; 54862b0c6f7SStefano Zampini for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--; 54962b0c6f7SStefano Zampini if (nconn > order) ends++; 55062b0c6f7SStefano Zampini else if (nconn == order) ints++; 55162b0c6f7SStefano Zampini else undef++; 55262b0c6f7SStefano Zampini } 55362b0c6f7SStefano Zampini if (undef || ends > 2 || ints != order -1) { 55462b0c6f7SStefano Zampini marks[cum++] = i; 55562b0c6f7SStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 55662b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 55762b0c6f7SStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 55862b0c6f7SStefano Zampini } 55962b0c6f7SStefano Zampini } 56062b0c6f7SStefano Zampini } 561a13144ffSStefano Zampini } 562dec27d64SStefano Zampini /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */ 563dec27d64SStefano Zampini if (!order && ii[i+1] != ii[i]) { 564dec27d64SStefano Zampini PetscScalar val = 1./(ii[i+1]-ii[i]-1); 565dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vals[j] = val; 566a13144ffSStefano Zampini } 567dec27d64SStefano Zampini } 56862b0c6f7SStefano Zampini ierr = PetscBTDestroy(&btee);CHKERRQ(ierr); 569dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGe,&vals);CHKERRQ(ierr); 5704e64d54eSstefano_zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 57162b0c6f7SStefano Zampini if (!conforming) { 57262b0c6f7SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 57362b0c6f7SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 57462b0c6f7SStefano Zampini } 5754e64d54eSstefano_zampini ierr = MatZeroRows(lGe,cum,marks,0.,NULL,NULL);CHKERRQ(ierr); 576637e8532SStefano Zampini 577b03ebc13SStefano Zampini /* identify splitpoints and corner candidates */ 5784e64d54eSstefano_zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 579a13144ffSStefano Zampini if (print) { 5804e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");CHKERRQ(ierr); 5814e64d54eSstefano_zampini ierr = MatView(lGe,NULL);CHKERRQ(ierr); 5824e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");CHKERRQ(ierr); 583a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 584a13144ffSStefano Zampini } 585a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 586dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGt,&vals);CHKERRQ(ierr); 587a13144ffSStefano Zampini for (i=0;i<nv;i++) { 588637e8532SStefano Zampini PetscInt ord = order, test = ii[i+1]-ii[i], vc = vcount[i]; 5897d871cd7SStefano Zampini PetscBool sneighs = PETSC_TRUE, bdir = PETSC_FALSE; 590b03ebc13SStefano Zampini if (!order) { /* variable order */ 591dec27d64SStefano Zampini PetscReal vorder = 0.; 592dec27d64SStefano Zampini 593dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]); 594dec27d64SStefano Zampini test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON); 595dec27d64SStefano Zampini if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%d)",vorder,test); 596dec27d64SStefano Zampini ord = 1; 597dec27d64SStefano Zampini } 598a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 599dec27d64SStefano Zampini if (test%ord) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected number of edge dofs %d connected with nodal dof %d with order %d",test,i,ord); 600a13144ffSStefano Zampini #endif 601637e8532SStefano Zampini for (j=ii[i];j<ii[i+1] && sneighs;j++) { 6027d871cd7SStefano Zampini if (PetscBTLookup(btbd,jj[j])) { 6037d871cd7SStefano Zampini bdir = PETSC_TRUE; 6047d871cd7SStefano Zampini break; 6057d871cd7SStefano Zampini } 606637e8532SStefano Zampini if (vc != ecount[jj[j]]) { 607637e8532SStefano Zampini sneighs = PETSC_FALSE; 608637e8532SStefano Zampini } else { 609637e8532SStefano Zampini PetscInt k,*vn = vneighs[i], *en = eneighs[jj[j]]; 610637e8532SStefano Zampini for (k=0;k<vc;k++) { 611637e8532SStefano Zampini if (vn[k] != en[k]) { 612637e8532SStefano Zampini sneighs = PETSC_FALSE; 613637e8532SStefano Zampini break; 614637e8532SStefano Zampini } 615637e8532SStefano Zampini } 616637e8532SStefano Zampini } 617637e8532SStefano Zampini } 6187d871cd7SStefano Zampini if (!sneighs || test >= 3*ord || bdir) { /* splitpoints */ 6197d871cd7SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %d (%d %d %d)\n",i,!sneighs,test >= 3*ord,bdir); 620a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 621dec27d64SStefano Zampini } else if (test == ord) { 622b03ebc13SStefano Zampini if (order == 1 || (!order && ii[i+1]-ii[i] == 1)) { 623a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %d\n",i); 624a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 625a13144ffSStefano Zampini } else { 626a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %d\n",i); 627a13144ffSStefano Zampini ierr = PetscBTSet(btvcand,i);CHKERRQ(ierr); 628a13144ffSStefano Zampini } 629a13144ffSStefano Zampini } 630a13144ffSStefano Zampini } 631637e8532SStefano Zampini ierr = PetscFree(ecount);CHKERRQ(ierr); 632637e8532SStefano Zampini ierr = PetscFree(vcount);CHKERRQ(ierr); 633637e8532SStefano Zampini if (ne) { 634637e8532SStefano Zampini ierr = PetscFree(eneighs[0]);CHKERRQ(ierr); 635637e8532SStefano Zampini } 636637e8532SStefano Zampini if (nv) { 637637e8532SStefano Zampini ierr = PetscFree(vneighs[0]);CHKERRQ(ierr); 638637e8532SStefano Zampini } 639637e8532SStefano Zampini ierr = PetscFree(eneighs);CHKERRQ(ierr); 640637e8532SStefano Zampini ierr = PetscFree(vneighs);CHKERRQ(ierr); 6417d871cd7SStefano Zampini ierr = PetscBTDestroy(&btbd);CHKERRQ(ierr); 642b03ebc13SStefano Zampini 643b03ebc13SStefano Zampini /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */ 644b03ebc13SStefano Zampini if (order != 1) { 645b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"INSPECTING CANDIDATES\n"); 646b03ebc13SStefano Zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 647b03ebc13SStefano Zampini for (i=0;i<nv;i++) { 648b03ebc13SStefano Zampini if (PetscBTLookup(btvcand,i)) { 649b03ebc13SStefano Zampini PetscBool found = PETSC_FALSE; 650b03ebc13SStefano Zampini for (j=ii[i];j<ii[i+1] && !found;j++) { 651b03ebc13SStefano Zampini PetscInt k,e = jj[j]; 652b03ebc13SStefano Zampini if (PetscBTLookup(bte,e)) continue; 653b03ebc13SStefano Zampini for (k=iit[e];k<iit[e+1];k++) { 654b03ebc13SStefano Zampini PetscInt v = jjt[k]; 655b03ebc13SStefano Zampini if (v != i && PetscBTLookup(btvcand,v)) { 656b03ebc13SStefano Zampini found = PETSC_TRUE; 657b03ebc13SStefano Zampini break; 658b03ebc13SStefano Zampini } 659b03ebc13SStefano Zampini } 660b03ebc13SStefano Zampini } 661b03ebc13SStefano Zampini if (!found) { 662b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d CLEARED\n",i); 663b03ebc13SStefano Zampini ierr = PetscBTClear(btvcand,i);CHKERRQ(ierr); 664b03ebc13SStefano Zampini } else { 665b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d ACCEPTED\n",i); 666b03ebc13SStefano Zampini } 667b03ebc13SStefano Zampini } 668b03ebc13SStefano Zampini } 669b03ebc13SStefano Zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 670b03ebc13SStefano Zampini } 671dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGt,&vals);CHKERRQ(ierr); 672a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 673b03ebc13SStefano Zampini ierr = MatDestroy(&lGe);CHKERRQ(ierr); 674a13144ffSStefano Zampini 675a13144ffSStefano Zampini /* Get the local G^T explicitly */ 6760569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 677a13144ffSStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 6784e64d54eSstefano_zampini ierr = MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 679a13144ffSStefano Zampini 6804e64d54eSstefano_zampini /* Mark interior nodal dofs */ 681a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 6824e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvi);CHKERRQ(ierr); 683a13144ffSStefano Zampini for (i=1;i<n_neigh;i++) { 684a13144ffSStefano Zampini for (j=0;j<n_shared[i];j++) { 6854e64d54eSstefano_zampini ierr = PetscBTSet(btvi,shared[i][j]);CHKERRQ(ierr); 686a13144ffSStefano Zampini } 687a13144ffSStefano Zampini } 688a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 689a13144ffSStefano Zampini 690a13144ffSStefano Zampini /* communicate corners and splitpoints */ 691a13144ffSStefano Zampini ierr = PetscMalloc1(nv,&vmarks);CHKERRQ(ierr); 692a13144ffSStefano Zampini ierr = PetscMemzero(sfvleaves,nv*sizeof(PetscInt));CHKERRQ(ierr); 693a13144ffSStefano Zampini ierr = PetscMemzero(sfvroots,Lv*sizeof(PetscInt));CHKERRQ(ierr); 694a13144ffSStefano Zampini for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1; 695a13144ffSStefano Zampini 696a13144ffSStefano Zampini if (print) { 697a13144ffSStefano Zampini IS tbz; 698a13144ffSStefano Zampini 699a13144ffSStefano Zampini cum = 0; 700a13144ffSStefano Zampini for (i=0;i<nv;i++) 701a13144ffSStefano Zampini if (sfvleaves[i]) 702a13144ffSStefano Zampini vmarks[cum++] = i; 703a13144ffSStefano Zampini 704a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 705a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");CHKERRQ(ierr); 706a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 707a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 708a13144ffSStefano Zampini } 709a13144ffSStefano Zampini 710a13144ffSStefano Zampini ierr = PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 711a13144ffSStefano Zampini ierr = PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 712a13144ffSStefano Zampini ierr = PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 713a13144ffSStefano Zampini ierr = PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 714a13144ffSStefano Zampini 7154e64d54eSstefano_zampini /* Zero rows of lGt corresponding to identified corners 7164e64d54eSstefano_zampini and interior nodal dofs */ 717a13144ffSStefano Zampini cum = 0; 718a13144ffSStefano Zampini for (i=0;i<nv;i++) { 719a13144ffSStefano Zampini if (sfvleaves[i]) { 720a13144ffSStefano Zampini vmarks[cum++] = i; 721a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 722a13144ffSStefano Zampini } 7234e64d54eSstefano_zampini if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i; 724a13144ffSStefano Zampini } 7254e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvi);CHKERRQ(ierr); 726a13144ffSStefano Zampini if (print) { 727a13144ffSStefano Zampini IS tbz; 728a13144ffSStefano Zampini 729a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 7304e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");CHKERRQ(ierr); 731a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 732a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 733a13144ffSStefano Zampini } 734a13144ffSStefano Zampini ierr = MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);CHKERRQ(ierr); 735a13144ffSStefano Zampini ierr = PetscFree(vmarks);CHKERRQ(ierr); 736a13144ffSStefano Zampini ierr = PetscSFDestroy(&sfv);CHKERRQ(ierr); 737a13144ffSStefano Zampini ierr = PetscFree2(sfvleaves,sfvroots);CHKERRQ(ierr); 738a13144ffSStefano Zampini 739a13144ffSStefano Zampini /* Recompute G */ 740a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 741a13144ffSStefano Zampini ierr = MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);CHKERRQ(ierr); 742a13144ffSStefano Zampini if (print) { 743a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"used_lG");CHKERRQ(ierr); 744a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 745a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lGt,"used_lGt");CHKERRQ(ierr); 746a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 747a13144ffSStefano Zampini } 748a13144ffSStefano Zampini 749a13144ffSStefano Zampini /* Get primal dofs (if any) */ 750a13144ffSStefano Zampini cum = 0; 751a13144ffSStefano Zampini for (i=0;i<ne;i++) { 752a13144ffSStefano Zampini if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i; 753a13144ffSStefano Zampini } 754c2151214SStefano Zampini if (fl2g) { 755c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);CHKERRQ(ierr); 756c2151214SStefano Zampini } 757a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 758a13144ffSStefano Zampini if (print) { 759a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");CHKERRQ(ierr); 760a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 761a13144ffSStefano Zampini } 762a13144ffSStefano Zampini ierr = PetscBTDestroy(&bte);CHKERRQ(ierr); 763c2151214SStefano Zampini /* TODO: what if the user passed in some of them ? */ 764a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 765a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 766a13144ffSStefano Zampini 767a13144ffSStefano Zampini /* Compute edge connectivity */ 768a13144ffSStefano Zampini ierr = PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");CHKERRQ(ierr); 769a13144ffSStefano Zampini ierr = MatMatMultSymbolic(lG,lGt,PETSC_DEFAULT,&conn);CHKERRQ(ierr); 770a13144ffSStefano Zampini ierr = MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 771c2151214SStefano Zampini if (fl2g) { 772c2151214SStefano Zampini PetscBT btf; 773c2151214SStefano Zampini PetscInt *iia,*jja,*iiu,*jju; 774c2151214SStefano Zampini PetscBool rest = PETSC_FALSE,free = PETSC_FALSE; 775c2151214SStefano Zampini 776c2151214SStefano Zampini /* create CSR for all local dofs */ 777c2151214SStefano Zampini ierr = PetscMalloc1(n+1,&iia);CHKERRQ(ierr); 778c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */ 779c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid size of CSR graph %d. Should be %d\n",pcbddc->mat_graph->nvtxs_csr,n); 780c2151214SStefano Zampini iiu = pcbddc->mat_graph->xadj; 781c2151214SStefano Zampini jju = pcbddc->mat_graph->adjncy; 782c2151214SStefano Zampini } else if (pcbddc->use_local_adj) { 783c2151214SStefano Zampini rest = PETSC_TRUE; 784c2151214SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 785c2151214SStefano Zampini } else { 786c2151214SStefano Zampini free = PETSC_TRUE; 787c2151214SStefano Zampini ierr = PetscMalloc2(n+1,&iiu,n,&jju);CHKERRQ(ierr); 788c2151214SStefano Zampini iiu[0] = 0; 789c2151214SStefano Zampini for (i=0;i<n;i++) { 790c2151214SStefano Zampini iiu[i+1] = i+1; 791c2151214SStefano Zampini jju[i] = -1; 792d904f53bSStefano Zampini } 793c2151214SStefano Zampini } 794c2151214SStefano Zampini 795c2151214SStefano Zampini /* import sizes of CSR */ 796c2151214SStefano Zampini iia[0] = 0; 797c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i]; 798c2151214SStefano Zampini 799c2151214SStefano Zampini /* overwrite entries corresponding to the Nedelec field */ 800c2151214SStefano Zampini ierr = PetscBTCreate(n,&btf);CHKERRQ(ierr); 801c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 802c2151214SStefano Zampini for (i=0;i<ne;i++) { 803c2151214SStefano Zampini ierr = PetscBTSet(btf,idxs[i]);CHKERRQ(ierr); 804c2151214SStefano Zampini iia[idxs[i]+1] = ii[i+1]-ii[i]; 805c2151214SStefano Zampini } 806c2151214SStefano Zampini 807c2151214SStefano Zampini /* iia in CSR */ 808c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] += iia[i]; 809c2151214SStefano Zampini 810c2151214SStefano Zampini /* jja in CSR */ 811c2151214SStefano Zampini ierr = PetscMalloc1(iia[n],&jja);CHKERRQ(ierr); 812c2151214SStefano Zampini for (i=0;i<n;i++) 813c2151214SStefano Zampini if (!PetscBTLookup(btf,i)) 814c2151214SStefano Zampini for (j=0;j<iiu[i+1]-iiu[i];j++) 815c2151214SStefano Zampini jja[iia[i]+j] = jju[iiu[i]+j]; 816c2151214SStefano Zampini 817c2151214SStefano Zampini /* map edge dofs connectivity */ 8181e0482f5SStefano Zampini if (jj) { 819c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);CHKERRQ(ierr); 820c2151214SStefano Zampini for (i=0;i<ne;i++) { 821c2151214SStefano Zampini PetscInt e = idxs[i]; 822c2151214SStefano Zampini for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j]; 823c2151214SStefano Zampini } 8241e0482f5SStefano Zampini } 825c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 826c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);CHKERRQ(ierr); 827c2151214SStefano Zampini if (rest) { 828c2151214SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 829c2151214SStefano Zampini } 830c2151214SStefano Zampini if (free) { 831c2151214SStefano Zampini ierr = PetscFree2(iiu,jju);CHKERRQ(ierr); 832c2151214SStefano Zampini } 833c2151214SStefano Zampini ierr = PetscBTDestroy(&btf);CHKERRQ(ierr); 834c2151214SStefano Zampini } else { 835c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);CHKERRQ(ierr); 836c2151214SStefano Zampini } 837c2151214SStefano Zampini 838a13144ffSStefano Zampini /* Analyze interface for edge dofs */ 839a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 840213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 841a13144ffSStefano Zampini 842a13144ffSStefano Zampini /* Get coarse edges in the edge space */ 843c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 844a13144ffSStefano Zampini ierr = MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 845a13144ffSStefano Zampini 846c2151214SStefano Zampini if (fl2g) { 847c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 848c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 849c2151214SStefano Zampini for (i=0;i<nee;i++) { 850c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 851c2151214SStefano Zampini } 852c2151214SStefano Zampini } else { 853c2151214SStefano Zampini eedges = alleedges; 854c2151214SStefano Zampini primals = allprimals; 855c2151214SStefano Zampini } 856c2151214SStefano Zampini 857a13144ffSStefano Zampini /* Mark fine edge dofs with their coarse edge id */ 858a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 859c2151214SStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 860c2151214SStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 861c2151214SStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = nee+1; 862c2151214SStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 863c2151214SStefano Zampini if (print) { 864c2151214SStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");CHKERRQ(ierr); 865c2151214SStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 866c2151214SStefano Zampini } 867c2151214SStefano Zampini 868c2151214SStefano Zampini maxsize = 0; 869a13144ffSStefano Zampini for (i=0;i<nee;i++) { 870a13144ffSStefano Zampini PetscInt size,mark = i+1; 871a13144ffSStefano Zampini 872a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 873a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 874a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 875a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 876a13144ffSStefano Zampini maxsize = PetscMax(maxsize,size); 877a13144ffSStefano Zampini } 878a13144ffSStefano Zampini 879a13144ffSStefano Zampini /* Find coarse edge endpoints */ 880a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 881a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 882a13144ffSStefano Zampini for (i=0;i<nee;i++) { 883a13144ffSStefano Zampini PetscInt mark = i+1,size; 884a13144ffSStefano Zampini 885a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 8861e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 8871e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 888a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 889a13144ffSStefano Zampini if (print) { 890a13144ffSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %d\n",i); 891a13144ffSStefano Zampini ISView(eedges[i],NULL); 892a13144ffSStefano Zampini } 893a13144ffSStefano Zampini for (j=0;j<size;j++) { 894a13144ffSStefano Zampini PetscInt k, ee = idxs[j]; 895a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," idx %d\n",ee); 896a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 897a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," inspect %d\n",jj[k]); 898a13144ffSStefano Zampini if (PetscBTLookup(btv,jj[k])) { 899a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %d\n",jj[k]); 900a13144ffSStefano Zampini } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */ 901a13144ffSStefano Zampini PetscInt k2; 902a13144ffSStefano Zampini PetscBool corner = PETSC_FALSE; 903a13144ffSStefano Zampini for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) { 904c2151214SStefano 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])); 905c2151214SStefano Zampini /* it's a corner if either is connected with an edge dof belonging to a different cc or 906c2151214SStefano Zampini if the edge dof lie on the natural part of the boundary */ 907c2151214SStefano Zampini if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) { 908a13144ffSStefano Zampini corner = PETSC_TRUE; 909a13144ffSStefano Zampini break; 910a13144ffSStefano Zampini } 911a13144ffSStefano Zampini } 912a13144ffSStefano Zampini if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */ 913a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found %d\n",jj[k]); 914a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 915a13144ffSStefano Zampini } else { 916a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n"); 917a13144ffSStefano Zampini } 918a13144ffSStefano Zampini } 919a13144ffSStefano Zampini } 920a13144ffSStefano Zampini } 921a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 922a13144ffSStefano Zampini } 923a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 924a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 925c2151214SStefano Zampini ierr = PetscBTDestroy(&btb);CHKERRQ(ierr); 926a13144ffSStefano Zampini 927a13144ffSStefano Zampini /* Reset marked primal dofs */ 928a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 929a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 930a13144ffSStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = 0; 931a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 932a13144ffSStefano Zampini 9330569b399SStefano Zampini /* Now use the initial lG */ 9340569b399SStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 9350569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 9360569b399SStefano Zampini lG = lGinit; 9370569b399SStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 9380569b399SStefano Zampini 939a13144ffSStefano Zampini /* Compute extended cols indices */ 940b03ebc13SStefano Zampini ierr = PetscBTCreate(nv,&btvc);CHKERRQ(ierr); 941b03ebc13SStefano Zampini ierr = PetscBTCreate(nee,&bter);CHKERRQ(ierr); 942a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 943a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lG,&i);CHKERRQ(ierr); 944a13144ffSStefano Zampini i *= maxsize; 945b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 946a13144ffSStefano Zampini ierr = PetscMalloc2(i,&extrow,i,&gidxs);CHKERRQ(ierr); 947a13144ffSStefano Zampini eerr = PETSC_FALSE; 948a13144ffSStefano Zampini for (i=0;i<nee;i++) { 949b03ebc13SStefano Zampini PetscInt size,found = 0; 950a13144ffSStefano Zampini 951a13144ffSStefano Zampini cum = 0; 952a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 9531e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 9541e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 955a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 956b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 957a13144ffSStefano Zampini for (j=0;j<size;j++) { 958a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 959b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 960b03ebc13SStefano Zampini PetscInt vv = jj[k]; 961b03ebc13SStefano Zampini if (!PetscBTLookup(btv,vv)) extrow[cum++] = vv; 962b03ebc13SStefano Zampini else if (!PetscBTLookupSet(btvc,vv)) found++; 963b03ebc13SStefano Zampini } 964a13144ffSStefano Zampini } 965a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 966a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 967a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 968a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 969a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 970a13144ffSStefano Zampini /* it may happen that endpoints are not defined at this point 971a13144ffSStefano Zampini if it is the case, mark this edge for a second pass */ 972b03ebc13SStefano Zampini if (cum != size -1 || found != 2) { 973b03ebc13SStefano Zampini ierr = PetscBTSet(bter,i);CHKERRQ(ierr); 974a13144ffSStefano Zampini if (print) { 975a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge");CHKERRQ(ierr); 976a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 977a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol");CHKERRQ(ierr); 978a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 979a13144ffSStefano Zampini } 980a13144ffSStefano Zampini eerr = PETSC_TRUE; 981a13144ffSStefano Zampini } 982a13144ffSStefano Zampini } 9834e64d54eSstefano_zampini /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */ 984a13144ffSStefano Zampini ierr = MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 985a13144ffSStefano Zampini if (done) { 986a13144ffSStefano Zampini PetscInt *newprimals; 987a13144ffSStefano Zampini 988a13144ffSStefano Zampini ierr = PetscMalloc1(ne,&newprimals);CHKERRQ(ierr); 989a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 990a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 991a13144ffSStefano Zampini ierr = PetscMemcpy(newprimals,idxs,cum*sizeof(PetscInt));CHKERRQ(ierr); 992a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 9930569b399SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 994b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"DOING SECOND PASS (eerr %d)\n",eerr); 995a13144ffSStefano Zampini for (i=0;i<nee;i++) { 996b03ebc13SStefano Zampini PetscBool has_candidates = PETSC_FALSE; 997b03ebc13SStefano Zampini if (PetscBTLookup(bter,i)) { 998a13144ffSStefano Zampini PetscInt size,mark = i+1; 999a13144ffSStefano Zampini 1000a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 1001a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1002c2151214SStefano Zampini /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */ 1003a13144ffSStefano Zampini for (j=0;j<size;j++) { 1004a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 1005b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"Inspecting edge dof %d [%d %d)\n",ee,ii[ee],ii[ee+1]); 1006a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 1007a13144ffSStefano Zampini /* set all candidates located on the edge as corners */ 1008a13144ffSStefano Zampini if (PetscBTLookup(btvcand,jj[k])) { 1009a13144ffSStefano Zampini PetscInt k2,vv = jj[k]; 1010b03ebc13SStefano Zampini has_candidates = PETSC_TRUE; 1011b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Candidate set to vertex %d\n",vv); 1012a13144ffSStefano Zampini ierr = PetscBTSet(btv,vv);CHKERRQ(ierr); 1013a13144ffSStefano Zampini /* set all edge dofs connected to candidate as primals */ 1014a13144ffSStefano Zampini for (k2=iit[vv];k2<iit[vv+1];k2++) { 1015a13144ffSStefano Zampini if (marks[jjt[k2]] == mark) { 1016a13144ffSStefano Zampini PetscInt k3,ee2 = jjt[k2]; 1017b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected edge dof set to primal %d\n",ee2); 1018a13144ffSStefano Zampini newprimals[cum++] = ee2; 1019a13144ffSStefano Zampini /* finally set the new corners */ 1020a13144ffSStefano Zampini for (k3=ii[ee2];k3<ii[ee2+1];k3++) { 1021b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected nodal dof set to vertex %d\n",jj[k3]); 1022a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k3]);CHKERRQ(ierr); 1023a13144ffSStefano Zampini } 1024a13144ffSStefano Zampini } 1025a13144ffSStefano Zampini } 1026b03ebc13SStefano Zampini } else { 1027b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Not a candidate vertex %d\n",jj[k]); 1028a13144ffSStefano Zampini } 1029a13144ffSStefano Zampini } 1030a13144ffSStefano Zampini } 1031b03ebc13SStefano Zampini if (!has_candidates) { /* circular edge */ 1032b03ebc13SStefano Zampini PetscInt k, ee = idxs[0],*tmarks; 1033b03ebc13SStefano Zampini 1034b03ebc13SStefano Zampini ierr = PetscCalloc1(ne,&tmarks);CHKERRQ(ierr); 1035b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Circular edge %d\n",i); 1036b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 1037b03ebc13SStefano Zampini PetscInt k2; 1038b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Set to corner %d\n",jj[k]); 1039b03ebc13SStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 1040b03ebc13SStefano Zampini for (k2=iit[jj[k]];k2<iit[jj[k]+1];k2++) tmarks[jjt[k2]]++; 1041b03ebc13SStefano Zampini } 1042b03ebc13SStefano Zampini for (j=0;j<size;j++) { 1043b03ebc13SStefano Zampini if (tmarks[idxs[j]] > 1) { 1044b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Edge dof set to primal %d\n",idxs[j]); 1045b03ebc13SStefano Zampini newprimals[cum++] = idxs[j]; 1046b03ebc13SStefano Zampini } 1047b03ebc13SStefano Zampini } 1048b03ebc13SStefano Zampini ierr = PetscFree(tmarks);CHKERRQ(ierr); 1049b03ebc13SStefano Zampini } 1050a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1051a13144ffSStefano Zampini } 1052a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 1053a13144ffSStefano Zampini } 1054b03ebc13SStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 10550569b399SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 1056a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,newprimals);CHKERRQ(ierr); 1057c2151214SStefano Zampini if (fl2g) { 1058c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);CHKERRQ(ierr); 1059c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1060c2151214SStefano Zampini for (i=0;i<nee;i++) { 1061c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1062c2151214SStefano Zampini } 1063c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1064c2151214SStefano Zampini } 1065c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1066a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 1067a13144ffSStefano Zampini ierr = PetscFree(newprimals);CHKERRQ(ierr); 1068a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 1069a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1070a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 1071213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 1072c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1073c2151214SStefano Zampini if (fl2g) { 1074c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 1075c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 1076c2151214SStefano Zampini for (i=0;i<nee;i++) { 1077c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 1078c2151214SStefano Zampini } 1079c2151214SStefano Zampini } else { 1080c2151214SStefano Zampini eedges = alleedges; 1081c2151214SStefano Zampini primals = allprimals; 1082c2151214SStefano Zampini } 1083b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 1084a13144ffSStefano Zampini 1085a13144ffSStefano Zampini /* Mark again */ 1086a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 1087a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1088a13144ffSStefano Zampini PetscInt size,mark = i+1; 1089a13144ffSStefano Zampini 1090a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 1091a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1092a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 1093a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1094a13144ffSStefano Zampini } 1095a13144ffSStefano Zampini if (print) { 1096a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");CHKERRQ(ierr); 1097a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 1098a13144ffSStefano Zampini } 1099a13144ffSStefano Zampini 1100a13144ffSStefano Zampini /* Recompute extended cols */ 1101a13144ffSStefano Zampini eerr = PETSC_FALSE; 1102a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1103a13144ffSStefano Zampini PetscInt size; 1104a13144ffSStefano Zampini 1105a13144ffSStefano Zampini cum = 0; 1106a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 11071e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 11081e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 1109a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1110a13144ffSStefano Zampini for (j=0;j<size;j++) { 1111a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 11121e0482f5SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) if (!PetscBTLookup(btv,jj[k])) extrow[cum++] = jj[k]; 1113a13144ffSStefano Zampini } 1114a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1115a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 1116a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 1117a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 1118a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 1119a13144ffSStefano Zampini if (cum != size -1) { 1120a13144ffSStefano Zampini if (print) { 1121a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");CHKERRQ(ierr); 1122a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 1123a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");CHKERRQ(ierr); 1124a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 1125a13144ffSStefano Zampini } 1126a13144ffSStefano Zampini eerr = PETSC_TRUE; 1127a13144ffSStefano Zampini } 1128a13144ffSStefano Zampini } 1129a13144ffSStefano Zampini } 1130a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1131a13144ffSStefano Zampini ierr = PetscFree2(extrow,gidxs);CHKERRQ(ierr); 1132b03ebc13SStefano Zampini ierr = PetscBTDestroy(&bter);CHKERRQ(ierr); 11337d871cd7SStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 1134a13144ffSStefano Zampini /* an error should not occur at this point */ 1135a13144ffSStefano Zampini if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS"); 1136a13144ffSStefano Zampini 11374e64d54eSstefano_zampini /* Check the number of endpoints */ 11380569b399SStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1139b03ebc13SStefano Zampini ierr = PetscMalloc1(2*nee,&corners);CHKERRQ(ierr); 1140b03ebc13SStefano Zampini ierr = PetscMalloc1(nee,&cedges);CHKERRQ(ierr); 11414e64d54eSstefano_zampini for (i=0;i<nee;i++) { 1142b03ebc13SStefano Zampini PetscInt size, found = 0, gc[2]; 11434e64d54eSstefano_zampini 1144b03ebc13SStefano Zampini /* init with defaults */ 1145b03ebc13SStefano Zampini cedges[i] = corners[i*2] = corners[i*2+1] = -1; 11464e64d54eSstefano_zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 11471e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 11481e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 11494e64d54eSstefano_zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1150b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 11514e64d54eSstefano_zampini for (j=0;j<size;j++) { 11524e64d54eSstefano_zampini PetscInt k,ee = idxs[j]; 11534e64d54eSstefano_zampini for (k=ii[ee];k<ii[ee+1];k++) { 11544e64d54eSstefano_zampini PetscInt vv = jj[k]; 11554e64d54eSstefano_zampini if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) { 1156b03ebc13SStefano Zampini if (found == 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found more then two corners for edge %d\n",i); 1157b03ebc13SStefano Zampini corners[i*2+found++] = vv; 11584e64d54eSstefano_zampini } 11594e64d54eSstefano_zampini } 11604e64d54eSstefano_zampini } 1161b03ebc13SStefano Zampini if (found != 2) { 1162b03ebc13SStefano Zampini PetscInt e; 1163b03ebc13SStefano Zampini if (fl2g) { 1164b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,1,idxs,&e);CHKERRQ(ierr); 1165b03ebc13SStefano Zampini } else { 1166b03ebc13SStefano Zampini e = idxs[0]; 1167b03ebc13SStefano Zampini } 1168b03ebc13SStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %d corners for edge %d (astart %d, estart %d)\n",found,i,e,idxs[0]); 1169b03ebc13SStefano Zampini } 1170eee23b56SStefano Zampini 1171eee23b56SStefano Zampini /* get primal dof index on this coarse edge */ 1172b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,2,corners+2*i,gc);CHKERRQ(ierr); 1173b03ebc13SStefano Zampini if (gc[0] > gc[1]) { 1174b03ebc13SStefano Zampini PetscInt swap = corners[2*i]; 1175b03ebc13SStefano Zampini corners[2*i] = corners[2*i+1]; 1176b03ebc13SStefano Zampini corners[2*i+1] = swap; 1177b03ebc13SStefano Zampini } 1178eee23b56SStefano Zampini cedges[i] = idxs[size-1]; 11794e64d54eSstefano_zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1180b03ebc13SStefano 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]); 11814e64d54eSstefano_zampini } 11820569b399SStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 11834e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvc);CHKERRQ(ierr); 11844e64d54eSstefano_zampini 1185a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 1186a13144ffSStefano Zampini /* Inspects columns of lG (rows of lGt) and make sure the change of basis will 1187a13144ffSStefano Zampini not interfere with neighbouring coarse edges */ 1188a13144ffSStefano Zampini ierr = PetscMalloc1(nee+1,&emarks);CHKERRQ(ierr); 1189a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1190a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1191a13144ffSStefano Zampini PetscInt emax = 0,eemax = 0; 1192a13144ffSStefano Zampini 1193a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1194a13144ffSStefano Zampini ierr = PetscMemzero(emarks,(nee+1)*sizeof(PetscInt));CHKERRQ(ierr); 1195a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++; 1196a13144ffSStefano Zampini for (j=1;j<nee+1;j++) { 1197a13144ffSStefano Zampini if (emax < emarks[j]) { 1198a13144ffSStefano Zampini emax = emarks[j]; 1199a13144ffSStefano Zampini eemax = j; 1200a13144ffSStefano Zampini } 1201a13144ffSStefano Zampini } 1202a13144ffSStefano Zampini /* not relevant for edges */ 1203a13144ffSStefano Zampini if (!eemax) continue; 1204a13144ffSStefano Zampini 1205a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 1206a13144ffSStefano Zampini if (marks[jj[j]] && marks[jj[j]] != eemax) { 1207c2151214SStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_SUP,"Found 2 coarse edges (id %d and %d) connected through the %d nodal dof at edge dof %d\n",marks[jj[j]]-1,eemax,i,jj[j]); 1208a13144ffSStefano Zampini } 1209a13144ffSStefano Zampini } 1210a13144ffSStefano Zampini } 1211a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 1212a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1213a13144ffSStefano Zampini #endif 1214a13144ffSStefano Zampini 1215a13144ffSStefano Zampini /* Compute extended rows indices for edge blocks of the change of basis */ 1216a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1217a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);CHKERRQ(ierr); 1218a13144ffSStefano Zampini extmem *= maxsize; 1219a13144ffSStefano Zampini ierr = PetscMalloc1(extmem*nee,&extrow);CHKERRQ(ierr); 1220a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extrows);CHKERRQ(ierr); 1221a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&extrowcum);CHKERRQ(ierr); 1222a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1223a13144ffSStefano Zampini PetscInt mark = 0,size,start; 1224213b8bfaSStefano Zampini 1225a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1226a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) 1227a13144ffSStefano Zampini if (marks[jj[j]] && !mark) 1228a13144ffSStefano Zampini mark = marks[jj[j]]; 1229a13144ffSStefano Zampini 1230a13144ffSStefano Zampini /* not relevant */ 1231a13144ffSStefano Zampini if (!mark) continue; 1232a13144ffSStefano Zampini 1233a13144ffSStefano Zampini /* import extended row */ 1234a13144ffSStefano Zampini mark--; 1235a13144ffSStefano Zampini start = mark*extmem+extrowcum[mark]; 1236a13144ffSStefano Zampini size = ii[i+1]-ii[i]; 1237a13144ffSStefano Zampini if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %d > %d",extrowcum[mark] + size,extmem); 1238a13144ffSStefano Zampini ierr = PetscMemcpy(extrow+start,jj+ii[i],size*sizeof(PetscInt));CHKERRQ(ierr); 1239a13144ffSStefano Zampini extrowcum[mark] += size; 1240a13144ffSStefano Zampini } 1241a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1242213b8bfaSStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 1243213b8bfaSStefano Zampini ierr = PetscFree(marks);CHKERRQ(ierr); 1244213b8bfaSStefano Zampini 1245213b8bfaSStefano Zampini /* Compress extrows */ 1246a13144ffSStefano Zampini cum = 0; 1247a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1248a13144ffSStefano Zampini PetscInt size = extrowcum[i],*start = extrow + i*extmem; 1249a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&size,start);CHKERRQ(ierr); 1250a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);CHKERRQ(ierr); 1251a13144ffSStefano Zampini cum = PetscMax(cum,size); 1252a13144ffSStefano Zampini } 1253a13144ffSStefano Zampini ierr = PetscFree(extrowcum);CHKERRQ(ierr); 1254a13144ffSStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 1255a13144ffSStefano Zampini ierr = PetscBTDestroy(&btvcand);CHKERRQ(ierr); 1256a13144ffSStefano Zampini 1257a13144ffSStefano Zampini /* Workspace for lapack inner calls and VecSetValues */ 1258a13144ffSStefano Zampini ierr = PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);CHKERRQ(ierr); 1259a13144ffSStefano Zampini 1260a13144ffSStefano Zampini /* Create change of basis matrix (preallocation can be improved) */ 1261a13144ffSStefano Zampini ierr = MatCreate(comm,&T);CHKERRQ(ierr); 1262c2151214SStefano Zampini ierr = MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n, 1263c2151214SStefano Zampini pc->pmat->rmap->N,pc->pmat->rmap->N);CHKERRQ(ierr); 1264a13144ffSStefano Zampini ierr = MatSetType(T,MATAIJ);CHKERRQ(ierr); 1265a13144ffSStefano Zampini ierr = MatSeqAIJSetPreallocation(T,10,NULL);CHKERRQ(ierr); 1266a13144ffSStefano Zampini ierr = MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);CHKERRQ(ierr); 12671e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(T,al2g,al2g);CHKERRQ(ierr); 1268a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1269a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 1270213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&al2g);CHKERRQ(ierr); 1271a13144ffSStefano Zampini 1272a13144ffSStefano Zampini /* Defaults to identity */ 1273c2151214SStefano Zampini ierr = MatCreateVecs(pc->pmat,&tvec,NULL);CHKERRQ(ierr); 1274a13144ffSStefano Zampini ierr = VecSet(tvec,1.0);CHKERRQ(ierr); 1275a13144ffSStefano Zampini ierr = MatDiagonalSet(T,tvec,INSERT_VALUES);CHKERRQ(ierr); 1276a13144ffSStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 1277a13144ffSStefano Zampini 12781e0482f5SStefano Zampini /* Create discrete gradient for the coarser level if needed */ 12791e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 12801e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 12811e0482f5SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 12821e0482f5SStefano Zampini ISLocalToGlobalMapping cel2g,cvl2g; 12831e0482f5SStefano Zampini IS wis,gwis; 12841e0482f5SStefano Zampini PetscInt cnv,cne; 12851e0482f5SStefano Zampini 12861e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,nee,cedges,PETSC_COPY_VALUES,&wis);CHKERRQ(ierr); 12871e0482f5SStefano Zampini if (fl2g) { 12881e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(fl2g,wis,&pcbddc->nedclocal);CHKERRQ(ierr); 12891e0482f5SStefano Zampini } else { 12901e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)wis);CHKERRQ(ierr); 12911e0482f5SStefano Zampini pcbddc->nedclocal = wis; 12921e0482f5SStefano Zampini } 12931e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(el2g,wis,&gwis);CHKERRQ(ierr); 12941e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12951e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cne,&wis);CHKERRQ(ierr); 12961e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cel2g);CHKERRQ(ierr); 12971e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12981e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 12991e0482f5SStefano Zampini 13001e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,2*nee,corners,PETSC_USE_POINTER,&wis);CHKERRQ(ierr); 13011e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(vl2g,wis,&gwis);CHKERRQ(ierr); 13021e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13031e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cnv,&wis);CHKERRQ(ierr); 13041e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cvl2g);CHKERRQ(ierr); 13051e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13061e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 13071e0482f5SStefano Zampini 13081e0482f5SStefano Zampini ierr = MatCreate(comm,&pcbddc->nedcG);CHKERRQ(ierr); 13091e0482f5SStefano Zampini ierr = MatSetSizes(pcbddc->nedcG,PETSC_DECIDE,PETSC_DECIDE,cne,cnv);CHKERRQ(ierr); 13101e0482f5SStefano Zampini ierr = MatSetType(pcbddc->nedcG,MATAIJ);CHKERRQ(ierr); 13111e0482f5SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->nedcG,2,NULL);CHKERRQ(ierr); 13121e0482f5SStefano Zampini ierr = MatMPIAIJSetPreallocation(pcbddc->nedcG,2,NULL,2,NULL);CHKERRQ(ierr); 13131e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->nedcG,cel2g,cvl2g);CHKERRQ(ierr); 13141e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cel2g);CHKERRQ(ierr); 13151e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cvl2g);CHKERRQ(ierr); 13161e0482f5SStefano Zampini } 1317213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&vl2g);CHKERRQ(ierr); 13181e0482f5SStefano Zampini 13191e0482f5SStefano Zampini #if defined(PRINT_GDET) 13201e0482f5SStefano Zampini inc = 0; 13211e0482f5SStefano Zampini lev = pcbddc->current_level; 13221e0482f5SStefano Zampini #endif 1323213b8bfaSStefano Zampini 1324213b8bfaSStefano Zampini /* Insert values in the change of basis matrix */ 1325a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1326a13144ffSStefano Zampini Mat Gins = NULL, GKins = NULL; 13271e0482f5SStefano Zampini IS cornersis = NULL; 13281e0482f5SStefano Zampini PetscScalar cvals[2]; 1329a13144ffSStefano Zampini 13301e0482f5SStefano Zampini if (pcbddc->nedcG) { 13311e0482f5SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,2,corners+2*i,PETSC_USE_POINTER,&cornersis);CHKERRQ(ierr); 13321e0482f5SStefano Zampini } 13331e0482f5SStefano Zampini ierr = PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],cornersis,&Gins,&GKins,cvals,work,rwork);CHKERRQ(ierr); 1334a13144ffSStefano Zampini if (Gins && GKins) { 1335a13144ffSStefano Zampini PetscScalar *data; 1336a13144ffSStefano Zampini const PetscInt *rows,*cols; 1337a13144ffSStefano Zampini PetscInt nrh,nch,nrc,ncc; 1338a13144ffSStefano Zampini 1339a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&cols);CHKERRQ(ierr); 1340a13144ffSStefano Zampini /* H1 */ 1341a13144ffSStefano Zampini ierr = ISGetIndices(extrows[i],&rows);CHKERRQ(ierr); 1342a13144ffSStefano Zampini ierr = MatGetSize(Gins,&nrh,&nch);CHKERRQ(ierr); 1343a13144ffSStefano Zampini ierr = MatDenseGetArray(Gins,&data);CHKERRQ(ierr); 1344a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);CHKERRQ(ierr); 1345a13144ffSStefano Zampini ierr = MatDenseRestoreArray(Gins,&data);CHKERRQ(ierr); 1346a13144ffSStefano Zampini ierr = ISRestoreIndices(extrows[i],&rows);CHKERRQ(ierr); 1347a13144ffSStefano Zampini /* complement */ 1348a13144ffSStefano Zampini ierr = MatGetSize(GKins,&nrc,&ncc);CHKERRQ(ierr); 13491e0482f5SStefano Zampini if (!ncc) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Constant function has not been generated for coarse edge %d",i); 1350213b8bfaSStefano 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); 1351213b8bfaSStefano 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); 1352a13144ffSStefano Zampini ierr = MatDenseGetArray(GKins,&data);CHKERRQ(ierr); 1353a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);CHKERRQ(ierr); 1354a13144ffSStefano Zampini ierr = MatDenseRestoreArray(GKins,&data);CHKERRQ(ierr); 13551e0482f5SStefano Zampini 13561e0482f5SStefano Zampini /* coarse discrete gradient */ 13571e0482f5SStefano Zampini if (pcbddc->nedcG) { 13581e0482f5SStefano Zampini PetscInt cols[2]; 13591e0482f5SStefano Zampini 13601e0482f5SStefano Zampini cols[0] = 2*i; 13611e0482f5SStefano Zampini cols[1] = 2*i+1; 13621e0482f5SStefano Zampini ierr = MatSetValuesLocal(pcbddc->nedcG,1,&i,2,cols,cvals,INSERT_VALUES);CHKERRQ(ierr); 13631e0482f5SStefano Zampini } 1364a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&cols);CHKERRQ(ierr); 1365a13144ffSStefano Zampini } 1366a13144ffSStefano Zampini ierr = ISDestroy(&extrows[i]);CHKERRQ(ierr); 1367a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 13681e0482f5SStefano Zampini ierr = ISDestroy(&cornersis);CHKERRQ(ierr); 1369a13144ffSStefano Zampini ierr = MatDestroy(&Gins);CHKERRQ(ierr); 1370a13144ffSStefano Zampini ierr = MatDestroy(&GKins);CHKERRQ(ierr); 1371a13144ffSStefano Zampini } 1372213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&el2g);CHKERRQ(ierr); 1373a13144ffSStefano Zampini 1374a13144ffSStefano Zampini /* Start assembling */ 1375a13144ffSStefano Zampini ierr = MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13761e0482f5SStefano Zampini if (pcbddc->nedcG) { 13771e0482f5SStefano Zampini ierr = MatAssemblyBegin(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13781e0482f5SStefano Zampini } 1379a13144ffSStefano Zampini 1380a13144ffSStefano Zampini /* Free */ 1381c2151214SStefano Zampini if (fl2g) { 1382c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1383c2151214SStefano Zampini for (i=0;i<nee;i++) { 1384c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1385c2151214SStefano Zampini } 1386c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1387c2151214SStefano Zampini } 1388eee23b56SStefano Zampini 1389eee23b56SStefano Zampini /* hack mat_graph with primal dofs on the coarse edges */ 1390eee23b56SStefano Zampini { 1391eee23b56SStefano Zampini PCBDDCGraph graph = pcbddc->mat_graph; 1392eee23b56SStefano Zampini PetscInt *oqueue = graph->queue; 1393eee23b56SStefano Zampini PetscInt *ocptr = graph->cptr; 1394eee23b56SStefano Zampini PetscInt ncc,*idxs; 1395eee23b56SStefano Zampini 1396eee23b56SStefano Zampini /* find first primal edge */ 1397eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1398eee23b56SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1399eee23b56SStefano Zampini } else { 1400eee23b56SStefano Zampini if (fl2g) { 1401eee23b56SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,nee,cedges,cedges);CHKERRQ(ierr); 1402eee23b56SStefano Zampini } 1403eee23b56SStefano Zampini idxs = cedges; 1404eee23b56SStefano Zampini } 1405eee23b56SStefano Zampini cum = 0; 1406eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1407eee23b56SStefano Zampini 1408eee23b56SStefano Zampini /* adapt connected components */ 1409eee23b56SStefano Zampini ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,ocptr[graph->ncc],&graph->queue);CHKERRQ(ierr); 1410eee23b56SStefano Zampini graph->cptr[0] = 0; 1411eee23b56SStefano Zampini for (i=0,ncc=0;i<graph->ncc;i++) { 1412eee23b56SStefano Zampini PetscInt lc = ocptr[i+1]-ocptr[i]; 1413eee23b56SStefano Zampini if (cum != nee && oqueue[ocptr[i+1]-1] == cedges[cum]) { /* this cc has a primal dof */ 1414eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc]+1; 1415eee23b56SStefano Zampini graph->queue[graph->cptr[ncc]] = cedges[cum]; 1416eee23b56SStefano Zampini ncc++; 1417eee23b56SStefano Zampini lc--; 1418eee23b56SStefano Zampini cum++; 1419eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1420eee23b56SStefano Zampini } 1421eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc] + lc; 1422eee23b56SStefano Zampini for (j=0;j<lc;j++) graph->queue[graph->cptr[ncc]+j] = oqueue[ocptr[i]+j]; 1423eee23b56SStefano Zampini ncc++; 1424eee23b56SStefano Zampini } 1425eee23b56SStefano Zampini graph->ncc = ncc; 1426eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1427eee23b56SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1428eee23b56SStefano Zampini } 1429eee23b56SStefano Zampini ierr = PetscFree2(ocptr,oqueue);CHKERRQ(ierr); 1430eee23b56SStefano Zampini } 1431213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&fl2g);CHKERRQ(ierr); 1432c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1433c2151214SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1434213b8bfaSStefano Zampini ierr = MatDestroy(&conn);CHKERRQ(ierr); 1435eee23b56SStefano Zampini 1436c2151214SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 1437a13144ffSStefano Zampini ierr = PetscFree(extrow);CHKERRQ(ierr); 1438a13144ffSStefano Zampini ierr = PetscFree2(work,rwork);CHKERRQ(ierr); 1439b03ebc13SStefano Zampini ierr = PetscFree(corners);CHKERRQ(ierr); 1440b03ebc13SStefano Zampini ierr = PetscFree(cedges);CHKERRQ(ierr); 1441a13144ffSStefano Zampini ierr = PetscFree(extrows);CHKERRQ(ierr); 1442a13144ffSStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 1443a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 1444a13144ffSStefano Zampini 1445a13144ffSStefano Zampini /* Complete assembling */ 1446a13144ffSStefano Zampini ierr = MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14471e0482f5SStefano Zampini if (pcbddc->nedcG) { 14481e0482f5SStefano Zampini ierr = MatAssemblyEnd(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14491e0482f5SStefano Zampini #if 0 14501e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->nedcG,"coarse_G");CHKERRQ(ierr); 14511e0482f5SStefano Zampini ierr = MatView(pcbddc->nedcG,NULL);CHKERRQ(ierr); 14521e0482f5SStefano Zampini #endif 14531e0482f5SStefano Zampini } 1454a13144ffSStefano Zampini 1455a13144ffSStefano Zampini /* set change of basis */ 1456213b8bfaSStefano Zampini ierr = PCBDDCSetChangeOfBasisMat(pc,T,singular);CHKERRQ(ierr); 1457a13144ffSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1458a13144ffSStefano Zampini 1459a13144ffSStefano Zampini PetscFunctionReturn(0); 1460a13144ffSStefano Zampini } 1461a13144ffSStefano Zampini 1462d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 1463d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 1464d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 1465d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 1466d8203eabSStefano Zampini { 1467d8203eabSStefano Zampini PetscErrorCode ierr; 1468d8203eabSStefano Zampini PetscInt i; 1469d8203eabSStefano Zampini 1470d8203eabSStefano Zampini PetscFunctionBegin; 1471d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 1472d8203eabSStefano Zampini PetscInt first,last; 1473d8203eabSStefano Zampini 1474d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 147586fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 1476d8203eabSStefano Zampini if (i>=first && i < last) { 1477d8203eabSStefano Zampini PetscScalar *data; 1478d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1479d8203eabSStefano Zampini if (!has_const) { 1480d8203eabSStefano Zampini data[i-first] = 1.; 1481d8203eabSStefano Zampini } else { 148286fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 148386fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 1484d8203eabSStefano Zampini } 1485d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1486d8203eabSStefano Zampini } 1487d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1488d8203eabSStefano Zampini } 1489d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 1490d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 1491d8203eabSStefano Zampini PetscInt first,last; 1492ddc40e2cSstefano_zampini ierr = VecLockPop(quad_vecs[i]);CHKERRQ(ierr); 1493d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 1494d8203eabSStefano Zampini if (i>=first && i < last) { 1495d8203eabSStefano Zampini PetscScalar *data; 1496d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1497d8203eabSStefano Zampini if (!has_const) { 1498d8203eabSStefano Zampini data[i-first] = 0.; 1499d8203eabSStefano Zampini } else { 150086fa73c5SStefano Zampini data[2*i-first] = 0.; 150186fa73c5SStefano Zampini data[2*i-first+1] = 0.; 1502d8203eabSStefano Zampini } 1503d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1504d8203eabSStefano Zampini } 1505d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1506ddc40e2cSstefano_zampini ierr = VecLockPush(quad_vecs[i]);CHKERRQ(ierr); 1507d8203eabSStefano Zampini } 1508d8203eabSStefano Zampini PetscFunctionReturn(0); 1509d8203eabSStefano Zampini } 1510d8203eabSStefano Zampini 15118ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 1512669cc0f4SStefano Zampini { 1513a198735bSStefano Zampini Mat loc_divudotp; 1514fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 15158ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 1516669cc0f4SStefano Zampini IS *faces,*edges; 1517669cc0f4SStefano Zampini PetscScalar *vals; 1518669cc0f4SStefano Zampini const PetscScalar *array; 1519669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 15201ae86dd6SStefano Zampini PetscMPIInt rank; 1521a198735bSStefano Zampini PetscErrorCode ierr; 1522669cc0f4SStefano Zampini 1523669cc0f4SStefano Zampini PetscFunctionBegin; 1524669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1525669cc0f4SStefano Zampini if (graph->twodim) { 1526669cc0f4SStefano Zampini lmaxneighs = 2; 1527669cc0f4SStefano Zampini } else { 1528669cc0f4SStefano Zampini lmaxneighs = 1; 1529669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1530669cc0f4SStefano Zampini const PetscInt *idxs; 1531669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1532669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 1533669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1534669cc0f4SStefano Zampini } 1535669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 1536669cc0f4SStefano Zampini } 1537669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1538669cc0f4SStefano Zampini maxsize = 0; 1539669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1540669cc0f4SStefano Zampini PetscInt nn; 1541669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1542669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1543669cc0f4SStefano Zampini } 1544669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1545669cc0f4SStefano Zampini PetscInt nn; 1546669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1547669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1548669cc0f4SStefano Zampini } 1549669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 1550669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 1551669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 15528ae0ca82SStefano Zampini if (!transpose) { 15538ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 15548ae0ca82SStefano Zampini } else { 15558ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 15568ae0ca82SStefano Zampini } 1557669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 15581ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 1559d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 1560669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1561ddc40e2cSstefano_zampini ierr = VecLockPop(quad_vecs[i]);CHKERRQ(ierr); 15628ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 1563669cc0f4SStefano Zampini } 1564d8203eabSStefano Zampini 1565669cc0f4SStefano Zampini /* compute local quad vec */ 1566a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 15678ae0ca82SStefano Zampini if (!transpose) { 1568a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 15698ae0ca82SStefano Zampini } else { 15708ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 15718ae0ca82SStefano Zampini } 1572669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 15738ae0ca82SStefano Zampini if (!transpose) { 1574a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 15758ae0ca82SStefano Zampini } else { 15768ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 15778ae0ca82SStefano Zampini } 1578fa23a32eSStefano Zampini if (vl2l) { 1579187c917aSStefano Zampini Mat lA; 1580187c917aSStefano Zampini VecScatter sc; 1581187c917aSStefano Zampini 1582187c917aSStefano Zampini ierr = MatISGetLocalMat(A,&lA);CHKERRQ(ierr); 1583187c917aSStefano Zampini ierr = MatCreateVecs(lA,&vins,NULL);CHKERRQ(ierr); 1584187c917aSStefano Zampini ierr = VecScatterCreate(v,vl2l,vins,NULL,&sc);CHKERRQ(ierr); 1585187c917aSStefano Zampini ierr = VecScatterBegin(sc,v,vins,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1586187c917aSStefano Zampini ierr = VecScatterEnd(sc,v,vins,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1587187c917aSStefano Zampini ierr = VecScatterDestroy(&sc);CHKERRQ(ierr); 1588fa23a32eSStefano Zampini } else { 1589fa23a32eSStefano Zampini vins = v; 1590fa23a32eSStefano Zampini } 1591fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 1592669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 15939a962809SStefano Zampini 15941ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 15951ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 1596669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1597669cc0f4SStefano Zampini const PetscInt *idxs; 1598669cc0f4SStefano Zampini PetscInt idx,nn,j; 1599669cc0f4SStefano Zampini 1600669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 1601669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1602669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 16031ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1604669cc0f4SStefano Zampini idx = -(idx+1); 1605669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1606669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 1607669cc0f4SStefano Zampini } 1608669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1609669cc0f4SStefano Zampini const PetscInt *idxs; 1610669cc0f4SStefano Zampini PetscInt idx,nn,j; 1611669cc0f4SStefano Zampini 1612669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1613669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1614669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 16151ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1616669cc0f4SStefano Zampini idx = -(idx+1); 1617669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1618669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1619669cc0f4SStefano Zampini } 1620c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1621fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 1622fa23a32eSStefano Zampini if (vl2l) { 1623187c917aSStefano Zampini ierr = VecDestroy(&vins);CHKERRQ(ierr); 1624fa23a32eSStefano Zampini } 1625669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1626669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1627669cc0f4SStefano Zampini 1628669cc0f4SStefano Zampini /* assemble near null space */ 1629669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1630669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 1631669cc0f4SStefano Zampini } 1632669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1633669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 1634ddc40e2cSstefano_zampini ierr = VecLockPush(quad_vecs[i]);CHKERRQ(ierr); 1635669cc0f4SStefano Zampini } 1636669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 1637669cc0f4SStefano Zampini PetscFunctionReturn(0); 1638669cc0f4SStefano Zampini } 1639669cc0f4SStefano Zampini 16407620a527SStefano Zampini PetscErrorCode PCBDDCAddPrimalVerticesLocalIS(PC pc, IS primalv) 16417620a527SStefano Zampini { 16427620a527SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16437620a527SStefano Zampini PetscErrorCode ierr; 16447620a527SStefano Zampini 16457620a527SStefano Zampini PetscFunctionBegin; 16467620a527SStefano Zampini if (primalv) { 16477620a527SStefano Zampini if (pcbddc->user_primal_vertices_local) { 16487620a527SStefano Zampini IS list[2], newp; 16497620a527SStefano Zampini 16507620a527SStefano Zampini list[0] = primalv; 16517620a527SStefano Zampini list[1] = pcbddc->user_primal_vertices_local; 16527620a527SStefano Zampini ierr = ISConcatenate(PetscObjectComm((PetscObject)pc),2,list,&newp);CHKERRQ(ierr); 16537620a527SStefano Zampini ierr = ISSortRemoveDups(newp);CHKERRQ(ierr); 16547620a527SStefano Zampini ierr = ISDestroy(&list[1]);CHKERRQ(ierr); 16557620a527SStefano Zampini pcbddc->user_primal_vertices_local = newp; 16567620a527SStefano Zampini } else { 16577620a527SStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primalv);CHKERRQ(ierr); 16587620a527SStefano Zampini } 16597620a527SStefano Zampini } 16607620a527SStefano Zampini PetscFunctionReturn(0); 16617620a527SStefano Zampini } 1662669cc0f4SStefano Zampini 16631f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 16641f4df5f7SStefano Zampini { 16651f4df5f7SStefano Zampini PetscErrorCode ierr; 16661f4df5f7SStefano Zampini Vec local,global; 16671f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16681f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 16695c5e10d6SStefano Zampini PetscBool monolithic = PETSC_FALSE; 16701f4df5f7SStefano Zampini 16711f4df5f7SStefano Zampini PetscFunctionBegin; 16725c5e10d6SStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC topology options","PC");CHKERRQ(ierr); 16735c5e10d6SStefano Zampini ierr = PetscOptionsBool("-pc_bddc_monolithic","Discard any information on dofs splitting",NULL,monolithic,&monolithic,NULL);CHKERRQ(ierr); 16745c5e10d6SStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 16751f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 167621ef3d20SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 16771f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 16785c5e10d6SStefano Zampini if (monolithic) goto boundary; 16795c5e10d6SStefano Zampini 16801f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 16811f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 16821f4df5f7SStefano Zampini PetscInt i; 16831f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16841f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 16851f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16861f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 16871f4df5f7SStefano Zampini } 16881f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 16891f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 16901f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 16911f4df5f7SStefano Zampini } 16921f4df5f7SStefano Zampini } else { 169321ef3d20SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present */ 169421ef3d20SStefano Zampini DM dm; 169521ef3d20SStefano Zampini 169621ef3d20SStefano Zampini ierr = PCGetDM(pc, &dm);CHKERRQ(ierr); 169721ef3d20SStefano Zampini if (!dm) { 169821ef3d20SStefano Zampini ierr = MatGetDM(pc->pmat, &dm);CHKERRQ(ierr); 169921ef3d20SStefano Zampini } 170021ef3d20SStefano Zampini if (dm) { 170121ef3d20SStefano Zampini IS *fields; 170221ef3d20SStefano Zampini PetscInt nf,i; 170321ef3d20SStefano Zampini ierr = DMCreateFieldDecomposition(dm,&nf,NULL,&fields,NULL);CHKERRQ(ierr); 170421ef3d20SStefano Zampini ierr = PetscMalloc1(nf,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 170521ef3d20SStefano Zampini for (i=0;i<nf;i++) { 170621ef3d20SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,fields[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 170721ef3d20SStefano Zampini ierr = ISDestroy(&fields[i]);CHKERRQ(ierr); 170821ef3d20SStefano Zampini } 170921ef3d20SStefano Zampini ierr = PetscFree(fields);CHKERRQ(ierr); 171021ef3d20SStefano Zampini pcbddc->n_ISForDofsLocal = nf; 171121ef3d20SStefano Zampini } else { /* See if MATIS has fields attached by the conversion from MatNest */ 171221ef3d20SStefano Zampini PetscContainer c; 171321ef3d20SStefano Zampini 171421ef3d20SStefano Zampini ierr = PetscObjectQuery((PetscObject)pc->pmat,"_convert_nest_lfields",(PetscObject*)&c);CHKERRQ(ierr); 171521ef3d20SStefano Zampini if (c) { 171621ef3d20SStefano Zampini MatISLocalFields lf; 171721ef3d20SStefano Zampini ierr = PetscContainerGetPointer(c,(void**)&lf);CHKERRQ(ierr); 171821ef3d20SStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,lf->nr,lf->rf);CHKERRQ(ierr); 171921ef3d20SStefano Zampini } else { /* fallback, create the default fields if bs > 1 */ 17201f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1721986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 172221ef3d20SStefano Zampini if (i > 1) { 1723986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 17241f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 17251f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 17261f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 17271f4df5f7SStefano Zampini } 17281f4df5f7SStefano Zampini } 172921ef3d20SStefano Zampini } 173021ef3d20SStefano Zampini } 17317a0e7b2cSstefano_zampini } else { 17327a0e7b2cSstefano_zampini PetscInt i; 17337a0e7b2cSstefano_zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 17347a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 17357a0e7b2cSstefano_zampini } 17361f4df5f7SStefano Zampini } 1737986cdee1SStefano Zampini } 17381f4df5f7SStefano Zampini 17395c5e10d6SStefano Zampini boundary: 17401f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 17411f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 17427a0e7b2cSstefano_zampini } else if (pcbddc->DirichletBoundariesLocal) { 17437a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 17441f4df5f7SStefano Zampini } 17451f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 17461f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 17477a0e7b2cSstefano_zampini } else if (pcbddc->NeumannBoundariesLocal) { 17487a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 17491f4df5f7SStefano Zampini } 17501f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 17511f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 17521f4df5f7SStefano Zampini } 17531f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 17541f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 17557620a527SStefano Zampini /* detect local disconnected subdomains if requested (use matis->A) */ 17567620a527SStefano Zampini if (pcbddc->detect_disconnected) { 17577620a527SStefano Zampini IS primalv = NULL; 17587620a527SStefano Zampini PetscInt i; 17597a0e7b2cSstefano_zampini 17607620a527SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 17617620a527SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 17627620a527SStefano Zampini } 17637620a527SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 17647620a527SStefano Zampini ierr = PCBDDCDetectDisconnectedComponents(pc,&pcbddc->n_local_subs,&pcbddc->local_subs,&primalv);CHKERRQ(ierr); 17657620a527SStefano Zampini ierr = PCBDDCAddPrimalVerticesLocalIS(pc,primalv);CHKERRQ(ierr); 17667620a527SStefano Zampini ierr = ISDestroy(&primalv);CHKERRQ(ierr); 17677620a527SStefano Zampini } 17687620a527SStefano Zampini /* early stage corner detection */ 17697620a527SStefano Zampini { 17707620a527SStefano Zampini DM dm; 17717620a527SStefano Zampini 17727620a527SStefano Zampini ierr = MatGetDM(pc->pmat,&dm);CHKERRQ(ierr); 17737620a527SStefano Zampini if (dm) { 17747620a527SStefano Zampini PetscBool isda; 17757620a527SStefano Zampini 17767620a527SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)dm,DMDA,&isda);CHKERRQ(ierr); 17777620a527SStefano Zampini if (isda) { 17787620a527SStefano Zampini ISLocalToGlobalMapping l2l; 17797620a527SStefano Zampini IS corners; 17807620a527SStefano Zampini Mat lA; 17817620a527SStefano Zampini 1782*d4a6ed37SStefano Zampini ierr = DMDAGetSubdomainCornersIS(dm,&corners);CHKERRQ(ierr); 17837620a527SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&lA);CHKERRQ(ierr); 17847620a527SStefano Zampini ierr = MatGetLocalToGlobalMapping(lA,&l2l,NULL);CHKERRQ(ierr); 17857620a527SStefano Zampini ierr = MatISRestoreLocalMat(pc->pmat,&lA);CHKERRQ(ierr); 17867620a527SStefano Zampini if (l2l) { 17877620a527SStefano Zampini const PetscInt *idx; 17887620a527SStefano Zampini PetscInt bs,*idxout,n; 17897620a527SStefano Zampini 17907620a527SStefano Zampini ierr = ISLocalToGlobalMappingGetBlockSize(l2l,&bs);CHKERRQ(ierr); 17917620a527SStefano Zampini ierr = ISGetLocalSize(corners,&n);CHKERRQ(ierr); 17927620a527SStefano Zampini ierr = ISGetIndices(corners,&idx);CHKERRQ(ierr); 17937620a527SStefano Zampini ierr = PetscMalloc1(n,&idxout);CHKERRQ(ierr); 17947620a527SStefano Zampini ierr = ISLocalToGlobalMappingApplyBlock(l2l,n,idx,idxout);CHKERRQ(ierr); 17957620a527SStefano Zampini ierr = ISRestoreIndices(corners,&idx);CHKERRQ(ierr); 1796*d4a6ed37SStefano Zampini ierr = DMDARestoreSubdomainCornersIS(dm,&corners);CHKERRQ(ierr); 17977620a527SStefano Zampini ierr = ISCreateBlock(PetscObjectComm((PetscObject)pc),bs,n,idxout,PETSC_OWN_POINTER,&corners);CHKERRQ(ierr); 17987620a527SStefano Zampini ierr = PCBDDCAddPrimalVerticesLocalIS(pc,corners);CHKERRQ(ierr); 17997620a527SStefano Zampini ierr = ISDestroy(&corners);CHKERRQ(ierr); 18007620a527SStefano Zampini } else { /* not from DMDA */ 1801*d4a6ed37SStefano Zampini ierr = DMDARestoreSubdomainCornersIS(dm,&corners);CHKERRQ(ierr); 18027620a527SStefano Zampini } 18037620a527SStefano Zampini } 18047620a527SStefano Zampini } 18057620a527SStefano Zampini } 18067a0e7b2cSstefano_zampini PetscFunctionReturn(0); 18077a0e7b2cSstefano_zampini } 18087a0e7b2cSstefano_zampini 18097a0e7b2cSstefano_zampini PetscErrorCode PCBDDCConsistencyCheckIS(PC pc, MPI_Op mop, IS *is) 18107a0e7b2cSstefano_zampini { 18117a0e7b2cSstefano_zampini Mat_IS *matis = (Mat_IS*)(pc->pmat->data); 18127a0e7b2cSstefano_zampini PetscErrorCode ierr; 18137a0e7b2cSstefano_zampini IS nis; 18147a0e7b2cSstefano_zampini const PetscInt *idxs; 18157a0e7b2cSstefano_zampini PetscInt i,nd,n = matis->A->rmap->n,*nidxs,nnd; 18167a0e7b2cSstefano_zampini PetscBool *ld; 18177a0e7b2cSstefano_zampini 18187a0e7b2cSstefano_zampini PetscFunctionBegin; 18197a0e7b2cSstefano_zampini if (mop != MPI_LAND && mop != MPI_LOR) SETERRQ(PetscObjectComm((PetscObject)(pc)),PETSC_ERR_SUP,"Supported are MPI_LAND and MPI_LOR"); 18207a0e7b2cSstefano_zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 18217a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 18227a0e7b2cSstefano_zampini /* init rootdata with true */ 18237a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_rootdata; 18247a0e7b2cSstefano_zampini for (i=0;i<pc->pmat->rmap->n;i++) ld[i] = PETSC_TRUE; 18257a0e7b2cSstefano_zampini } else { 18267a0e7b2cSstefano_zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscBool));CHKERRQ(ierr); 18277a0e7b2cSstefano_zampini } 18287a0e7b2cSstefano_zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscBool));CHKERRQ(ierr); 18297a0e7b2cSstefano_zampini ierr = ISGetLocalSize(*is,&nd);CHKERRQ(ierr); 18307a0e7b2cSstefano_zampini ierr = ISGetIndices(*is,&idxs);CHKERRQ(ierr); 18317a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_leafdata; 18327a0e7b2cSstefano_zampini for (i=0;i<nd;i++) 18337a0e7b2cSstefano_zampini if (-1 < idxs[i] && idxs[i] < n) 18347a0e7b2cSstefano_zampini ld[idxs[i]] = PETSC_TRUE; 18357a0e7b2cSstefano_zampini ierr = ISRestoreIndices(*is,&idxs);CHKERRQ(ierr); 18367a0e7b2cSstefano_zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 18377a0e7b2cSstefano_zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 18387a0e7b2cSstefano_zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 18397a0e7b2cSstefano_zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 18407a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 18417a0e7b2cSstefano_zampini ierr = PetscMalloc1(nd,&nidxs);CHKERRQ(ierr); 18427a0e7b2cSstefano_zampini } else { 18437a0e7b2cSstefano_zampini ierr = PetscMalloc1(n,&nidxs);CHKERRQ(ierr); 18447a0e7b2cSstefano_zampini } 18457a0e7b2cSstefano_zampini for (i=0,nnd=0;i<n;i++) 18467a0e7b2cSstefano_zampini if (ld[i]) 18477a0e7b2cSstefano_zampini nidxs[nnd++] = i; 18487a0e7b2cSstefano_zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(*is)),nnd,nidxs,PETSC_OWN_POINTER,&nis);CHKERRQ(ierr); 18497a0e7b2cSstefano_zampini ierr = ISDestroy(is);CHKERRQ(ierr); 18507a0e7b2cSstefano_zampini *is = nis; 18511f4df5f7SStefano Zampini PetscFunctionReturn(0); 18521f4df5f7SStefano Zampini } 18531f4df5f7SStefano Zampini 18543e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 18553e589ea0SStefano Zampini { 18563e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 18573e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 18583e589ea0SStefano Zampini PetscErrorCode ierr; 18593e589ea0SStefano Zampini 18603e589ea0SStefano Zampini PetscFunctionBegin; 18613e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 18623e589ea0SStefano Zampini PetscFunctionReturn(0); 18633e589ea0SStefano Zampini } 18643e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 18653e589ea0SStefano Zampini Vec swap; 18663e589ea0SStefano Zampini 18673e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 18683e589ea0SStefano Zampini swap = pcbddc->work_change; 18693e589ea0SStefano Zampini pcbddc->work_change = r; 18703e589ea0SStefano Zampini r = swap; 18713e589ea0SStefano Zampini } 18723e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 18733e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 18743e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 18753e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 18763e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 18773e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 18783e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1879f913dca9SStefano Zampini pcbddc->work_change = r; 18803e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 18813e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 18823e589ea0SStefano Zampini } 18833e589ea0SStefano Zampini PetscFunctionReturn(0); 18843e589ea0SStefano Zampini } 18853e589ea0SStefano Zampini 1886a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1887a3df083aSStefano Zampini { 1888a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1889a3df083aSStefano Zampini PetscErrorCode ierr; 1890a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1891a3df083aSStefano Zampini 1892a3df083aSStefano Zampini PetscFunctionBegin; 1893a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1894a3df083aSStefano Zampini if (transpose) { 1895a3df083aSStefano Zampini apply_right = ctx->apply_left; 1896a3df083aSStefano Zampini apply_left = ctx->apply_right; 1897a3df083aSStefano Zampini } else { 1898a3df083aSStefano Zampini apply_right = ctx->apply_right; 1899a3df083aSStefano Zampini apply_left = ctx->apply_left; 1900a3df083aSStefano Zampini } 1901a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1902a3df083aSStefano Zampini if (apply_right) { 1903a3df083aSStefano Zampini const PetscScalar *ax; 1904a3df083aSStefano Zampini PetscInt nl,i; 1905a3df083aSStefano Zampini 1906a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1907a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1908a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 1909a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1910a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1911a3df083aSStefano Zampini PetscScalar sum,val; 1912a3df083aSStefano Zampini const PetscInt *idxs; 1913a3df083aSStefano Zampini PetscInt nz,j; 1914a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1915a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1916a3df083aSStefano Zampini sum = 0.; 1917a3df083aSStefano Zampini if (ctx->apply_p0) { 1918a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1919a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1920a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1921a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1922a3df083aSStefano Zampini } 1923a3df083aSStefano Zampini } else { 1924a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1925a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1926a3df083aSStefano Zampini } 1927a3df083aSStefano Zampini } 1928a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1929a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1930a3df083aSStefano Zampini } 1931a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1932a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1933a3df083aSStefano Zampini } 1934a3df083aSStefano Zampini if (transpose) { 1935a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1936a3df083aSStefano Zampini } else { 1937a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1938a3df083aSStefano Zampini } 1939a3df083aSStefano Zampini if (reset_x) { 1940a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1941a3df083aSStefano Zampini } 1942a3df083aSStefano Zampini if (apply_left) { 1943a3df083aSStefano Zampini PetscScalar *ay; 1944a3df083aSStefano Zampini PetscInt i; 1945a3df083aSStefano Zampini 1946a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1947a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1948a3df083aSStefano Zampini PetscScalar sum,val; 1949a3df083aSStefano Zampini const PetscInt *idxs; 1950a3df083aSStefano Zampini PetscInt nz,j; 1951a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1952a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1953a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 1954a3df083aSStefano Zampini if (ctx->apply_p0) { 1955a3df083aSStefano Zampini sum = 0.; 1956a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1957a3df083aSStefano Zampini sum += ay[idxs[j]]; 1958a3df083aSStefano Zampini ay[idxs[j]] += val; 1959a3df083aSStefano Zampini } 1960a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 1961a3df083aSStefano Zampini } else { 1962a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1963a3df083aSStefano Zampini ay[idxs[j]] += val; 1964a3df083aSStefano Zampini } 1965a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 1966a3df083aSStefano Zampini } 1967a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1968a3df083aSStefano Zampini } 1969a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 1970a3df083aSStefano Zampini } 1971a3df083aSStefano Zampini PetscFunctionReturn(0); 1972a3df083aSStefano Zampini } 1973a3df083aSStefano Zampini 1974a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 1975a3df083aSStefano Zampini { 1976a3df083aSStefano Zampini PetscErrorCode ierr; 1977a3df083aSStefano Zampini 1978a3df083aSStefano Zampini PetscFunctionBegin; 1979a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 1980a3df083aSStefano Zampini PetscFunctionReturn(0); 1981a3df083aSStefano Zampini } 1982a3df083aSStefano Zampini 1983a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 1984a3df083aSStefano Zampini { 1985a3df083aSStefano Zampini PetscErrorCode ierr; 1986a3df083aSStefano Zampini 1987a3df083aSStefano Zampini PetscFunctionBegin; 1988a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 1989a3df083aSStefano Zampini PetscFunctionReturn(0); 1990a3df083aSStefano Zampini } 1991a3df083aSStefano Zampini 1992a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 1993a3df083aSStefano Zampini { 1994a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1995a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1996a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1997a3df083aSStefano Zampini PetscErrorCode ierr; 1998a3df083aSStefano Zampini 1999a3df083aSStefano Zampini PetscFunctionBegin; 2000a3df083aSStefano Zampini if (!restore) { 20011dd7afcfSStefano Zampini Mat A_IB,A_BI; 2002a3df083aSStefano Zampini PetscScalar *work; 2003b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 2004a3df083aSStefano Zampini 20059a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 20069a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 2007a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 2008a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 2009a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2010a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 2011a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 2012a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 2013a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 2014a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 2015a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 2016a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 2017a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 2018a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 2019059032f7SStefano Zampini if (reuse) { 2020a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 20211dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 2022059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 2023059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 2024059032f7SStefano Zampini PetscInt i; 2025059032f7SStefano Zampini 2026059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 2027059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 2028059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2029059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 2030059032f7SStefano Zampini } 2031059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 20321dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 2033059032f7SStefano Zampini } 2034a3df083aSStefano Zampini ctx->A = pcis->A_IB; 2035a3df083aSStefano Zampini ctx->work = work; 2036a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 2037a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2038a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2039a3df083aSStefano Zampini pcis->A_IB = A_IB; 2040a3df083aSStefano Zampini 2041a3df083aSStefano Zampini /* A_BI as A_IB^T */ 2042a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 2043a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 2044a3df083aSStefano Zampini pcis->A_BI = A_BI; 2045a3df083aSStefano Zampini } else { 20461dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 20471dd7afcfSStefano Zampini PetscFunctionReturn(0); 20481dd7afcfSStefano Zampini } 2049a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 2050a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 2051a3df083aSStefano Zampini pcis->A_IB = ctx->A; 20521dd7afcfSStefano Zampini ctx->A = NULL; 20531dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 20541dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 20551dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 20561dd7afcfSStefano Zampini if (ctx->free) { 2057059032f7SStefano Zampini PetscInt i; 20581dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 2059059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 2060059032f7SStefano Zampini } 2061059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 2062059032f7SStefano Zampini } 2063a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 2064a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 2065a3df083aSStefano Zampini } 2066a3df083aSStefano Zampini PetscFunctionReturn(0); 2067a3df083aSStefano Zampini } 2068a3df083aSStefano Zampini 2069a3df083aSStefano Zampini /* used just in bddc debug mode */ 2070a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 2071a3df083aSStefano Zampini { 2072a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2073a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2074a3df083aSStefano Zampini Mat An; 2075a3df083aSStefano Zampini PetscErrorCode ierr; 2076a3df083aSStefano Zampini 2077a3df083aSStefano Zampini PetscFunctionBegin; 2078a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 2079a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 2080a3df083aSStefano Zampini if (is1) { 20817dae84e0SHong Zhang ierr = MatCreateSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 2082a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 2083a3df083aSStefano Zampini } else { 2084a3df083aSStefano Zampini *B = An; 2085a3df083aSStefano Zampini } 2086a3df083aSStefano Zampini PetscFunctionReturn(0); 2087a3df083aSStefano Zampini } 2088a3df083aSStefano Zampini 20891cf9b237SStefano Zampini /* TODO: add reuse flag */ 20901cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 20911cf9b237SStefano Zampini { 20921cf9b237SStefano Zampini Mat Bt; 20931cf9b237SStefano Zampini PetscScalar *a,*bdata; 20941cf9b237SStefano Zampini const PetscInt *ii,*ij; 20951cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 20961cf9b237SStefano Zampini PetscBool flg_row; 20971cf9b237SStefano Zampini PetscErrorCode ierr; 20981cf9b237SStefano Zampini 20991cf9b237SStefano Zampini PetscFunctionBegin; 21001cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 21011cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 21021cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 21031cf9b237SStefano Zampini nnz = n; 21041cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 21051cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 21061cf9b237SStefano Zampini } 21071cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 21081cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 21091cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 21101cf9b237SStefano Zampini nnz = 0; 21111cf9b237SStefano Zampini bii[0] = 0; 21121cf9b237SStefano Zampini for (i=0;i<n;i++) { 21131cf9b237SStefano Zampini PetscInt j; 21141cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 21151cf9b237SStefano Zampini PetscScalar entry = a[j]; 21161cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 21171cf9b237SStefano Zampini bij[nnz] = ij[j]; 21181cf9b237SStefano Zampini bdata[nnz] = entry; 21191cf9b237SStefano Zampini nnz++; 21201cf9b237SStefano Zampini } 21211cf9b237SStefano Zampini } 21221cf9b237SStefano Zampini bii[i+1] = nnz; 21231cf9b237SStefano Zampini } 21241cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 21251cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 21261cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 21271cf9b237SStefano Zampini { 21281cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 21291cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 21301cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 21311cf9b237SStefano Zampini } 21321cf9b237SStefano Zampini *B = Bt; 21331cf9b237SStefano Zampini PetscFunctionReturn(0); 21341cf9b237SStefano Zampini } 21351cf9b237SStefano Zampini 2136c80a6c00SStefano Zampini PetscErrorCode PCBDDCDetectDisconnectedComponents(PC pc, PetscInt *ncc, IS* cc[], IS* primalv) 21374f1b2e48SStefano Zampini { 2138c80a6c00SStefano Zampini Mat B = NULL; 2139c80a6c00SStefano Zampini DM dm; 21404f1b2e48SStefano Zampini IS is_dummy,*cc_n; 21414f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 21424f1b2e48SStefano Zampini PCBDDCGraph graph; 2143c80a6c00SStefano Zampini PetscInt *xadj_filtered = NULL,*adjncy_filtered = NULL; 21444f1b2e48SStefano Zampini PetscInt i,n; 21454f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2146c80a6c00SStefano Zampini PetscBool isplex = PETSC_FALSE; 21474f1b2e48SStefano Zampini PetscErrorCode ierr; 21484f1b2e48SStefano Zampini 21494f1b2e48SStefano Zampini PetscFunctionBegin; 2150a2eca866SStefano Zampini if (ncc) *ncc = 0; 2151a2eca866SStefano Zampini if (cc) *cc = NULL; 2152a2eca866SStefano Zampini if (primalv) *primalv = NULL; 2153c80a6c00SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 2154c80a6c00SStefano Zampini ierr = PCGetDM(pc,&dm);CHKERRQ(ierr); 2155c80a6c00SStefano Zampini if (!dm) { 2156c80a6c00SStefano Zampini ierr = MatGetDM(pc->pmat,&dm);CHKERRQ(ierr); 2157c80a6c00SStefano Zampini } 2158c80a6c00SStefano Zampini if (dm) { 2159c80a6c00SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)dm,DMPLEX,&isplex);CHKERRQ(ierr); 2160c80a6c00SStefano Zampini } 2161c80a6c00SStefano Zampini if (isplex) { /* this code has been modified from plexpartition.c */ 2162c80a6c00SStefano Zampini PetscInt p, pStart, pEnd, a, adjSize, idx, size, nroots; 2163c80a6c00SStefano Zampini PetscInt *adj = NULL; 2164c80a6c00SStefano Zampini IS cellNumbering; 2165c80a6c00SStefano Zampini const PetscInt *cellNum; 2166c80a6c00SStefano Zampini PetscBool useCone, useClosure; 2167c80a6c00SStefano Zampini PetscSection section; 2168c80a6c00SStefano Zampini PetscSegBuffer adjBuffer; 2169c80a6c00SStefano Zampini PetscSF sfPoint; 2170c80a6c00SStefano Zampini PetscErrorCode ierr; 2171c80a6c00SStefano Zampini 2172c80a6c00SStefano Zampini PetscFunctionBegin; 2173c80a6c00SStefano Zampini ierr = DMPlexGetHeightStratum(dm, 0, &pStart, &pEnd);CHKERRQ(ierr); 2174c80a6c00SStefano Zampini ierr = DMGetPointSF(dm, &sfPoint);CHKERRQ(ierr); 2175c80a6c00SStefano Zampini ierr = PetscSFGetGraph(sfPoint, &nroots, NULL, NULL, NULL);CHKERRQ(ierr); 2176c80a6c00SStefano Zampini /* Build adjacency graph via a section/segbuffer */ 2177c80a6c00SStefano Zampini ierr = PetscSectionCreate(PetscObjectComm((PetscObject) dm), §ion);CHKERRQ(ierr); 2178c80a6c00SStefano Zampini ierr = PetscSectionSetChart(section, pStart, pEnd);CHKERRQ(ierr); 2179c80a6c00SStefano Zampini ierr = PetscSegBufferCreate(sizeof(PetscInt),1000,&adjBuffer);CHKERRQ(ierr); 2180c80a6c00SStefano Zampini /* Always use FVM adjacency to create partitioner graph */ 2181c80a6c00SStefano Zampini ierr = DMPlexGetAdjacencyUseCone(dm, &useCone);CHKERRQ(ierr); 2182c80a6c00SStefano Zampini ierr = DMPlexGetAdjacencyUseClosure(dm, &useClosure);CHKERRQ(ierr); 2183c80a6c00SStefano Zampini ierr = DMPlexSetAdjacencyUseCone(dm, PETSC_TRUE);CHKERRQ(ierr); 2184c80a6c00SStefano Zampini ierr = DMPlexSetAdjacencyUseClosure(dm, PETSC_FALSE);CHKERRQ(ierr); 2185956e2312SStefano Zampini ierr = DMPlexGetCellNumbering(dm, &cellNumbering);CHKERRQ(ierr); 2186c80a6c00SStefano Zampini ierr = ISGetIndices(cellNumbering, &cellNum);CHKERRQ(ierr); 2187c80a6c00SStefano Zampini for (n = 0, p = pStart; p < pEnd; p++) { 2188c80a6c00SStefano Zampini /* Skip non-owned cells in parallel (ParMetis expects no overlap) */ 2189c80a6c00SStefano Zampini if (nroots > 0) {if (cellNum[p] < 0) continue;} 2190c80a6c00SStefano Zampini adjSize = PETSC_DETERMINE; 2191c80a6c00SStefano Zampini ierr = DMPlexGetAdjacency(dm, p, &adjSize, &adj);CHKERRQ(ierr); 2192c80a6c00SStefano Zampini for (a = 0; a < adjSize; ++a) { 2193c80a6c00SStefano Zampini const PetscInt point = adj[a]; 21945cef3d0dSStefano Zampini if (pStart <= point && point < pEnd) { 2195c80a6c00SStefano Zampini PetscInt *PETSC_RESTRICT pBuf; 2196c80a6c00SStefano Zampini ierr = PetscSectionAddDof(section, p, 1);CHKERRQ(ierr); 2197c80a6c00SStefano Zampini ierr = PetscSegBufferGetInts(adjBuffer, 1, &pBuf);CHKERRQ(ierr); 2198c80a6c00SStefano Zampini *pBuf = point; 2199c80a6c00SStefano Zampini } 2200c80a6c00SStefano Zampini } 2201c80a6c00SStefano Zampini n++; 2202c80a6c00SStefano Zampini } 2203c80a6c00SStefano Zampini ierr = DMPlexSetAdjacencyUseCone(dm, useCone);CHKERRQ(ierr); 2204c80a6c00SStefano Zampini ierr = DMPlexSetAdjacencyUseClosure(dm, useClosure);CHKERRQ(ierr); 2205c80a6c00SStefano Zampini /* Derive CSR graph from section/segbuffer */ 2206c80a6c00SStefano Zampini ierr = PetscSectionSetUp(section);CHKERRQ(ierr); 2207c80a6c00SStefano Zampini ierr = PetscSectionGetStorageSize(section, &size);CHKERRQ(ierr); 2208c80a6c00SStefano Zampini ierr = PetscMalloc1(n+1, &xadj);CHKERRQ(ierr); 2209c80a6c00SStefano Zampini for (idx = 0, p = pStart; p < pEnd; p++) { 2210c80a6c00SStefano Zampini if (nroots > 0) {if (cellNum[p] < 0) continue;} 2211c80a6c00SStefano Zampini ierr = PetscSectionGetOffset(section, p, &(xadj[idx++]));CHKERRQ(ierr); 2212c80a6c00SStefano Zampini } 2213c80a6c00SStefano Zampini xadj[n] = size; 2214c80a6c00SStefano Zampini ierr = PetscSegBufferExtractAlloc(adjBuffer, &adjncy);CHKERRQ(ierr); 2215c80a6c00SStefano Zampini /* Clean up */ 2216c80a6c00SStefano Zampini ierr = PetscSegBufferDestroy(&adjBuffer);CHKERRQ(ierr); 2217c80a6c00SStefano Zampini ierr = PetscSectionDestroy(§ion);CHKERRQ(ierr); 2218c80a6c00SStefano Zampini ierr = PetscFree(adj);CHKERRQ(ierr); 2219c80a6c00SStefano Zampini graph->xadj = xadj; 2220c80a6c00SStefano Zampini graph->adjncy = adjncy; 2221c80a6c00SStefano Zampini } else { 2222c80a6c00SStefano Zampini Mat A; 2223c80a6c00SStefano Zampini PetscBool filter = PETSC_FALSE, isseqaij, flg_row; 2224c80a6c00SStefano Zampini 2225c80a6c00SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 222663c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 2227a2eca866SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 222863c961adSStefano Zampini PetscFunctionReturn(0); 222963c961adSStefano Zampini } 22304f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 22314f1b2e48SStefano Zampini if (!isseqaij && filter) { 22321cf9b237SStefano Zampini PetscBool isseqdense; 22331cf9b237SStefano Zampini 22341cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 22351cf9b237SStefano Zampini if (!isseqdense) { 22364f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 22371cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 22381cf9b237SStefano Zampini PetscScalar *array; 22391cf9b237SStefano Zampini PetscReal chop=1.e-6; 22401cf9b237SStefano Zampini 22411cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 22421cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 22431cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 22441cf9b237SStefano Zampini for (i=0;i<n;i++) { 22451cf9b237SStefano Zampini PetscInt j; 22461cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 22471cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 22481cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 22491cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 22501cf9b237SStefano Zampini } 22511cf9b237SStefano Zampini } 22521cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 22539d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 22541cf9b237SStefano Zampini } 22554f1b2e48SStefano Zampini } else { 2256c80a6c00SStefano Zampini ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); 22574f1b2e48SStefano Zampini B = A; 22584f1b2e48SStefano Zampini } 22594f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 22604f1b2e48SStefano Zampini 22614f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 22624f1b2e48SStefano Zampini if (filter) { 22634f1b2e48SStefano Zampini PetscScalar *data; 22644f1b2e48SStefano Zampini PetscInt j,cum; 22654f1b2e48SStefano Zampini 22664f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 22674f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 22684f1b2e48SStefano Zampini cum = 0; 22694f1b2e48SStefano Zampini for (i=0;i<n;i++) { 22704f1b2e48SStefano Zampini PetscInt t; 22714f1b2e48SStefano Zampini 22724f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 22734f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 22744f1b2e48SStefano Zampini continue; 22754f1b2e48SStefano Zampini } 22764f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 22774f1b2e48SStefano Zampini } 22784f1b2e48SStefano Zampini t = xadj_filtered[i]; 22794f1b2e48SStefano Zampini xadj_filtered[i] = cum; 22804f1b2e48SStefano Zampini cum += t; 22814f1b2e48SStefano Zampini } 22824f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 22834f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 22844f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 22854f1b2e48SStefano Zampini } else { 22864f1b2e48SStefano Zampini graph->xadj = xadj; 22874f1b2e48SStefano Zampini graph->adjncy = adjncy; 22884f1b2e48SStefano Zampini } 2289c80a6c00SStefano Zampini } 2290c80a6c00SStefano Zampini /* compute local connected components using PCBDDCGraph */ 2291c80a6c00SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 2292c80a6c00SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 2293c80a6c00SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2294c80a6c00SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 2295c80a6c00SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 22964f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 22974f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 2298c80a6c00SStefano Zampini 22994f1b2e48SStefano Zampini /* partial clean up */ 23004f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 2301c80a6c00SStefano Zampini if (B) { 2302c80a6c00SStefano Zampini PetscBool flg_row; 23034f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 23044f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 23054f1b2e48SStefano Zampini } 2306c80a6c00SStefano Zampini if (isplex) { 2307c80a6c00SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 2308c80a6c00SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 2309c80a6c00SStefano Zampini } 23104f1b2e48SStefano Zampini 23114f1b2e48SStefano Zampini /* get back data */ 2312c80a6c00SStefano Zampini if (isplex) { 2313c80a6c00SStefano Zampini if (ncc) *ncc = graph->ncc; 2314c80a6c00SStefano Zampini if (cc || primalv) { 2315c80a6c00SStefano Zampini Mat A; 2316c80a6c00SStefano Zampini PetscBT btv,btvt; 2317c80a6c00SStefano Zampini PetscSection subSection; 2318c80a6c00SStefano Zampini PetscInt *ids,cum,cump,*cids,*pids; 2319c80a6c00SStefano Zampini 2320c80a6c00SStefano Zampini ierr = DMPlexGetSubdomainSection(dm,&subSection);CHKERRQ(ierr); 2321c80a6c00SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 2322c80a6c00SStefano Zampini ierr = PetscMalloc3(A->rmap->n,&ids,graph->ncc+1,&cids,A->rmap->n,&pids);CHKERRQ(ierr); 2323c80a6c00SStefano Zampini ierr = PetscBTCreate(A->rmap->n,&btv);CHKERRQ(ierr); 2324c80a6c00SStefano Zampini ierr = PetscBTCreate(A->rmap->n,&btvt);CHKERRQ(ierr); 2325c80a6c00SStefano Zampini 2326c80a6c00SStefano Zampini cids[0] = 0; 2327c80a6c00SStefano Zampini for (i = 0, cump = 0, cum = 0; i < graph->ncc; i++) { 2328c80a6c00SStefano Zampini PetscInt j; 2329c80a6c00SStefano Zampini 2330c80a6c00SStefano Zampini ierr = PetscBTMemzero(A->rmap->n,btvt);CHKERRQ(ierr); 2331c80a6c00SStefano Zampini for (j = graph->cptr[i]; j < graph->cptr[i+1]; j++) { 2332c80a6c00SStefano Zampini PetscInt k, size, *closure = NULL, cell = graph->queue[j]; 2333c80a6c00SStefano Zampini 2334c80a6c00SStefano Zampini ierr = DMPlexGetTransitiveClosure(dm,cell,PETSC_TRUE,&size,&closure);CHKERRQ(ierr); 2335c80a6c00SStefano Zampini for (k = 0; k < 2*size; k += 2) { 2336c80a6c00SStefano Zampini PetscInt s, p = closure[k], off, dof, cdof; 2337c80a6c00SStefano Zampini 2338c80a6c00SStefano Zampini ierr = PetscSectionGetConstraintDof(subSection, p, &cdof);CHKERRQ(ierr); 2339c80a6c00SStefano Zampini ierr = PetscSectionGetOffset(subSection,p,&off);CHKERRQ(ierr); 2340c80a6c00SStefano Zampini ierr = PetscSectionGetDof(subSection,p,&dof);CHKERRQ(ierr); 2341c80a6c00SStefano Zampini for (s = 0; s < dof-cdof; s++) { 2342c80a6c00SStefano Zampini if (PetscBTLookupSet(btvt,off+s)) continue; 2343c80a6c00SStefano Zampini if (!PetscBTLookup(btv,off+s)) { 2344c80a6c00SStefano Zampini ids[cum++] = off+s; 2345c80a6c00SStefano Zampini } else { /* cross-vertex */ 2346c80a6c00SStefano Zampini pids[cump++] = off+s; 2347c80a6c00SStefano Zampini } 2348c80a6c00SStefano Zampini } 2349c80a6c00SStefano Zampini } 2350c80a6c00SStefano Zampini ierr = DMPlexRestoreTransitiveClosure(dm,cell,PETSC_TRUE,&size,&closure);CHKERRQ(ierr); 2351c80a6c00SStefano Zampini } 2352c80a6c00SStefano Zampini cids[i+1] = cum; 2353c80a6c00SStefano Zampini /* mark dofs as already assigned */ 2354c80a6c00SStefano Zampini for (j = cids[i]; j < cids[i+1]; j++) { 2355c80a6c00SStefano Zampini ierr = PetscBTSet(btv,ids[j]);CHKERRQ(ierr); 2356c80a6c00SStefano Zampini } 2357c80a6c00SStefano Zampini } 2358c80a6c00SStefano Zampini if (cc) { 2359c80a6c00SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 2360c80a6c00SStefano Zampini for (i = 0; i < graph->ncc; i++) { 2361c80a6c00SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cids[i+1]-cids[i],ids+cids[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 2362c80a6c00SStefano Zampini } 2363c80a6c00SStefano Zampini *cc = cc_n; 2364c80a6c00SStefano Zampini } 2365c80a6c00SStefano Zampini if (primalv) { 2366c80a6c00SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),cump,pids,PETSC_COPY_VALUES,primalv);CHKERRQ(ierr); 2367c80a6c00SStefano Zampini } 2368c80a6c00SStefano Zampini ierr = PetscFree3(ids,cids,pids);CHKERRQ(ierr); 2369c80a6c00SStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 2370c80a6c00SStefano Zampini ierr = PetscBTDestroy(&btvt);CHKERRQ(ierr); 2371c80a6c00SStefano Zampini } 2372c80a6c00SStefano Zampini } else { 23731cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 23741cf9b237SStefano Zampini if (cc) { 23754f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 23764f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 23774f1b2e48SStefano 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); 23784f1b2e48SStefano Zampini } 23794f1b2e48SStefano Zampini *cc = cc_n; 23801cf9b237SStefano Zampini } 2381c80a6c00SStefano Zampini } 23824f1b2e48SStefano Zampini /* clean up graph */ 23834f1b2e48SStefano Zampini graph->xadj = 0; 23844f1b2e48SStefano Zampini graph->adjncy = 0; 23854f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 23864f1b2e48SStefano Zampini PetscFunctionReturn(0); 23874f1b2e48SStefano Zampini } 23884f1b2e48SStefano Zampini 23895408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 23905408967cSStefano Zampini { 23915408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 23925408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2393dee84bffSStefano Zampini IS dirIS = NULL; 23944f1b2e48SStefano Zampini PetscInt i; 23955408967cSStefano Zampini PetscErrorCode ierr; 23965408967cSStefano Zampini 23975408967cSStefano Zampini PetscFunctionBegin; 2398dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 23995408967cSStefano Zampini if (zerodiag) { 24005408967cSStefano Zampini Mat A; 24015408967cSStefano Zampini Vec vec3_N; 24025408967cSStefano Zampini PetscScalar *vals; 24035408967cSStefano Zampini const PetscInt *idxs; 2404d12d3064SStefano Zampini PetscInt nz,*count; 24055408967cSStefano Zampini 24065408967cSStefano Zampini /* p0 */ 24075408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 24085408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 24095408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 24105408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 24114f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 24125408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24135408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 24145408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 24155408967cSStefano Zampini /* v_I */ 24165408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 24175408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 24185408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24195408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 24205408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 24215408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 24225408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24235408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 24245408967cSStefano Zampini if (dirIS) { 24255408967cSStefano Zampini PetscInt n; 24265408967cSStefano Zampini 24275408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 24285408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 24295408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 24305408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 24315408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 24325408967cSStefano Zampini } 24335408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 24345408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 24355408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 24365408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 2437669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 24385408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 24395408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 24409a962809SStefano 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])); 24415408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 24425408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 2443d12d3064SStefano Zampini 2444d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 2445d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 2446d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2447d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 2448d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2449d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 24509a962809SStefano 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]); 2451d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 2452d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 24535408967cSStefano Zampini } 2454dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 24555408967cSStefano Zampini 24565408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 24575408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 24584f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 24595408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 24604f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 24615408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 2462f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2463f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 246413903a91SSatish Balay if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %d instead of %g\n",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i); 2465f2a566d8SStefano Zampini } 24665408967cSStefano Zampini PetscFunctionReturn(0); 24675408967cSStefano Zampini } 24685408967cSStefano Zampini 2469339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 2470339f8db1SStefano Zampini { 2471339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 24724edc6404Sstefano_zampini IS pressures,zerodiag,zerodiag_save,*zerodiag_subs; 2473b0f5fe93SStefano Zampini PetscInt nz,n; 24744edc6404Sstefano_zampini PetscInt *interior_dofs,n_interior_dofs,nneu; 24754edc6404Sstefano_zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag,checkb; 2476339f8db1SStefano Zampini PetscErrorCode ierr; 2477339f8db1SStefano Zampini 2478339f8db1SStefano Zampini PetscFunctionBegin; 24799f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 24809f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 2481a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 2482a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 2483a3df083aSStefano Zampini } 2484a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 2485a3df083aSStefano Zampini pcbddc->benign_n = 0; 248628b8efb1Sstefano_zampini 248728b8efb1Sstefano_zampini /* if a local info on dofs is present, uses the last field for "pressures" (or fid by command line) 24884f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 24894f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 24904f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 24911ae86dd6SStefano Zampini since the local Schur complements are already SPD 24924f1b2e48SStefano Zampini */ 24934f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 24944f1b2e48SStefano Zampini have_null = PETSC_TRUE; 249540fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 24967fbe2174Sstefano_zampini IS iP = NULL; 24974f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 24984f1b2e48SStefano Zampini 249928b8efb1Sstefano_zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC benign options","PC");CHKERRQ(ierr); 250028b8efb1Sstefano_zampini ierr = PetscOptionsInt("-pc_bddc_pressure_field","Field id for pressures",NULL,p,&p,NULL);CHKERRQ(ierr); 250128b8efb1Sstefano_zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 250228b8efb1Sstefano_zampini if (p < 0 || p > pcbddc->n_ISForDofsLocal-1) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Invalid field id for pressures %D",p); 25034f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 25044f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 25054f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 25064f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 2507ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 25087fbe2174Sstefano_zampini /* remove zeroed out pressures if we are setting up a BDDC solver for a saddle-point FETI-DP */ 25097fbe2174Sstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lP",(PetscObject*)&iP);CHKERRQ(ierr); 25107fbe2174Sstefano_zampini if (iP) { 25117fbe2174Sstefano_zampini IS newpressures; 25127fbe2174Sstefano_zampini 25137fbe2174Sstefano_zampini ierr = ISDifference(pressures,iP,&newpressures);CHKERRQ(ierr); 25147fbe2174Sstefano_zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 25157fbe2174Sstefano_zampini pressures = newpressures; 25167fbe2174Sstefano_zampini } 251740fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 251840fa8d13SStefano Zampini if (!sorted) { 251940fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 252040fa8d13SStefano Zampini } 252140fa8d13SStefano Zampini } else { 252240fa8d13SStefano Zampini pressures = NULL; 252340fa8d13SStefano Zampini } 252497d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 252597d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 252627b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 252797d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 2528339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 2529339f8db1SStefano Zampini if (!sorted) { 2530339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 2531339f8db1SStefano Zampini } 25324edc6404Sstefano_zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 25334edc6404Sstefano_zampini zerodiag_save = zerodiag; 2534339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 25354f1b2e48SStefano Zampini if (!nz) { 25364f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 25374f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 253840fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 253940fa8d13SStefano Zampini } 25404f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 25414f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 25424f1b2e48SStefano Zampini zerodiag_subs = NULL; 25434f1b2e48SStefano Zampini pcbddc->benign_n = 0; 25441f4df5f7SStefano Zampini n_interior_dofs = 0; 25451f4df5f7SStefano Zampini interior_dofs = NULL; 25464edc6404Sstefano_zampini nneu = 0; 25474edc6404Sstefano_zampini if (pcbddc->NeumannBoundariesLocal) { 25484edc6404Sstefano_zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nneu);CHKERRQ(ierr); 25494edc6404Sstefano_zampini } 25503369cb78Sstefano_zampini checkb = (PetscBool)(!pcbddc->NeumannBoundariesLocal || pcbddc->current_level); 25514edc6404Sstefano_zampini if (checkb) { /* need to compute interior nodes */ 25521f4df5f7SStefano Zampini PetscInt n,i,j; 25531f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 25541f4df5f7SStefano Zampini PetscInt *iwork; 25551f4df5f7SStefano Zampini 25561f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 25571f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 25581f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 25591f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 256090648384SStefano Zampini for (i=1;i<n_neigh;i++) 25611f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 25621f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 25631f4df5f7SStefano Zampini for (i=0;i<n;i++) 25641f4df5f7SStefano Zampini if (!iwork[i]) 25651f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 25661f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 25671f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 25681f4df5f7SStefano Zampini } 25694f1b2e48SStefano Zampini if (has_null_pressures) { 25704f1b2e48SStefano Zampini IS *subs; 25714edc6404Sstefano_zampini PetscInt nsubs,i,j,nl; 25721f4df5f7SStefano Zampini const PetscInt *idxs; 25731f4df5f7SStefano Zampini PetscScalar *array; 25741f4df5f7SStefano Zampini Vec *work; 25751f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 25764f1b2e48SStefano Zampini 25774f1b2e48SStefano Zampini subs = pcbddc->local_subs; 25784f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 25791f4df5f7SStefano 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) */ 25804edc6404Sstefano_zampini if (checkb) { 25811f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 25821f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 25831f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 25841f4df5f7SStefano Zampini /* work[0] = 1_p */ 25851f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 25861f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 25871f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 25881f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 25891f4df5f7SStefano Zampini /* work[0] = 1_v */ 25901f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 25911f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 25921f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 25931f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 25941f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 25951f4df5f7SStefano Zampini } 25964f1b2e48SStefano Zampini if (nsubs > 1) { 25974f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 25984f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 25994f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 26004f1b2e48SStefano Zampini IS t_zerodiag_subs; 26014f1b2e48SStefano Zampini PetscInt nl; 26024f1b2e48SStefano Zampini 26034f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 26044f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 26054f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 26064f1b2e48SStefano Zampini if (nl) { 26074f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 26084f1b2e48SStefano Zampini 26094edc6404Sstefano_zampini if (checkb) { 26101f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 26111f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 26121f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 26131f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 26141f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 26151f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 26161f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 26171f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 26181f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 26191f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 26201f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 26211f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 26221f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 26231f4df5f7SStefano Zampini valid = PETSC_FALSE; 26241f4df5f7SStefano Zampini break; 26251f4df5f7SStefano Zampini } 26261f4df5f7SStefano Zampini } 26271f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 26281f4df5f7SStefano Zampini } 26296632bad2Sstefano_zampini if (valid && nneu) { 26306632bad2Sstefano_zampini const PetscInt *idxs; 26311f4df5f7SStefano Zampini PetscInt nzb; 26321f4df5f7SStefano Zampini 26336632bad2Sstefano_zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 26346632bad2Sstefano_zampini ierr = ISGlobalToLocalMappingApply(l2g,IS_GTOLM_DROP,nneu,idxs,&nzb,NULL);CHKERRQ(ierr); 26356632bad2Sstefano_zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 26361f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 26371f4df5f7SStefano Zampini } 26381f4df5f7SStefano Zampini if (valid && pressures) { 26394f1b2e48SStefano Zampini IS t_pressure_subs; 26404f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 26414f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 26424f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 26434f1b2e48SStefano Zampini } 26444f1b2e48SStefano Zampini if (valid) { 26454f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 26464f1b2e48SStefano Zampini pcbddc->benign_n++; 26474f1b2e48SStefano Zampini } else { 26484f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 26494f1b2e48SStefano Zampini } 26504f1b2e48SStefano Zampini } 26514f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 26524f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 26534f1b2e48SStefano Zampini } 26544f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 26554f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 26561f4df5f7SStefano Zampini 26576632bad2Sstefano_zampini if (nneu) valid = PETSC_FALSE; 26581f4df5f7SStefano Zampini if (valid && pressures) { 26594f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 26604f1b2e48SStefano Zampini } 26614edc6404Sstefano_zampini if (valid && checkb) { 26621f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 26631f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 26641f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 26651f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 26661f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 26671f4df5f7SStefano Zampini valid = PETSC_FALSE; 26681f4df5f7SStefano Zampini break; 26691f4df5f7SStefano Zampini } 26701f4df5f7SStefano Zampini } 26711f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 26721f4df5f7SStefano Zampini } 26734f1b2e48SStefano Zampini if (valid) { 26744f1b2e48SStefano Zampini pcbddc->benign_n = 1; 2675ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 26764f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 26774f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 26784f1b2e48SStefano Zampini } 26794f1b2e48SStefano Zampini } 26804edc6404Sstefano_zampini if (checkb) { 26811f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 26824f1b2e48SStefano Zampini } 26831f4df5f7SStefano Zampini } 26841f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 26854f1b2e48SStefano Zampini 26864f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2687b9b0e38cSStefano Zampini PetscInt n; 2688b9b0e38cSStefano Zampini 26894f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 26904f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2691b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2692b9b0e38cSStefano Zampini if (n) { 26934f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 26944f1b2e48SStefano Zampini have_null = PETSC_FALSE; 26954f1b2e48SStefano Zampini } 2696b9b0e38cSStefano Zampini } 26974f1b2e48SStefano Zampini 26984f1b2e48SStefano Zampini /* final check for null pressures */ 26994f1b2e48SStefano Zampini if (zerodiag && pressures) { 27004f1b2e48SStefano Zampini PetscInt nz,np; 27014f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 27024f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 27034f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 27044f1b2e48SStefano Zampini } 27054f1b2e48SStefano Zampini 27064f1b2e48SStefano Zampini if (recompute_zerodiag) { 27074f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 27084f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 27094f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 27104f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 27114f1b2e48SStefano Zampini } else { 27124f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 27134f1b2e48SStefano Zampini 27144f1b2e48SStefano Zampini nzn = 0; 27154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 27164f1b2e48SStefano Zampini PetscInt ns; 27174f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 27184f1b2e48SStefano Zampini nzn += ns; 27194f1b2e48SStefano Zampini } 27204f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 27214f1b2e48SStefano Zampini nzn = 0; 27224f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 27234f1b2e48SStefano Zampini PetscInt ns,*idxs; 27244f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 27254f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 27264f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 27274f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 27284f1b2e48SStefano Zampini nzn += ns; 27294f1b2e48SStefano Zampini } 27304f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 27314f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 27324f1b2e48SStefano Zampini } 27334f1b2e48SStefano Zampini have_null = PETSC_FALSE; 27344f1b2e48SStefano Zampini } 27354f1b2e48SStefano Zampini 2736669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2737a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2738a198735bSStefano Zampini Mat A,loc_divudotp; 2739a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2740a198735bSStefano Zampini IS row,col,isused = NULL; 2741a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2742a198735bSStefano Zampini 27431f4df5f7SStefano Zampini if (pressures) { 27441f4df5f7SStefano Zampini isused = pressures; 27451f4df5f7SStefano Zampini } else { 27464edc6404Sstefano_zampini isused = zerodiag_save; 27471f4df5f7SStefano Zampini } 2748a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2749669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 27501ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 27511ae86dd6SStefano 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"); 2752a198735bSStefano Zampini n_isused = 0; 2753a198735bSStefano Zampini if (isused) { 2754a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2755a198735bSStefano Zampini } 2756a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2757a198735bSStefano Zampini st = st-n_isused; 27581ae86dd6SStefano Zampini if (n) { 2759a198735bSStefano Zampini const PetscInt *gidxs; 2760a198735bSStefano Zampini 27617dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2762a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2763a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2764a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2765a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2766a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 27671ae86dd6SStefano Zampini } else { 2768a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2769a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2770a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2771a198735bSStefano Zampini } 2772a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2773a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2774a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2775a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2776a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2777a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2778a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2779a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2780a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2781a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2782a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2783a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2784a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2785a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 27861ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 27871ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 27881ae86dd6SStefano Zampini } 27894edc6404Sstefano_zampini ierr = ISDestroy(&zerodiag_save);CHKERRQ(ierr); 2790b3afcdbeSStefano Zampini 2791b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 27924f1b2e48SStefano Zampini if (has_null_pressures) { 27934f1b2e48SStefano Zampini IS zerodiagc; 27944f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 27954f1b2e48SStefano Zampini PetscInt i,s,*nnz; 27964f1b2e48SStefano Zampini 27974f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2798339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 2799339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 2800339f8db1SStefano Zampini /* local change of basis for pressures */ 2801339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 280297d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2803339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2804339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2805339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 28064f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 28074f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 28084f1b2e48SStefano Zampini PetscInt nzs,j; 28094f1b2e48SStefano Zampini 28104f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 28114f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 28124f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 28134f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 28144f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 28154f1b2e48SStefano Zampini } 2816339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2817339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2818339f8db1SStefano Zampini /* set identity on velocities */ 2819339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 2820339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 2821339f8db1SStefano Zampini } 28224f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 28234f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 28249f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 28254f1b2e48SStefano 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); 2826339f8db1SStefano Zampini /* set change on pressures */ 28274f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 28284f1b2e48SStefano Zampini PetscScalar *array; 28294f1b2e48SStefano Zampini PetscInt nzs; 28304f1b2e48SStefano Zampini 28314f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 28324f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 28334f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2834339f8db1SStefano Zampini PetscScalar vals[2]; 2835339f8db1SStefano Zampini PetscInt cols[2]; 2836339f8db1SStefano Zampini 2837339f8db1SStefano Zampini cols[0] = idxs[i]; 28384f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2839339f8db1SStefano Zampini vals[0] = 1.; 2840b0f5fe93SStefano Zampini vals[1] = 1.; 28414f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2842339f8db1SStefano Zampini } 28434f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 28444f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 28454f1b2e48SStefano Zampini array[nzs-1] = 1.; 28464f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 28474f1b2e48SStefano Zampini /* store local idxs for p0 */ 28484f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 28494f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2850339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 28514f1b2e48SStefano Zampini } 2852339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2853339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2854a3df083aSStefano Zampini /* project if needed */ 2855a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 28561dd7afcfSStefano Zampini Mat M; 28571dd7afcfSStefano Zampini 28581dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2859339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 28601dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 28611dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2862a3df083aSStefano Zampini } 28634f1b2e48SStefano Zampini /* store global idxs for p0 */ 28644f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2865339f8db1SStefano Zampini } 2866ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 28674f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 2868b0f5fe93SStefano Zampini 2869b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 2870b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 287127b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 287227b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2873339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2874339f8db1SStefano Zampini PetscFunctionReturn(0); 2875339f8db1SStefano Zampini } 2876339f8db1SStefano Zampini 2877015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 2878efc2fbd9SStefano Zampini { 2879efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2880de9d7bd0SStefano Zampini PetscScalar *array; 2881efc2fbd9SStefano Zampini PetscErrorCode ierr; 2882efc2fbd9SStefano Zampini 2883efc2fbd9SStefano Zampini PetscFunctionBegin; 2884efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 2885efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 28864f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2887efc2fbd9SStefano Zampini } 2888de9d7bd0SStefano Zampini if (get) { 2889efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 28904f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 28914f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 2892efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 2893de9d7bd0SStefano Zampini } else { 2894de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 2895de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2896de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2897de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 2898efc2fbd9SStefano Zampini } 2899efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2900efc2fbd9SStefano Zampini } 2901efc2fbd9SStefano Zampini 2902c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 2903c263805aSStefano Zampini { 2904c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2905c263805aSStefano Zampini PetscErrorCode ierr; 2906c263805aSStefano Zampini 2907c263805aSStefano Zampini PetscFunctionBegin; 2908c263805aSStefano Zampini /* TODO: add error checking 2909c263805aSStefano Zampini - avoid nested pop (or push) calls. 2910c263805aSStefano Zampini - cannot push before pop. 29111c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 2912c263805aSStefano Zampini */ 29134f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2914efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2915efc2fbd9SStefano Zampini } 2916c263805aSStefano Zampini if (pop) { 2917a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 29184f1b2e48SStefano Zampini IS is_p0; 29194f1b2e48SStefano Zampini MatReuse reuse; 2920c263805aSStefano Zampini 2921c263805aSStefano Zampini /* extract B_0 */ 29224f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 29234f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 29244f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 29254f1b2e48SStefano Zampini } 29264f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 29277dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 2928c263805aSStefano Zampini /* remove rows and cols from local problem */ 2929c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 293097d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 29314f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 29324f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 2933a3df083aSStefano Zampini } else { 2934a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2935a3df083aSStefano Zampini PetscScalar *vals; 2936a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 2937a3df083aSStefano Zampini 2938a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 2939a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 2940a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 29410b5adadeSStefano Zampini PetscInt *nnz; 2942a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 2943a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 2944a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2945331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 2946331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2947331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 2948331e053bSStefano Zampini nnz[i] = n - nnz[i]; 2949331e053bSStefano Zampini } 2950331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 2951331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2952331e053bSStefano Zampini } 2953a3df083aSStefano Zampini 2954a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2955a3df083aSStefano Zampini PetscScalar *array; 2956a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 2957a3df083aSStefano Zampini 2958a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 2959a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2960a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2961a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 2962a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2963a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 2964a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 2965a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 2966a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 2967a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 2968a3df083aSStefano Zampini cum = 0; 2969a3df083aSStefano Zampini for (j=0;j<n;j++) { 297022db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 2971a3df083aSStefano Zampini vals[cum] = array[j]; 2972a3df083aSStefano Zampini idxs_ins[cum] = j; 2973a3df083aSStefano Zampini cum++; 2974a3df083aSStefano Zampini } 2975a3df083aSStefano Zampini } 2976a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 2977a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 2978a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2979a3df083aSStefano Zampini } 2980a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2981a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2982a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 2983a3df083aSStefano Zampini } 2984c263805aSStefano Zampini } else { /* push */ 2985a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 29864f1b2e48SStefano Zampini PetscInt i; 29874f1b2e48SStefano Zampini 29884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 29894f1b2e48SStefano Zampini PetscScalar *B0_vals; 29904f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 29914f1b2e48SStefano Zampini 29924f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 29934f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 29947b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 29954f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 29964f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 29974f1b2e48SStefano Zampini } 2998c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2999c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3000a3df083aSStefano Zampini } else { 3001a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 3002a3df083aSStefano Zampini } 3003c263805aSStefano Zampini } 3004c263805aSStefano Zampini PetscFunctionReturn(0); 3005c263805aSStefano Zampini } 3006c263805aSStefano Zampini 300708122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 3008b1b3d7a2SStefano Zampini { 3009b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 301008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 301108122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 301208122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 301308122e43SStefano Zampini PetscScalar *work,lwork; 301408122e43SStefano Zampini PetscScalar *St,*S,*eigv; 301508122e43SStefano Zampini PetscScalar *Sarray,*Starray; 301608122e43SStefano Zampini PetscReal *eigs,thresh; 30171b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 3018f6f667cfSStefano Zampini PetscBool allocated_S_St; 301908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 302008122e43SStefano Zampini PetscReal *rwork; 302108122e43SStefano Zampini #endif 3022b1b3d7a2SStefano Zampini PetscErrorCode ierr; 3023b1b3d7a2SStefano Zampini 3024b1b3d7a2SStefano Zampini PetscFunctionBegin; 3025b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 3026af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 3027d750e150Sstefano_zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\nRerun with -sub_schurs_hermitian 1 -sub_schurs_posdef 1 if the problem is SPD",sub_schurs->is_hermitian,sub_schurs->is_posdef); 302806a4e24aSStefano Zampini 3029fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3030fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3031fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 3032fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 30331575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3034fd14bc51SStefano Zampini } 3035fd14bc51SStefano Zampini 3036e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 3037e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 3038e496cd5dSStefano Zampini } 3039e496cd5dSStefano Zampini 304008122e43SStefano Zampini /* max size of subsets */ 304108122e43SStefano Zampini mss = 0; 304208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 304308122e43SStefano Zampini PetscInt subset_size; 3044862806e4SStefano Zampini 304508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 304608122e43SStefano Zampini mss = PetscMax(mss,subset_size); 304708122e43SStefano Zampini } 304808122e43SStefano Zampini 304908122e43SStefano Zampini /* min/max and threshold */ 305008122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 3051f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 305208122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 3053f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 3054f6f667cfSStefano Zampini if (nmin) { 3055f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 3056f6f667cfSStefano Zampini } 305708122e43SStefano Zampini 305808122e43SStefano Zampini /* allocate lapack workspace */ 305908122e43SStefano Zampini cum = cum2 = 0; 306008122e43SStefano Zampini maxneigs = 0; 306108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 306208122e43SStefano Zampini PetscInt n,subset_size; 3063f6f667cfSStefano Zampini 306408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 306508122e43SStefano Zampini n = PetscMin(subset_size,nmax); 30669162d606SStefano Zampini cum += subset_size; 30679162d606SStefano Zampini cum2 += subset_size*n; 306808122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 306908122e43SStefano Zampini } 307008122e43SStefano Zampini if (mss) { 30719ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 307208122e43SStefano Zampini PetscBLASInt B_itype = 1; 307308122e43SStefano Zampini PetscBLASInt B_N = mss; 30744c6709b3SStefano Zampini PetscReal zero = 0.0; 30754c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 307608122e43SStefano Zampini 307708122e43SStefano Zampini B_lwork = -1; 307808122e43SStefano Zampini S = NULL; 307908122e43SStefano Zampini St = NULL; 3080a58a30b4SStefano Zampini eigs = NULL; 3081a58a30b4SStefano Zampini eigv = NULL; 3082a58a30b4SStefano Zampini B_iwork = NULL; 3083a58a30b4SStefano Zampini B_ifail = NULL; 3084d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3085d1710679SStefano Zampini rwork = NULL; 3086d1710679SStefano Zampini #endif 30878bec7fa6SStefano Zampini thresh = 1.0; 308808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 308908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 309008122e43SStefano 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)); 309108122e43SStefano Zampini #else 309208122e43SStefano 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)); 309308122e43SStefano Zampini #endif 309408122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 309508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 309608122e43SStefano Zampini } else { 309708122e43SStefano Zampini /* TODO */ 309808122e43SStefano Zampini } 309908122e43SStefano Zampini } else { 310008122e43SStefano Zampini lwork = 0; 310108122e43SStefano Zampini } 310208122e43SStefano Zampini 310308122e43SStefano Zampini nv = 0; 3104d62866d3SStefano 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) */ 3105d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 310608122e43SStefano Zampini } 31074c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 3108f6f667cfSStefano Zampini if (allocated_S_St) { 3109f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 3110f6f667cfSStefano Zampini } 3111f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 311208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 311308122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 311408122e43SStefano Zampini #endif 31159162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 31169162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 31179162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 311808122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 31199162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 312008122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 312108122e43SStefano Zampini 312208122e43SStefano Zampini maxneigs = 0; 312372b8c272SStefano Zampini cum = cumarray = 0; 31249162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 31259162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 3126d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 312708122e43SStefano Zampini const PetscInt *idxs; 312808122e43SStefano Zampini 3129d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 313008122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 313108122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 313208122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 313308122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 31349162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 31359162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 313608122e43SStefano Zampini } 3137d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 313808122e43SStefano Zampini } 313908122e43SStefano Zampini 314008122e43SStefano Zampini if (mss) { /* multilevel */ 314108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 314208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 314308122e43SStefano Zampini } 314408122e43SStefano Zampini 3145ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 314608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 314708122e43SStefano Zampini const PetscInt *idxs; 31489d54b7f4SStefano Zampini PetscReal upper,lower; 3149862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 315008122e43SStefano Zampini PetscBLASInt B_N; 3151aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 315208122e43SStefano Zampini 31539d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 31549d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 31559d54b7f4SStefano Zampini lower = thresh; 31569d54b7f4SStefano Zampini } else { 31579d54b7f4SStefano Zampini upper = 1./thresh; 31589d54b7f4SStefano Zampini lower = 0.; 31599d54b7f4SStefano Zampini } 3160862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 3161ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 3162f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 3163f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 31649ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 3165aff50787SStefano Zampini PetscInt j,k; 3166aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 3167aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 3168aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 316908122e43SStefano Zampini } 317008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 3171aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 3172aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 3173aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 3174aff50787SStefano Zampini } 317508122e43SStefano Zampini } 317608122e43SStefano Zampini } else { 317708122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 317808122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 317908122e43SStefano Zampini } 31808bec7fa6SStefano Zampini } else { 3181f6f667cfSStefano Zampini S = Sarray + cumarray; 3182f6f667cfSStefano Zampini St = Starray + cumarray; 31838bec7fa6SStefano Zampini } 3184aff50787SStefano Zampini /* see if we can save some work */ 3185b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 3186aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 3187aff50787SStefano Zampini } 3188aff50787SStefano Zampini 3189b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 3190aff50787SStefano Zampini B_neigs = 0; 3191aff50787SStefano Zampini } else { 31929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 319308122e43SStefano Zampini PetscBLASInt B_itype = 1; 3194f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 31954c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 31969552c7c7SStefano Zampini PetscInt nmin_s; 3197b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 319808122e43SStefano Zampini 3199fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3200eee23b56SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d size %d count %d fid %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]+1,pcbddc->mat_graph->which_dof[idxs[0]]); 3201fd14bc51SStefano Zampini } 3202d16cbb6bSStefano Zampini 3203b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 3204b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 3205b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 3206b7ab4a40SStefano Zampini } 3207b7ab4a40SStefano Zampini 320808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3209b7ab4a40SStefano Zampini if (compute_range) { 3210d16cbb6bSStefano Zampini 3211d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 321208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32139d54b7f4SStefano 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)); 321408122e43SStefano Zampini #else 32159d54b7f4SStefano 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)); 321608122e43SStefano Zampini #endif 3217b7ab4a40SStefano Zampini } else if (!same_data) { 3218d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 3219d16cbb6bSStefano Zampini B_IL = 1; 3220d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 32219d54b7f4SStefano 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)); 3222d16cbb6bSStefano Zampini #else 32239d54b7f4SStefano 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)); 3224d16cbb6bSStefano Zampini #endif 3225b03ebc13SStefano Zampini } else { /* same_data is true, so just get the adaptive functional requested by the user */ 3226b7ab4a40SStefano Zampini PetscInt k; 3227b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 3228b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 3229b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 3230b7ab4a40SStefano Zampini nmin = nmax; 3231b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 3232b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 3233b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 3234b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 3235b7ab4a40SStefano Zampini } 3236d16cbb6bSStefano Zampini } 323708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 323808122e43SStefano Zampini if (B_ierr) { 32396c4ed002SBarry 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); 32406c4ed002SBarry 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); 32416c4ed002SBarry 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); 324208122e43SStefano Zampini } 324308122e43SStefano Zampini 324408122e43SStefano Zampini if (B_neigs > nmax) { 3245fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3246fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 3247fd14bc51SStefano Zampini } 32489d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 324908122e43SStefano Zampini B_neigs = nmax; 325008122e43SStefano Zampini } 325108122e43SStefano Zampini 32529552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 32539552c7c7SStefano Zampini if (B_neigs < nmin_s) { 325408122e43SStefano Zampini PetscBLASInt B_neigs2; 325508122e43SStefano Zampini 32569d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3257f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 32589d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 32599d54b7f4SStefano Zampini } else { 32609d54b7f4SStefano Zampini B_IL = B_neigs + 1; 32619d54b7f4SStefano Zampini B_IU = nmin_s; 32629d54b7f4SStefano Zampini } 3263fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3264fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 3265fd14bc51SStefano Zampini } 32669ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 32671ae86dd6SStefano Zampini PetscInt j,k; 326808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 32691ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 32701ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 32711ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 327208122e43SStefano Zampini } 327308122e43SStefano Zampini } 327408122e43SStefano Zampini } else { 327508122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 327608122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 327708122e43SStefano Zampini } 327808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 327908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32809d54b7f4SStefano 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)); 328108122e43SStefano Zampini #else 32829d54b7f4SStefano 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)); 328308122e43SStefano Zampini #endif 328408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 328508122e43SStefano Zampini B_neigs += B_neigs2; 328608122e43SStefano Zampini } 328708122e43SStefano Zampini if (B_ierr) { 32886c4ed002SBarry 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); 32896c4ed002SBarry 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); 32906c4ed002SBarry 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); 329108122e43SStefano Zampini } 3292fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3293ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 329408122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 329508122e43SStefano Zampini if (eigs[j] == 0.0) { 3296ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 329708122e43SStefano Zampini } else { 32989d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3299ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 33009d54b7f4SStefano Zampini } else { 33019d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 33029d54b7f4SStefano Zampini } 3303fd14bc51SStefano Zampini } 330408122e43SStefano Zampini } 330508122e43SStefano Zampini } 330608122e43SStefano Zampini } else { 330708122e43SStefano Zampini /* TODO */ 330808122e43SStefano Zampini } 3309aff50787SStefano Zampini } 33106c3e6151SStefano Zampini /* change the basis back to the original one */ 33116c3e6151SStefano Zampini if (sub_schurs->change) { 331272b8c272SStefano Zampini Mat change,phi,phit; 33136c3e6151SStefano Zampini 331403dfb2d7SStefano Zampini if (pcbddc->dbg_flag > 2) { 33156c3e6151SStefano Zampini PetscInt ii; 33166c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 33176c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 33186c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 3319684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3320684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 3321684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 3322684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3323684229deSStefano Zampini #else 33246c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 3325684229deSStefano Zampini #endif 33266c3e6151SStefano Zampini } 33276c3e6151SStefano Zampini } 33286c3e6151SStefano Zampini } 332972b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 33306c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 333172b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 33326c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 33336c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 33346c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 33356c3e6151SStefano Zampini } 33368bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 33378bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 33389162d606SStefano Zampini if (B_neigs) { 33399162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 3340fd14bc51SStefano Zampini 3341fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 33429552c7c7SStefano Zampini PetscInt ii; 33439552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 3344ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 33459552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 3346ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3347ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3348ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3349ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3350ac47001eSStefano Zampini #else 3351ac47001eSStefano 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); 3352ac47001eSStefano Zampini #endif 33539552c7c7SStefano Zampini } 33549552c7c7SStefano Zampini } 3355fd14bc51SStefano Zampini } 33569162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 33579162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 33589162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 33599162d606SStefano Zampini cum++; 336008122e43SStefano Zampini } 336108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 336208122e43SStefano Zampini /* shift for next computation */ 336308122e43SStefano Zampini cumarray += subset_size*subset_size; 336408122e43SStefano Zampini } 3365fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3366fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3367fd14bc51SStefano Zampini } 336808122e43SStefano Zampini 336908122e43SStefano Zampini if (mss) { 337008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 337108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 3372f6f667cfSStefano Zampini /* destroy matrices (junk) */ 3373f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 3374f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 337508122e43SStefano Zampini } 3376f6f667cfSStefano Zampini if (allocated_S_St) { 3377f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 3378f6f667cfSStefano Zampini } 3379f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 338008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 338108122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 338208122e43SStefano Zampini #endif 338308122e43SStefano Zampini if (pcbddc->dbg_flag) { 33841b968477SStefano Zampini PetscInt maxneigs_r; 3385b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 33869b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 338708122e43SStefano Zampini } 338808122e43SStefano Zampini PetscFunctionReturn(0); 338908122e43SStefano Zampini } 3390b1b3d7a2SStefano Zampini 3391c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 3392c8587f34SStefano Zampini { 33938629588bSStefano Zampini PetscScalar *coarse_submat_vals; 3394c8587f34SStefano Zampini PetscErrorCode ierr; 3395c8587f34SStefano Zampini 3396c8587f34SStefano Zampini PetscFunctionBegin; 3397f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 33985e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 3399c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 3400c8587f34SStefano Zampini 3401684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 34020fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 3403684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3404c8587f34SStefano Zampini 34058629588bSStefano Zampini /* 34068629588bSStefano Zampini Setup local correction and local part of coarse basis. 34078629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 34088629588bSStefano Zampini */ 340947f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 34108629588bSStefano Zampini 34118629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 34128629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 34138629588bSStefano Zampini 34148629588bSStefano Zampini /* free */ 34158629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 3416c8587f34SStefano Zampini PetscFunctionReturn(0); 3417c8587f34SStefano Zampini } 3418c8587f34SStefano Zampini 3419674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 3420674ae819SStefano Zampini { 3421674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3422674ae819SStefano Zampini PetscErrorCode ierr; 3423674ae819SStefano Zampini 3424674ae819SStefano Zampini PetscFunctionBegin; 3425674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 342630368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3427674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 3428785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3429674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 3430f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3431f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3432785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 343363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 343463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 3435674ae819SStefano Zampini PetscFunctionReturn(0); 3436674ae819SStefano Zampini } 3437674ae819SStefano Zampini 3438674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 3439674ae819SStefano Zampini { 3440674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 34414f1b2e48SStefano Zampini PetscInt i; 3442674ae819SStefano Zampini PetscErrorCode ierr; 3443674ae819SStefano Zampini 3444674ae819SStefano Zampini PetscFunctionBegin; 34451e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 34461e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 3447a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 3448b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3449674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 345016909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 34511dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 3452674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3453669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 3454fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 34559326c5c6Sstefano_zampini ierr = PCBDDCGraphDestroy(&pcbddc->mat_graph);CHKERRQ(ierr); 34564f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 34574f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 34584f1b2e48SStefano Zampini } 3459e68a0315Sstefano_zampini pcbddc->n_local_subs = 0; 34604f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 3461e68a0315Sstefano_zampini ierr = PCBDDCSubSchursDestroy(&pcbddc->sub_schurs);CHKERRQ(ierr); 3462c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 34638af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 3464674ae819SStefano Zampini PetscFunctionReturn(0); 3465674ae819SStefano Zampini } 3466674ae819SStefano Zampini 3467674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 3468674ae819SStefano Zampini { 3469674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3470674ae819SStefano Zampini PetscErrorCode ierr; 3471674ae819SStefano Zampini 3472674ae819SStefano Zampini PetscFunctionBegin; 3473674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 347458da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 3475ca92afb2SStefano Zampini PetscScalar *array; 347606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 347706656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 347858da7f69SStefano Zampini } 3479674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3480674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 348115aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 348215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3483674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 3484674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 3485674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 348606656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 3487674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3488674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 34898ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 3490674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 3491674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 3492674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 34939326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_D);CHKERRQ(ierr); 34949326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 34959326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3496f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3497727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 34980e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3499f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 350070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 350181d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 35020369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 35031dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 35044f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 35058b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 3506ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 3507ca92afb2SStefano Zampini PetscInt i; 3508ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3509ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 3510ca92afb2SStefano Zampini } 3511ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 3512ca92afb2SStefano Zampini } 35134f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 3514674ae819SStefano Zampini PetscFunctionReturn(0); 3515674ae819SStefano Zampini } 3516674ae819SStefano Zampini 3517f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 35186bfb1811SStefano Zampini { 35196bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 35206bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 35216bfb1811SStefano Zampini VecType impVecType; 35224f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 35236bfb1811SStefano Zampini PetscErrorCode ierr; 35246bfb1811SStefano Zampini 35256bfb1811SStefano Zampini PetscFunctionBegin; 35264f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 3527b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 35286bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 3529e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 3530e7b262bdSStefano Zampini /* R nodes */ 3531e7b262bdSStefano Zampini old_size = -1; 3532e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 3533e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 3534e7b262bdSStefano Zampini } 3535e7b262bdSStefano Zampini if (n_R != old_size) { 3536e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3537e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 35386bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 35396bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 35406bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 35416bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 3542e7b262bdSStefano Zampini } 3543e7b262bdSStefano Zampini /* local primal dofs */ 3544e7b262bdSStefano Zampini old_size = -1; 3545e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 3546e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 3547e7b262bdSStefano Zampini } 3548e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 3549e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 355083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 3551e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 35526bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 3553e7b262bdSStefano Zampini } 3554e7b262bdSStefano Zampini /* local explicit constraints */ 3555e7b262bdSStefano Zampini old_size = -1; 3556e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 3557e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 3558e7b262bdSStefano Zampini } 3559e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 3560e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 356183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 356283b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 356383b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 356483b7ccabSStefano Zampini } 35656bfb1811SStefano Zampini PetscFunctionReturn(0); 35666bfb1811SStefano Zampini } 35676bfb1811SStefano Zampini 356847f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 356988ebb749SStefano Zampini { 357025084f0cSStefano Zampini PetscErrorCode ierr; 357125084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 357288ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 357388ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3574d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 357525084f0cSStefano Zampini /* submatrices of local problem */ 357680677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 357706656605SStefano Zampini /* submatrices of local coarse problem */ 357806656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 357925084f0cSStefano Zampini /* working matrices */ 358006656605SStefano Zampini Mat C_CR; 358125084f0cSStefano Zampini /* additional working stuff */ 358206656605SStefano Zampini PC pc_R; 3583c58f9fdbSStefano Zampini Mat F,Brhs = NULL; 35845cbda25cSStefano Zampini Vec dummy_vec; 3585c58f9fdbSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction,sparserhs; 358625084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 358706656605SStefano Zampini PetscScalar *work; 358806656605SStefano Zampini PetscInt *idx_V_B; 3589ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 359006656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 3591ffd830a3SStefano Zampini 359225084f0cSStefano Zampini /* some shortcuts to scalars */ 359306656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 359488ebb749SStefano Zampini 359588ebb749SStefano Zampini PetscFunctionBegin; 35969a962809SStefano 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"); 3597ffd830a3SStefano Zampini 3598ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 3599b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 36004f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 3601b371cd4fSStefano Zampini n_B = pcis->n_B; 3602b371cd4fSStefano Zampini n_D = pcis->n - n_B; 360388ebb749SStefano Zampini n_R = pcis->n - n_vertices; 360488ebb749SStefano Zampini 360588ebb749SStefano Zampini /* vertices in boundary numbering */ 3606785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 36070e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 36086c4ed002SBarry Smith if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D\n",n_vertices,i); 360988ebb749SStefano Zampini 361006656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 3611019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 361206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 361306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 361406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 361506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 361606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 361706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 361806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 361906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 362006656605SStefano Zampini 362106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 362206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 362306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 362406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 362506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 3626ffd830a3SStefano Zampini lda_rhs = n_R; 3627a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 362806656605SStefano Zampini if (isLU || isILU || isCHOL) { 362906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 3630b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 3631df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3632d62866d3SStefano Zampini MatFactorType type; 3633d62866d3SStefano Zampini 3634df4d28bfSStefano Zampini F = reuse_solver->F; 36356816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 3636d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 3637ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 363822db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 363906656605SStefano Zampini } else { 364006656605SStefano Zampini F = NULL; 364106656605SStefano Zampini } 364206656605SStefano Zampini 3643c58f9fdbSStefano Zampini /* determine if we can use a sparse right-hand side */ 3644c58f9fdbSStefano Zampini sparserhs = PETSC_FALSE; 3645c58f9fdbSStefano Zampini if (F) { 3646ea799195SBarry Smith MatSolverType solver; 3647c58f9fdbSStefano Zampini 36483ca39a21SBarry Smith ierr = MatFactorGetSolverType(F,&solver);CHKERRQ(ierr); 3649c58f9fdbSStefano Zampini ierr = PetscStrcmp(solver,MATSOLVERMUMPS,&sparserhs);CHKERRQ(ierr); 3650c58f9fdbSStefano Zampini } 3651c58f9fdbSStefano Zampini 3652ffd830a3SStefano Zampini /* allocate workspace */ 3653ffd830a3SStefano Zampini n = 0; 3654ffd830a3SStefano Zampini if (n_constraints) { 3655ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 3656ffd830a3SStefano Zampini } 3657ffd830a3SStefano Zampini if (n_vertices) { 3658ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 3659ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 3660ffd830a3SStefano Zampini } 36612a3a6641Sstefano_zampini if (!pcbddc->symmetric_primal) { 36622a3a6641Sstefano_zampini n = PetscMax(2*lda_rhs*pcbddc->local_primal_size,n); 36632a3a6641Sstefano_zampini } 3664ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 3665ffd830a3SStefano Zampini 36665cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 36675cbda25cSStefano Zampini dummy_vec = NULL; 36685cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 36695cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 36705cbda25cSStefano Zampini } 36715cbda25cSStefano Zampini 367288ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 367388ebb749SStefano Zampini if (n_constraints) { 367472b8c272SStefano Zampini Mat M1,M2,M3,C_B; 367506656605SStefano Zampini IS is_aux; 367680677318SStefano Zampini PetscScalar *array,*array2; 367706656605SStefano Zampini 3678f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 367980677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 368088ebb749SStefano Zampini 368125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 368225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 36837dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 36847dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 368588ebb749SStefano Zampini 368680677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 368780677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 3688c58f9fdbSStefano Zampini if (!sparserhs) { 3689ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 369088ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 369106656605SStefano Zampini const PetscScalar *row_cmat_values; 369206656605SStefano Zampini const PetscInt *row_cmat_indices; 369306656605SStefano Zampini PetscInt size_of_constraint,j; 369488ebb749SStefano Zampini 369506656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 369606656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 3697ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 369806656605SStefano Zampini } 369906656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 370006656605SStefano Zampini } 3701c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&Brhs);CHKERRQ(ierr); 3702c58f9fdbSStefano Zampini } else { 3703c58f9fdbSStefano Zampini Mat tC_CR; 3704c58f9fdbSStefano Zampini 3705c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 3706c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 3707c58f9fdbSStefano Zampini PetscScalar *aa; 3708c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 3709c58f9fdbSStefano Zampini PetscBool done; 3710c58f9fdbSStefano Zampini 3711c58f9fdbSStefano Zampini ierr = MatGetRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 371213903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed"); 3713c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(C_CR,&aa);CHKERRQ(ierr); 3714c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_constraints,lda_rhs,ii,jj,aa,&tC_CR);CHKERRQ(ierr); 3715c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 371613903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed"); 3717c58f9fdbSStefano Zampini } else { 3718c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)C_CR);CHKERRQ(ierr); 3719c58f9fdbSStefano Zampini tC_CR = C_CR; 3720c58f9fdbSStefano Zampini } 3721c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tC_CR,&Brhs);CHKERRQ(ierr); 3722c58f9fdbSStefano Zampini ierr = MatDestroy(&tC_CR);CHKERRQ(ierr); 3723c58f9fdbSStefano Zampini } 3724ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 372506656605SStefano Zampini if (F) { 3726a3df083aSStefano Zampini if (need_benign_correction) { 3727df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3728a3df083aSStefano Zampini 372972b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 373072b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 3731a3df083aSStefano Zampini } 3732c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,local_auxmat2_R);CHKERRQ(ierr); 3733a3df083aSStefano Zampini if (need_benign_correction) { 3734a3df083aSStefano Zampini PetscScalar *marr; 3735df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3736a3df083aSStefano Zampini 3737a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 37385cbda25cSStefano Zampini if (lda_rhs != n_R) { 37395cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 37405cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 37415cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 37425cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 37435cbda25cSStefano Zampini } 37445cbda25cSStefano Zampini } else { 3745a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 3746a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 37475cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3748a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3749a3df083aSStefano Zampini } 37505cbda25cSStefano Zampini } 3751a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 3752a3df083aSStefano Zampini } 375306656605SStefano Zampini } else { 375480677318SStefano Zampini PetscScalar *marr; 375580677318SStefano Zampini 375680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 375706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 3758ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3759ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 376006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 376106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 376206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 376306656605SStefano Zampini } 376480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 376506656605SStefano Zampini } 3766c58f9fdbSStefano Zampini if (sparserhs) { 3767c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 3768c58f9fdbSStefano Zampini } 3769c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 377080677318SStefano Zampini if (!pcbddc->switch_static) { 377180677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 377280677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 377380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 377480677318SStefano Zampini for (i=0;i<n_constraints;i++) { 3775ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 377680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 377780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 377880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 377980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 378080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 378180677318SStefano Zampini } 378280677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 378380677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 378472b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 378580677318SStefano Zampini } else { 3786ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3787ffd830a3SStefano Zampini IS dummy; 3788ffd830a3SStefano Zampini 3789ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 37907dae84e0SHong Zhang ierr = MatCreateSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 3791ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 3792ffd830a3SStefano Zampini } else { 379380677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 379480677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 3795ffd830a3SStefano Zampini } 379625084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 379780677318SStefano Zampini } 379880677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 379980677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 380080677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 380106656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 380206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 380380677318SStefano Zampini if (isCHOL) { 380480677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 380580677318SStefano Zampini } else { 380625084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 380780677318SStefano Zampini } 380880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 380906656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 381025084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 381125084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 381225084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 381380677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 381472b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 381572b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 381606656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 381706656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 3818f4ddd8eeSStefano Zampini } 3819fc227af8SStefano Zampini 3820fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 382188ebb749SStefano Zampini if (n_vertices) { 382206656605SStefano Zampini IS is_aux; 3823c58f9fdbSStefano Zampini PetscBool isseqaij; 38243a50541eSStefano Zampini 3825b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 38266816873aSStefano Zampini IS tis; 38276816873aSStefano Zampini 38286816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 38296816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 38306816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 38316816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 38326816873aSStefano Zampini } else { 38333a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 38346816873aSStefano Zampini } 38357dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 38367dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3837c58f9fdbSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3838c58f9fdbSStefano Zampini if (!isseqaij) { /* MatMatMult(A_VR,A_RRmA_RV) below will raise an error */ 3839c58f9fdbSStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 3840c58f9fdbSStefano Zampini } 38417dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 384225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 384388ebb749SStefano Zampini } 384488ebb749SStefano Zampini 384588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 3846f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 384706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 384806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 384906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 385006656605SStefano Zampini } 3851f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 385206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 385306656605SStefano Zampini PetscScalar *marray; 385406656605SStefano Zampini 385506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 385606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 3857f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3858f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3859f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3860f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3861f4ddd8eeSStefano Zampini } 3862f4ddd8eeSStefano Zampini } 386306656605SStefano Zampini 3864f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 3865a6e023c1Sstefano_zampini PetscScalar *marr; 386688ebb749SStefano Zampini 3867a6e023c1Sstefano_zampini /* memory size */ 386806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 3869a6e023c1Sstefano_zampini if (pcbddc->switch_static || pcbddc->dbg_flag) n += n_D*pcbddc->local_primal_size; 3870a6e023c1Sstefano_zampini if (!pcbddc->symmetric_primal) n *= 2; 3871a6e023c1Sstefano_zampini ierr = PetscCalloc1(n,&marr);CHKERRQ(ierr); 3872a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3873a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 38748eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3875a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3876a6e023c1Sstefano_zampini marr += n_D*pcbddc->local_primal_size; 387788ebb749SStefano Zampini } 38783301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 3879a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3880a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 38818eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3882a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 388388ebb749SStefano Zampini } 388488ebb749SStefano Zampini } else { 3885c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 3886c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 38871b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3888c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 3889c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 3890c0553b1fSStefano Zampini } 389188ebb749SStefano Zampini } 389206656605SStefano Zampini } 3893019a44ceSStefano Zampini 389406656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 38954f1b2e48SStefano Zampini p0_lidx_I = NULL; 38964f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 3897d12edf2fSStefano Zampini const PetscInt *idxs; 3898d12edf2fSStefano Zampini 3899d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 39004f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 39014f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 39024f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 39034f1b2e48SStefano Zampini } 3904d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 3905d12edf2fSStefano Zampini } 3906d16cbb6bSStefano Zampini 390706656605SStefano Zampini /* vertices */ 390806656605SStefano Zampini if (n_vertices) { 3909c58f9fdbSStefano Zampini PetscBool restoreavr = PETSC_FALSE; 391016f15bc4SStefano Zampini 3911af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 391204708bb6SStefano Zampini 391316f15bc4SStefano Zampini if (n_R) { 391414393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 391506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 391616f15bc4SStefano Zampini PetscScalar *x,*y; 391706656605SStefano Zampini 391821eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 391914393ed6SStefano Zampini if (need_benign_correction) { 392014393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 392114393ed6SStefano Zampini IS is_p0; 392214393ed6SStefano Zampini PetscInt *idxs_p0,n; 392314393ed6SStefano Zampini 392414393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 392514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 392614393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 3927af25d912SStefano Zampini if (n != pcbddc->benign_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 392814393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 392914393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 39307dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 393114393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 393214393ed6SStefano Zampini } 393314393ed6SStefano Zampini 3934c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3935c58f9fdbSStefano Zampini if (!sparserhs || need_benign_correction) { 3936ffd830a3SStefano Zampini if (lda_rhs == n_R) { 3937af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3938ffd830a3SStefano Zampini } else { 3939ca92afb2SStefano Zampini PetscScalar *av,*array; 3940ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 3941ca92afb2SStefano Zampini PetscInt n; 3942ca92afb2SStefano Zampini PetscBool flg_row; 3943ffd830a3SStefano Zampini 3944ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 3945ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 39469d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3947ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3948ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 3949ca92afb2SStefano Zampini for (i=0;i<n;i++) { 3950ca92afb2SStefano Zampini PetscInt j; 3951ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 3952ffd830a3SStefano Zampini } 3953ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3954ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3955ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 3956ffd830a3SStefano Zampini } 3957a3df083aSStefano Zampini if (need_benign_correction) { 3958df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3959a3df083aSStefano Zampini PetscScalar *marr; 3960a3df083aSStefano Zampini 3961a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 396214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 396314393ed6SStefano Zampini 396414393ed6SStefano Zampini | 0 0 0 | (V) 396514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 396614393ed6SStefano Zampini | 0 0 -1 | (p0) 396714393ed6SStefano Zampini 396814393ed6SStefano Zampini */ 3969df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 397014393ed6SStefano Zampini const PetscScalar *vals; 397114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 397214393ed6SStefano Zampini PetscInt n,j,nz; 397314393ed6SStefano Zampini 3974df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3975df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 397614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 397714393ed6SStefano Zampini for (j=0;j<n;j++) { 397814393ed6SStefano Zampini PetscScalar val = vals[j]; 397914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 398014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 398114393ed6SStefano Zampini } 398214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3983df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 398414393ed6SStefano Zampini } 398572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 398672b8c272SStefano Zampini } 3987c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RV);CHKERRQ(ierr); 3988c58f9fdbSStefano Zampini Brhs = A_RV; 3989c58f9fdbSStefano Zampini } else { 3990c58f9fdbSStefano Zampini Mat tA_RVT,A_RVT; 3991c58f9fdbSStefano Zampini 3992c58f9fdbSStefano Zampini if (!pcbddc->symmetric_primal) { 3993c58f9fdbSStefano Zampini ierr = MatTranspose(A_RV,MAT_INITIAL_MATRIX,&A_RVT);CHKERRQ(ierr); 3994c58f9fdbSStefano Zampini } else { 3995c58f9fdbSStefano Zampini restoreavr = PETSC_TRUE; 3996c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 3997c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_VR);CHKERRQ(ierr); 3998c58f9fdbSStefano Zampini A_RVT = A_VR; 3999c58f9fdbSStefano Zampini } 4000c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 4001c58f9fdbSStefano Zampini PetscScalar *aa; 4002c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 4003c58f9fdbSStefano Zampini PetscBool done; 4004c58f9fdbSStefano Zampini 4005c58f9fdbSStefano Zampini ierr = MatGetRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 400613903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed"); 4007c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(A_RVT,&aa);CHKERRQ(ierr); 4008c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_vertices,lda_rhs,ii,jj,aa,&tA_RVT);CHKERRQ(ierr); 4009c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 401013903a91SSatish Balay if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed"); 4011c58f9fdbSStefano Zampini } else { 4012c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RVT);CHKERRQ(ierr); 4013c58f9fdbSStefano Zampini tA_RVT = A_RVT; 4014c58f9fdbSStefano Zampini } 4015c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tA_RVT,&Brhs);CHKERRQ(ierr); 4016c58f9fdbSStefano Zampini ierr = MatDestroy(&tA_RVT);CHKERRQ(ierr); 4017c58f9fdbSStefano Zampini ierr = MatDestroy(&A_RVT);CHKERRQ(ierr); 4018c58f9fdbSStefano Zampini } 401972b8c272SStefano Zampini if (F) { 402014393ed6SStefano Zampini /* need to correct the rhs */ 402172b8c272SStefano Zampini if (need_benign_correction) { 402272b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 402372b8c272SStefano Zampini PetscScalar *marr; 402472b8c272SStefano Zampini 4025c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&marr);CHKERRQ(ierr); 40265cbda25cSStefano Zampini if (lda_rhs != n_R) { 40275cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 40285cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 40295cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 40305cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 40315cbda25cSStefano Zampini } 40325cbda25cSStefano Zampini } else { 4033a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 4034a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 40355cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 4036a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4037a3df083aSStefano Zampini } 40385cbda25cSStefano Zampini } 4039c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&marr);CHKERRQ(ierr); 4040a3df083aSStefano Zampini } 4041c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,A_RRmA_RV);CHKERRQ(ierr); 4042c58f9fdbSStefano Zampini if (restoreavr) { 4043c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 4044c58f9fdbSStefano Zampini } 404514393ed6SStefano Zampini /* need to correct the solution */ 4046a3df083aSStefano Zampini if (need_benign_correction) { 4047df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4048a3df083aSStefano Zampini PetscScalar *marr; 4049a3df083aSStefano Zampini 4050a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 40515cbda25cSStefano Zampini if (lda_rhs != n_R) { 40525cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 40535cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 40545cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 40555cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 40565cbda25cSStefano Zampini } 40575cbda25cSStefano Zampini } else { 4058a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 4059a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 40605cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 4061a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4062a3df083aSStefano Zampini } 40635cbda25cSStefano Zampini } 4064a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 4065a3df083aSStefano Zampini } 406606656605SStefano Zampini } else { 4067c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&y);CHKERRQ(ierr); 406806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 4069ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 4070ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 407106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 407206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 407306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 407406656605SStefano Zampini } 4075c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&y);CHKERRQ(ierr); 407606656605SStefano Zampini } 407780677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 4078c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 4079ffd830a3SStefano Zampini /* S_VV and S_CV */ 408006656605SStefano Zampini if (n_constraints) { 408106656605SStefano Zampini Mat B; 408280677318SStefano Zampini 4083ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 408480677318SStefano Zampini for (i=0;i<n_vertices;i++) { 4085ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 4086ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 408780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 409080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 409180677318SStefano Zampini } 4092ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 409380677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 409480677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 4095ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 409680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 409706656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 4098ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 4099ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 410006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 410106656605SStefano Zampini } 4102ffd830a3SStefano Zampini if (lda_rhs != n_R) { 4103ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 4104ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 4105ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 4106ffd830a3SStefano Zampini } 410706656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 410814393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 410914393ed6SStefano Zampini if (need_benign_correction) { 4110df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 411114393ed6SStefano Zampini PetscScalar *marr,*sums; 411214393ed6SStefano Zampini 411314393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 4114f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 4115df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 411614393ed6SStefano Zampini const PetscScalar *vals; 411714393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 411814393ed6SStefano Zampini PetscInt n,j,nz; 411914393ed6SStefano Zampini 4120df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 4121df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 412214393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 412314393ed6SStefano Zampini PetscInt k; 412414393ed6SStefano Zampini sums[j] = 0.; 412514393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 412614393ed6SStefano Zampini } 412714393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 412814393ed6SStefano Zampini for (j=0;j<n;j++) { 412914393ed6SStefano Zampini PetscScalar val = vals[j]; 413014393ed6SStefano Zampini PetscInt k; 413114393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 413214393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 413314393ed6SStefano Zampini } 413414393ed6SStefano Zampini } 413514393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 4136df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 413714393ed6SStefano Zampini } 413814393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 4139f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 414014393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 414114393ed6SStefano Zampini } 414280677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 414306656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 414406656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 414506656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 414606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 414706656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 414806656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 414906656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4150d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 4151019a44ceSStefano Zampini } else { 4152d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4153d16cbb6bSStefano Zampini } 415421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 4155d16cbb6bSStefano Zampini 415606656605SStefano Zampini /* coarse basis functions */ 415706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 415816f15bc4SStefano Zampini PetscScalar *y; 415916f15bc4SStefano Zampini 4160ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 416106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 416206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 416306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 416406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 416506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 416606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 416706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 416806656605SStefano Zampini 416906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 41704f1b2e48SStefano Zampini PetscInt j; 41714f1b2e48SStefano Zampini 417206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 417306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 417406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 417506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 417606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 41774f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 417806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 417906656605SStefano Zampini } 418006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 418106656605SStefano Zampini } 418204708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 418304708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 418406656605SStefano Zampini } 41855cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 418606656605SStefano Zampini 418706656605SStefano Zampini if (n_constraints) { 418806656605SStefano Zampini Mat B; 418906656605SStefano Zampini 4190ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 419106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 419280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 419306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 419406656605SStefano Zampini if (n_vertices) { 419503dfb2d7SStefano Zampini if (isCHOL || need_benign_correction) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 419680677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 419780677318SStefano Zampini } else { 419880677318SStefano Zampini Mat S_VCt; 419980677318SStefano Zampini 4200ffd830a3SStefano Zampini if (lda_rhs != n_R) { 4201ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 420272b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 4203ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 4204ffd830a3SStefano Zampini } 420580677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 420680677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 420780677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 420880677318SStefano Zampini } 420906656605SStefano Zampini } 421006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 421106656605SStefano Zampini /* coarse basis functions */ 421206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 421306656605SStefano Zampini PetscScalar *y; 421406656605SStefano Zampini 4215ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 421606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 421706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 421806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 421906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 422006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 422106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 422206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 42234f1b2e48SStefano Zampini PetscInt j; 42244f1b2e48SStefano Zampini 422506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 422606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 422706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 422806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 422906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 42304f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 423106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 423206656605SStefano Zampini } 423306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 423406656605SStefano Zampini } 423506656605SStefano Zampini } 423680677318SStefano Zampini if (n_constraints) { 423780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 423880677318SStefano Zampini } 42394f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 424072b8c272SStefano Zampini 424172b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 424272b8c272SStefano Zampini if (pcbddc->benign_n) { 424372b8c272SStefano Zampini Mat B0_B,B0_BPHI; 424472b8c272SStefano Zampini IS is_dummy; 424572b8c272SStefano Zampini PetscScalar *data; 424672b8c272SStefano Zampini PetscInt j; 424772b8c272SStefano Zampini 424872b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 42497dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 425072b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 425172b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 425286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 425372b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 425472b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 425572b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 425672b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 425772b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 425872b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 425972b8c272SStefano Zampini } 426072b8c272SStefano Zampini } 426172b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 426272b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 426372b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 426472b8c272SStefano Zampini } 4265019a44ceSStefano Zampini 426606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 42673301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 4268ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 4269ffd830a3SStefano Zampini PetscScalar *marray; 427006656605SStefano Zampini 427106656605SStefano Zampini if (n_constraints) { 4272ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 427306656605SStefano Zampini 4274abc8f43dSstefano_zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 427506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 4276ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 427716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 427806656605SStefano Zampini if (n_vertices) { 4279ffd830a3SStefano Zampini Mat S_VCT; 428006656605SStefano Zampini 428106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 4282ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 428316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 428406656605SStefano Zampini } 4285ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 42865b782168SStefano Zampini } else { 42875b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 428806656605SStefano Zampini } 428916f15bc4SStefano Zampini if (n_vertices && n_R) { 4290ffd830a3SStefano Zampini PetscScalar *av,*marray; 4291ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 4292ffd830a3SStefano Zampini PetscInt n; 4293ffd830a3SStefano Zampini PetscBool flg_row; 429406656605SStefano Zampini 4295ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 4296af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 4297ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4298ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 4299ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4300ffd830a3SStefano Zampini for (i=0;i<n;i++) { 4301ffd830a3SStefano Zampini PetscInt j; 4302ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 4303ffd830a3SStefano Zampini } 4304ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4305ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4306ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 430706656605SStefano Zampini } 430806656605SStefano Zampini 4309ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 4310abc8f43dSstefano_zampini if (n_vertices) { 4311ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4312ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 4313ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 4314ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 431506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 431606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 431706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 431806656605SStefano Zampini } 4319ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4320abc8f43dSstefano_zampini } 43215b782168SStefano Zampini if (B_C) { 4322ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 4323ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 4324ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 4325ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 4326ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4327ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4328ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 432906656605SStefano Zampini } 4330ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 43315b782168SStefano Zampini } 433206656605SStefano Zampini /* coarse basis functions */ 433306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 433406656605SStefano Zampini PetscScalar *y; 433506656605SStefano Zampini 4336ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 433706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 433806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 433906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 434006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 434106656605SStefano Zampini if (i<n_vertices) { 434206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 434306656605SStefano Zampini } 434406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 434506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 434606656605SStefano Zampini 434706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 434806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 434906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 435006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 435106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 435206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 435306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 435406656605SStefano Zampini } 435506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 435606656605SStefano Zampini } 4357ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 4358ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 435906656605SStefano Zampini } 4360a6e023c1Sstefano_zampini 4361d62866d3SStefano Zampini /* free memory */ 436288ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 436306656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 436406656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 436506656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 436606656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 4367d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4368d62866d3SStefano Zampini if (n_vertices) { 4369d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 4370d62866d3SStefano Zampini } 4371d62866d3SStefano Zampini if (n_constraints) { 4372d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 4373d62866d3SStefano Zampini } 437488ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 437588ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 437688ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 4377d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 437888ebb749SStefano Zampini Mat coarse_sub_mat; 437925084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 438088ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 438188ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 438288ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 43838bec7fa6SStefano Zampini Mat C_B,CPHI; 43848bec7fa6SStefano Zampini IS is_dummy; 43858bec7fa6SStefano Zampini Vec mones; 438688ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 438788ebb749SStefano Zampini PetscReal real_value; 438888ebb749SStefano Zampini 4389a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 4390a3df083aSStefano Zampini Mat A; 4391a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 43927dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 43937dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 43947dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 43957dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4396a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 4397a3df083aSStefano Zampini } else { 439888ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 439988ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 440088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 440188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4402a3df083aSStefano Zampini } 440388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 440488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 4405ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 440688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 440788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 440888ebb749SStefano Zampini } 440988ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 441088ebb749SStefano Zampini 441125084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44123301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 441325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4414ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 441588ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 441688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 441788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 441888ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 441988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 442088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 442188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 442288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 442388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 442488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 442588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 442688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 442788ebb749SStefano Zampini } else { 442888ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 442988ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 443088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 443188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 443288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 443388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 443488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 443588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 443688ebb749SStefano Zampini } 443788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 443888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 443988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 4440511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 44414f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4442fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 4443d12edf2fSStefano Zampini PetscScalar *data,*data2; 44444f1b2e48SStefano Zampini PetscInt j; 4445d12edf2fSStefano Zampini 44464f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 44477dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 4448d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 444986c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 4450d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 4451d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 44524f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 44534f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 4454d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 44554f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 44564f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 44574f1b2e48SStefano Zampini } 4458d12edf2fSStefano Zampini } 4459d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 4460d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 4461d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 4462d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 4463d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 4464d12edf2fSStefano Zampini } 4465d12edf2fSStefano Zampini #if 0 4466d12edf2fSStefano Zampini { 4467d12edf2fSStefano Zampini PetscViewer viewer; 4468d12edf2fSStefano Zampini char filename[256]; 4469ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4470d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4471a7414863SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4472ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 4473ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 4474ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 4475d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 4476a7414863SStefano Zampini if (pcbddc->coarse_phi_B) { 4477ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 4478ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 447972b8c272SStefano Zampini } 4480ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 4481ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 4482ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 4483ffd830a3SStefano Zampini } 4484ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 4485ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 4486ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 4487ffd830a3SStefano Zampini } 448872b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 4489ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 4490ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 4491ffd830a3SStefano Zampini } 4492d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4493d12edf2fSStefano Zampini } 4494d12edf2fSStefano Zampini #endif 449581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 44968bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 44971575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 449806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 44998bec7fa6SStefano Zampini 45008bec7fa6SStefano Zampini /* check constraints */ 4501a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 45027dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 45034f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 45048bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4505a00504b5SStefano Zampini } else { 4506a00504b5SStefano Zampini PetscScalar *data; 4507a00504b5SStefano Zampini Mat tmat; 4508a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4509a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 4510a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4511a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4512a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4513a00504b5SStefano Zampini } 45148bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 45158bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 45168bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 45178bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4518bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 4519ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 4520bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4521bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 4522bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 4523bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4524bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 452588ebb749SStefano Zampini } 45268bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 45278bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 45288bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 45298bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 453025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 453188ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 453288ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 453388ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 453488ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 453588ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 453688ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 453788ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 453888ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 453988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 454088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 4541ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 454288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 454388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 454488ebb749SStefano Zampini } 454588ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 454688ebb749SStefano Zampini } 45478629588bSStefano Zampini /* get back data */ 45488629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 454988ebb749SStefano Zampini PetscFunctionReturn(0); 455088ebb749SStefano Zampini } 455188ebb749SStefano Zampini 45527dae84e0SHong Zhang PetscErrorCode MatCreateSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 4553aa0d41d4SStefano Zampini { 4554d65f70fdSStefano Zampini Mat *work_mat; 4555d65f70fdSStefano Zampini IS isrow_s,iscol_s; 4556d65f70fdSStefano Zampini PetscBool rsorted,csorted; 4557c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 4558aa0d41d4SStefano Zampini PetscErrorCode ierr; 4559aa0d41d4SStefano Zampini 4560aa0d41d4SStefano Zampini PetscFunctionBegin; 4561d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 4562d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 4563d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 4564d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 4565aa0d41d4SStefano Zampini 4566d65f70fdSStefano Zampini if (!rsorted) { 4567906d46d4SStefano Zampini const PetscInt *idxs; 4568906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 4569aa0d41d4SStefano Zampini 4570d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 4571d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 4572d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4573d65f70fdSStefano Zampini idxs_perm_r[i] = i; 4574aa0d41d4SStefano Zampini } 4575d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 4576d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 4577d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4578d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 4579aa0d41d4SStefano Zampini } 4580d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 4581d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 4582d65f70fdSStefano Zampini } else { 4583d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 4584d65f70fdSStefano Zampini isrow_s = isrow; 4585aa0d41d4SStefano Zampini } 4586906d46d4SStefano Zampini 4587d65f70fdSStefano Zampini if (!csorted) { 4588d65f70fdSStefano Zampini if (isrow == iscol) { 4589d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 4590d65f70fdSStefano Zampini iscol_s = isrow_s; 4591d65f70fdSStefano Zampini } else { 4592d65f70fdSStefano Zampini const PetscInt *idxs; 4593d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 4594906d46d4SStefano Zampini 4595d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 4596d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 4597d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4598d65f70fdSStefano Zampini idxs_perm_c[i] = i; 4599d65f70fdSStefano Zampini } 4600d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 4601d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 4602d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4603d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 4604d65f70fdSStefano Zampini } 4605d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 4606d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 4607d65f70fdSStefano Zampini } 4608d65f70fdSStefano Zampini } else { 4609d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 4610d65f70fdSStefano Zampini iscol_s = iscol; 4611d65f70fdSStefano Zampini } 4612d65f70fdSStefano Zampini 46137dae84e0SHong Zhang ierr = MatCreateSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4614d65f70fdSStefano Zampini 4615d65f70fdSStefano Zampini if (!rsorted || !csorted) { 4616906d46d4SStefano Zampini Mat new_mat; 4617d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 4618906d46d4SStefano Zampini 4619d65f70fdSStefano Zampini if (!rsorted) { 4620d65f70fdSStefano Zampini PetscInt *idxs_r,i; 4621d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 4622d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4623d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 4624906d46d4SStefano Zampini } 4625d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 4626d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 4627d65f70fdSStefano Zampini } else { 4628d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 4629906d46d4SStefano Zampini } 4630d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 4631d65f70fdSStefano Zampini 4632d65f70fdSStefano Zampini if (!csorted) { 4633d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 4634d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 4635d65f70fdSStefano Zampini is_perm_c = is_perm_r; 4636d65f70fdSStefano Zampini } else { 4637d65f70fdSStefano Zampini PetscInt *idxs_c,i; 4638f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 4639d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 4640d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4641d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 4642d65f70fdSStefano Zampini } 4643d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 4644d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 4645d65f70fdSStefano Zampini } 4646d65f70fdSStefano Zampini } else { 4647d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 4648d65f70fdSStefano Zampini } 4649d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 4650d65f70fdSStefano Zampini 4651d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 4652d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 4653d65f70fdSStefano Zampini work_mat[0] = new_mat; 4654d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 4655d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 4656d65f70fdSStefano Zampini } 4657d65f70fdSStefano Zampini 4658d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 4659d65f70fdSStefano Zampini *B = work_mat[0]; 4660d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 4661d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 4662d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 4663d65f70fdSStefano Zampini PetscFunctionReturn(0); 4664d65f70fdSStefano Zampini } 4665d65f70fdSStefano Zampini 46665e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 4667aa0d41d4SStefano Zampini { 4668aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 46695e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4670022d8d2bSstefano_zampini Mat new_mat,lA; 46715e8657edSStefano Zampini IS is_local,is_global; 4672d65f70fdSStefano Zampini PetscInt local_size; 4673d65f70fdSStefano Zampini PetscBool isseqaij; 4674aa0d41d4SStefano Zampini PetscErrorCode ierr; 4675aa0d41d4SStefano Zampini 4676aa0d41d4SStefano Zampini PetscFunctionBegin; 4677aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 46785e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 46795e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 4680b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 4681aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 46827dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 4683aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 4684906d46d4SStefano Zampini 4685906d46d4SStefano Zampini /* check */ 4686906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4687906d46d4SStefano Zampini Vec x,x_change; 4688906d46d4SStefano Zampini PetscReal error; 4689906d46d4SStefano Zampini 46905e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 4691906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 46925e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 4693e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4694e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4695d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 469688428137SStefano Zampini if (!pcbddc->change_interior) { 469788428137SStefano Zampini const PetscScalar *x,*y,*v; 469888428137SStefano Zampini PetscReal lerror = 0.; 469988428137SStefano Zampini PetscInt i; 470088428137SStefano Zampini 470188428137SStefano Zampini ierr = VecGetArrayRead(matis->x,&x);CHKERRQ(ierr); 470288428137SStefano Zampini ierr = VecGetArrayRead(matis->y,&y);CHKERRQ(ierr); 470388428137SStefano Zampini ierr = VecGetArrayRead(matis->counter,&v);CHKERRQ(ierr); 470488428137SStefano Zampini for (i=0;i<local_size;i++) 470588428137SStefano Zampini if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror) 470688428137SStefano Zampini lerror = PetscAbsScalar(x[i]-y[i]); 470788428137SStefano Zampini ierr = VecRestoreArrayRead(matis->x,&x);CHKERRQ(ierr); 470888428137SStefano Zampini ierr = VecRestoreArrayRead(matis->y,&y);CHKERRQ(ierr); 470988428137SStefano Zampini ierr = VecRestoreArrayRead(matis->counter,&v);CHKERRQ(ierr); 471088428137SStefano Zampini ierr = MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4711637e8532SStefano Zampini if (error > PETSC_SMALL) { 4712637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4713637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on I: %1.6e\n",error); 4714637e8532SStefano Zampini } else { 4715637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on I: %1.6e\n",error); 4716637e8532SStefano Zampini } 4717637e8532SStefano Zampini } 471888428137SStefano Zampini } 4719e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4720e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4721906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4722906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4723637e8532SStefano Zampini if (error > PETSC_SMALL) { 4724637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4725637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 4726637e8532SStefano Zampini } else { 4727637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on N: %1.6e\n",error); 4728637e8532SStefano Zampini } 4729637e8532SStefano Zampini } 4730906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4731906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4732906d46d4SStefano Zampini } 4733906d46d4SStefano Zampini 4734022d8d2bSstefano_zampini /* lA is present if we are setting up an inner BDDC for a saddle point FETI-DP */ 4735022d8d2bSstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject*)&lA);CHKERRQ(ierr); 4736022d8d2bSstefano_zampini 473722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 47389b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 473922d5777bSStefano Zampini if (isseqaij) { 4740a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4741a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 4742022d8d2bSstefano_zampini if (lA) { 4743022d8d2bSstefano_zampini Mat work; 4744022d8d2bSstefano_zampini ierr = MatPtAP(lA,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 4745022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 4746022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 4747022d8d2bSstefano_zampini } 4748aa0d41d4SStefano Zampini } else { 4749a00504b5SStefano Zampini Mat work_mat; 47501cf9b237SStefano Zampini 4751a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4752aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4753a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 47541d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 4755022d8d2bSstefano_zampini if (lA) { 4756022d8d2bSstefano_zampini Mat work; 4757022d8d2bSstefano_zampini ierr = MatConvert(lA,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4758022d8d2bSstefano_zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 4759022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 4760022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 4761022d8d2bSstefano_zampini } 4762aa0d41d4SStefano Zampini } 47633301b35fSStefano Zampini if (matis->A->symmetric_set) { 47643301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 4765e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 47663301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 4767e496cd5dSStefano Zampini #endif 47683301b35fSStefano Zampini } 4769d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 4770aa0d41d4SStefano Zampini PetscFunctionReturn(0); 4771aa0d41d4SStefano Zampini } 4772aa0d41d4SStefano Zampini 47738ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 4774a64d13efSStefano Zampini { 4775a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4776a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4777d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 477853892102SStefano Zampini PetscInt *idx_R_local=NULL; 47793a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 47803a50541eSStefano Zampini PetscInt vbs,bs; 47816816873aSStefano Zampini PetscBT bitmask=NULL; 4782a64d13efSStefano Zampini PetscErrorCode ierr; 4783a64d13efSStefano Zampini 4784a64d13efSStefano Zampini PetscFunctionBegin; 4785b23d619eSStefano Zampini /* 4786b23d619eSStefano Zampini No need to setup local scatters if 4787b23d619eSStefano Zampini - primal space is unchanged 4788b23d619eSStefano Zampini AND 4789b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 4790b23d619eSStefano Zampini AND 4791b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 4792b23d619eSStefano Zampini */ 4793b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 4794f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 4795f4ddd8eeSStefano Zampini } 4796f4ddd8eeSStefano Zampini /* destroy old objects */ 4797f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 4798f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 4799f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 4800a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 4801b371cd4fSStefano Zampini n_B = pcis->n_B; 4802b371cd4fSStefano Zampini n_D = pcis->n - n_B; 4803b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 48043a50541eSStefano Zampini 4805a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 48066816873aSStefano Zampini 480753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 4808b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4809854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 4810a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 4811a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 48120e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 4813a64d13efSStefano Zampini } 4814a64d13efSStefano Zampini 4815a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 48164641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 48176816873aSStefano Zampini idx_R_local[n_R++] = i; 4818a64d13efSStefano Zampini } 4819a64d13efSStefano Zampini } 4820df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 4821df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 48226816873aSStefano Zampini 4823df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4824df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 48256816873aSStefano Zampini } 48263a50541eSStefano Zampini 48273a50541eSStefano Zampini /* Block code */ 48283a50541eSStefano Zampini vbs = 1; 48293a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 48303a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 48313a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 48323a50541eSStefano Zampini PetscInt *vary; 4833b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4834785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 48353a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 4836d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 4837d3df7717SStefano 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 */ 48380e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 4839d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 48403a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 48413a50541eSStefano Zampini is_blocked = PETSC_FALSE; 48423a50541eSStefano Zampini break; 48433a50541eSStefano Zampini } 48443a50541eSStefano Zampini } 4845d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 4846d3df7717SStefano Zampini } else { 4847d3df7717SStefano Zampini /* Verify directly the R set */ 4848d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 4849d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 4850d3df7717SStefano Zampini for (j=1; j<bs; j++) { 4851d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 4852d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 4853d3df7717SStefano Zampini break; 4854d3df7717SStefano Zampini } 4855d3df7717SStefano Zampini } 4856d3df7717SStefano Zampini } 4857d3df7717SStefano Zampini } 48583a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 48593a50541eSStefano Zampini vbs = bs; 48603a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 48613a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 48623a50541eSStefano Zampini } 48633a50541eSStefano Zampini } 48643a50541eSStefano Zampini } 48653a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 4866b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4867df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 486853892102SStefano Zampini 4869df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4870df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 487153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 4872df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 487353892102SStefano Zampini } else { 48743a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 487553892102SStefano Zampini } 4876a64d13efSStefano Zampini 4877a64d13efSStefano Zampini /* print some info if requested */ 4878a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 4879a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4880a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 48811575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4882a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 4883a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 48844f1b2e48SStefano 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); 4885a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4886a64d13efSStefano Zampini } 4887a64d13efSStefano Zampini 4888a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 4889b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 48906816873aSStefano Zampini IS is_aux1,is_aux2; 48916816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 48926816873aSStefano Zampini 48933a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4894854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 4895854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 4896a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 48974641a718SStefano Zampini for (i=0; i<n_D; i++) { 48984641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 48994641a718SStefano Zampini } 4900a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4901a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 49024641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 49034641a718SStefano Zampini aux_array1[j++] = i; 4904a64d13efSStefano Zampini } 4905a64d13efSStefano Zampini } 4906a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4907a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4908a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 49094641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 49104641a718SStefano Zampini aux_array2[j++] = i; 4911a64d13efSStefano Zampini } 4912a64d13efSStefano Zampini } 4913a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4914a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 4915a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 4916a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4917a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 4918a64d13efSStefano Zampini 49198eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4920785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 4921a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 49224641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 49234641a718SStefano Zampini aux_array1[j++] = i; 4924a64d13efSStefano Zampini } 4925a64d13efSStefano Zampini } 4926a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4927a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 4928a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4929a64d13efSStefano Zampini } 49304641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 49313a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4932d62866d3SStefano Zampini } else { 4933df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 49346816873aSStefano Zampini IS tis; 49356816873aSStefano Zampini PetscInt schur_size; 49366816873aSStefano Zampini 4937df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 49386816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 4939df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 49406816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 49416816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 49426816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 49436816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 49446816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 4945d62866d3SStefano Zampini } 4946d62866d3SStefano Zampini } 4947a64d13efSStefano Zampini PetscFunctionReturn(0); 4948a64d13efSStefano Zampini } 4949a64d13efSStefano Zampini 4950304d26faSStefano Zampini 4951684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 4952304d26faSStefano Zampini { 4953304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4954304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4955304d26faSStefano Zampini PC pc_temp; 4956304d26faSStefano Zampini Mat A_RR; 4957f4ddd8eeSStefano Zampini MatReuse reuse; 4958304d26faSStefano Zampini PetscScalar m_one = -1.0; 4959304d26faSStefano Zampini PetscReal value; 496004708bb6SStefano Zampini PetscInt n_D,n_R; 496135529e7bSStefano Zampini PetscBool check_corr,issbaij; 4962304d26faSStefano Zampini PetscErrorCode ierr; 4963e604994aSStefano Zampini /* prefixes stuff */ 4964312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 4965e604994aSStefano Zampini size_t len; 4966304d26faSStefano Zampini 4967304d26faSStefano Zampini PetscFunctionBegin; 4968304d26faSStefano Zampini 4969e604994aSStefano Zampini /* compute prefixes */ 4970e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 4971e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 4972e604994aSStefano Zampini if (!pcbddc->current_level) { 4973e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4974e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4975e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4976e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4977e604994aSStefano Zampini } else { 497835529e7bSStefano Zampini ierr = PetscSNPrintf(str_level,sizeof(str_level),"l%d_",(int)(pcbddc->current_level));CHKERRQ(ierr); 4979e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4980e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 4981312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4982312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 498334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 498434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4985e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4986e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4987e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 4988e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 4989e604994aSStefano Zampini } 4990e604994aSStefano Zampini 4991304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 4992684f6988SStefano Zampini if (dirichlet) { 4993d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4994450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 49959a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 4996450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 4997a3df083aSStefano Zampini Mat A_IIn; 4998a3df083aSStefano Zampini 4999a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 5000a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 5001a3df083aSStefano Zampini pcis->A_II = A_IIn; 5002a3df083aSStefano Zampini } 5003450f8f5eSStefano Zampini } 50043301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 50053301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 5006964fefecSStefano Zampini } 5007ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 5008964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 5009304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 5010304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 5011304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 5012304d26faSStefano Zampini /* default */ 5013304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 5014e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 50159577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 5016304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 50179577ea80SStefano Zampini if (issbaij) { 50189577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 50199577ea80SStefano Zampini } else { 5020304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 50219577ea80SStefano Zampini } 5022304d26faSStefano Zampini /* Allow user's customization */ 5023304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 5024304d26faSStefano Zampini } 5025d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 5026b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 5027df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5028d62866d3SStefano Zampini 5029df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 5030d5574798SStefano Zampini } 5031304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 5032304d26faSStefano Zampini if (!n_D) { 5033304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 5034304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 5035304d26faSStefano Zampini } 5036304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 5037304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 5038304d26faSStefano Zampini /* set ksp_D into pcis data */ 5039304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 5040304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 5041304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 5042684f6988SStefano Zampini } 5043304d26faSStefano Zampini 5044304d26faSStefano Zampini /* NEUMANN PROBLEM */ 5045684f6988SStefano Zampini A_RR = 0; 5046684f6988SStefano Zampini if (neumann) { 5047d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 504804708bb6SStefano Zampini PetscInt ibs,mbs; 50490aa714b2SStefano Zampini PetscBool issbaij, reuse_neumann_solver; 505004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 50510aa714b2SStefano Zampini 50520aa714b2SStefano Zampini reuse_neumann_solver = PETSC_FALSE; 50530aa714b2SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 50540aa714b2SStefano Zampini IS iP; 50550aa714b2SStefano Zampini 50560aa714b2SStefano Zampini reuse_neumann_solver = PETSC_TRUE; 50570aa714b2SStefano Zampini ierr = PetscObjectQuery((PetscObject)sub_schurs->A,"__KSPFETIDP_iP",(PetscObject*)&iP);CHKERRQ(ierr); 50580aa714b2SStefano Zampini if (iP) reuse_neumann_solver = PETSC_FALSE; 50590aa714b2SStefano Zampini } 5060f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 50618ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 5062f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 5063f4ddd8eeSStefano Zampini PetscInt nn_R; 506481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 5065f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 5066f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 5067f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 5068f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 5069f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5070f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5071f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 5072727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 5073f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5074f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5075f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 5076f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 5077f4ddd8eeSStefano Zampini } 5078f4ddd8eeSStefano Zampini } 5079f4ddd8eeSStefano Zampini /* last check */ 5080d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 5081f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5082f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5083f4ddd8eeSStefano Zampini } 5084f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 5085f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 5086f4ddd8eeSStefano Zampini } 5087a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 5088af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 5089af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 509004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 509104708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 509204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 509304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 509404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 5095af732b37SStefano Zampini } else { 5096511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 50976816873aSStefano Zampini } 509804708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 509904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 510004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 510104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 510204708bb6SStefano Zampini } else { 5103511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 510404708bb6SStefano Zampini } 510504708bb6SStefano Zampini } 5106a00504b5SStefano Zampini /* extract A_RR */ 51070aa714b2SStefano Zampini if (reuse_neumann_solver) { 5108a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5109a00504b5SStefano Zampini 5110a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 511116e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5112a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 511316e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 511416e386b8SStefano Zampini } else { 51157dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 5116a00504b5SStefano Zampini } 5117a00504b5SStefano Zampini } else { 5118a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5119a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 5120a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 5121a00504b5SStefano Zampini } 5122a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 51237dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 512416e386b8SStefano Zampini } 51253301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 51263301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 51276816873aSStefano Zampini } 5128f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 5129304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 5130304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 5131304d26faSStefano Zampini /* default */ 5132304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 5133e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 5134304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 51359577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 51369577ea80SStefano Zampini if (issbaij) { 51379577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 51389577ea80SStefano Zampini } else { 5139304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 51409577ea80SStefano Zampini } 5141304d26faSStefano Zampini /* Allow user's customization */ 5142304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 5143304d26faSStefano Zampini } 5144304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 5145304d26faSStefano Zampini if (!n_R) { 5146304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 5147304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 5148304d26faSStefano Zampini } 51495cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 5150df4d28bfSStefano Zampini /* Reuse solver if it is present */ 51510aa714b2SStefano Zampini if (reuse_neumann_solver) { 5152df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5153d62866d3SStefano Zampini 5154df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 5155d62866d3SStefano Zampini } 5156304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 5157304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 5158684f6988SStefano Zampini } 5159304d26faSStefano Zampini 5160684f6988SStefano Zampini if (pcbddc->dbg_flag) { 5161684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 51621575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5163684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5164684f6988SStefano Zampini } 5165c7017625SStefano Zampini 5166c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 516735529e7bSStefano Zampini check_corr = PETSC_FALSE; 5168c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 5169c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 5170c7017625SStefano Zampini } 5171c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 517235529e7bSStefano Zampini check_corr = PETSC_TRUE; 5173c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 5174c7017625SStefano Zampini } 5175c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 517635529e7bSStefano Zampini check_corr = PETSC_TRUE; 5177c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 5178c7017625SStefano Zampini } 5179c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 5180c7017625SStefano Zampini if (pcbddc->dbg_flag) { 5181684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 51820fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 51830fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 51840fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 51850fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 51860fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 5187e604994aSStefano 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); 518835529e7bSStefano Zampini if (check_corr) { 5189c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 5190c7017625SStefano Zampini } 5191304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5192304d26faSStefano Zampini } 5193684f6988SStefano Zampini if (neumann) { /* Neumann */ 51940fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 51950fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 51960fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 51970fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 51980fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 5199e604994aSStefano 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); 520035529e7bSStefano Zampini if (check_corr) { 5201c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 5202c7017625SStefano Zampini } 5203304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5204304d26faSStefano Zampini } 5205684f6988SStefano Zampini } 52065cbda25cSStefano Zampini /* free Neumann problem's matrix */ 52075cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 5208304d26faSStefano Zampini PetscFunctionReturn(0); 5209304d26faSStefano Zampini } 5210304d26faSStefano Zampini 521180677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 5212674ae819SStefano Zampini { 5213674ae819SStefano Zampini PetscErrorCode ierr; 5214674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 5215be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5216b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 5217674ae819SStefano Zampini 5218674ae819SStefano Zampini PetscFunctionBegin; 5219b334f244SStefano Zampini if (!reuse_solver) { 522080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 522120c7b377SStefano Zampini } 522280677318SStefano Zampini if (!pcbddc->switch_static) { 522380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 522480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 522580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 522620c7b377SStefano Zampini } 5227b334f244SStefano Zampini if (!reuse_solver) { 522880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 522980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 523020c7b377SStefano Zampini } else { 5231df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5232be83ff47SStefano Zampini 5233df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5234df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 523520c7b377SStefano Zampini } 5236be83ff47SStefano Zampini } else { 523780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 523880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 523980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 524080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 524180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 524280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 524380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 524480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 524580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5246674ae819SStefano Zampini } 5247674ae819SStefano Zampini } 5248b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 524980677318SStefano Zampini if (applytranspose) { 525080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 525180677318SStefano Zampini } else { 525280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 525380677318SStefano Zampini } 5254be83ff47SStefano Zampini } else { 5255df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5256be83ff47SStefano Zampini 5257be83ff47SStefano Zampini if (applytranspose) { 5258df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 5259be83ff47SStefano Zampini } else { 5260df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 5261be83ff47SStefano Zampini } 5262be83ff47SStefano Zampini } 526380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 526480677318SStefano Zampini if (!pcbddc->switch_static) { 5265b334f244SStefano Zampini if (!reuse_solver) { 526680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 526780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5268be83ff47SStefano Zampini } else { 5269df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5270be83ff47SStefano Zampini 5271df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5272df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5273be83ff47SStefano Zampini } 527480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 527580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 527680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 527780677318SStefano Zampini } 527880677318SStefano Zampini } else { 527980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 528080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 528180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 528280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 528380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 528480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 528580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 528680677318SStefano Zampini } 528780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 528880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 528980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 529080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5291674ae819SStefano Zampini } 5292674ae819SStefano Zampini PetscFunctionReturn(0); 5293674ae819SStefano Zampini } 5294674ae819SStefano Zampini 5295dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 5296dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 5297674ae819SStefano Zampini { 5298674ae819SStefano Zampini PetscErrorCode ierr; 5299674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 5300674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 5301674ae819SStefano Zampini const PetscScalar zero = 0.0; 5302674ae819SStefano Zampini 5303674ae819SStefano Zampini PetscFunctionBegin; 5304*d4a6ed37SStefano Zampini PetscBool ss = PETSC_FALSE; 5305*d4a6ed37SStefano Zampini ierr = PetscOptionsGetBool(NULL,NULL,"-swap",&ss,NULL);CHKERRQ(ierr); 5306*d4a6ed37SStefano Zampini if (ss) { 5307*d4a6ed37SStefano Zampini Mat save_B = pcbddc->coarse_phi_B; 5308*d4a6ed37SStefano Zampini pcbddc->coarse_phi_B = pcbddc->coarse_psi_B; 5309*d4a6ed37SStefano Zampini pcbddc->coarse_psi_B = save_B; 5310*d4a6ed37SStefano Zampini Mat save_D = pcbddc->coarse_phi_D; 5311*d4a6ed37SStefano Zampini pcbddc->coarse_phi_D = pcbddc->coarse_psi_D; 5312*d4a6ed37SStefano Zampini pcbddc->coarse_psi_D = save_D; 5313*d4a6ed37SStefano Zampini } 5314dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 53154fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5316dc359a40SStefano Zampini if (applytranspose) { 5317674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 53188eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 5319dc359a40SStefano Zampini } else { 5320674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 5321674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 532215aaf578SStefano Zampini } 53234fee134fSStefano Zampini } else { 53244fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 53254fee134fSStefano Zampini } 5326efc2fbd9SStefano Zampini 5327efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 53284f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5329efc2fbd9SStefano Zampini PetscScalar *array; 53304f1b2e48SStefano Zampini PetscInt j; 5331efc2fbd9SStefano Zampini 5332efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 53334f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 5334efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5335efc2fbd9SStefano Zampini } 5336efc2fbd9SStefano Zampini 533712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 533812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 533912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 534012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 534112edc857SStefano Zampini 53429f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 534312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 534451694757SStefano Zampini Mat coarse_mat; 5345964fefecSStefano Zampini Vec rhs,sol; 534651694757SStefano Zampini MatNullSpace nullsp; 534727b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 5348964fefecSStefano Zampini 534927b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 535027b6a85dSStefano Zampini PC coarse_pc; 535127b6a85dSStefano Zampini 535227b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 535327b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 535427b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 535527b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 535627b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 535727b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 53583bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 535927b6a85dSStefano Zampini } 536027b6a85dSStefano Zampini } 5361964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 5362964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 536351694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 536451694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 536551694757SStefano Zampini if (nullsp) { 536651694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 536751694757SStefano Zampini } 536812edc857SStefano Zampini if (applytranspose) { 53699a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 5370964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 53712701bc32SStefano Zampini } else { 53721f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 53732701bc32SStefano Zampini PC coarse_pc; 53742701bc32SStefano Zampini 53752701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 53762701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 53773e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 53782701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 537912edc857SStefano Zampini } else { 5380964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 538112edc857SStefano Zampini } 53822701bc32SStefano Zampini } 53831d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 538427b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 538527b6a85dSStefano Zampini PC coarse_pc; 538627b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 538727b6a85dSStefano Zampini 538827b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 538927b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 539027b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 53913bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 539227b6a85dSStefano Zampini } 539351694757SStefano Zampini if (nullsp) { 539451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 539551694757SStefano Zampini } 539612edc857SStefano Zampini } 5397674ae819SStefano Zampini 5398674ae819SStefano Zampini /* Local solution on R nodes */ 53994fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 540080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 54019f00e9b4SStefano Zampini } 54029f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 54039f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 540412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5405674ae819SStefano Zampini 54064fee134fSStefano Zampini /* Sum contributions from the two levels */ 54074fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5408dc359a40SStefano Zampini if (applytranspose) { 5409dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5410dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5411dc359a40SStefano Zampini } else { 5412674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 54138eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5414dc359a40SStefano Zampini } 5415efc2fbd9SStefano Zampini /* store p0 */ 54164f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5417efc2fbd9SStefano Zampini PetscScalar *array; 54184f1b2e48SStefano Zampini PetscInt j; 5419efc2fbd9SStefano Zampini 5420efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 54214f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 5422efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5423efc2fbd9SStefano Zampini } 54244fee134fSStefano Zampini } else { /* expand the coarse solution */ 54254fee134fSStefano Zampini if (applytranspose) { 54264fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 54274fee134fSStefano Zampini } else { 54284fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 54294fee134fSStefano Zampini } 54304fee134fSStefano Zampini } 5431*d4a6ed37SStefano Zampini if (ss) { 5432*d4a6ed37SStefano Zampini Mat save_B = pcbddc->coarse_phi_B; 5433*d4a6ed37SStefano Zampini pcbddc->coarse_phi_B = pcbddc->coarse_psi_B; 5434*d4a6ed37SStefano Zampini pcbddc->coarse_psi_B = save_B; 5435*d4a6ed37SStefano Zampini Mat save_D = pcbddc->coarse_phi_D; 5436*d4a6ed37SStefano Zampini pcbddc->coarse_phi_D = pcbddc->coarse_psi_D; 5437*d4a6ed37SStefano Zampini pcbddc->coarse_psi_D = save_D; 5438*d4a6ed37SStefano Zampini } 5439674ae819SStefano Zampini PetscFunctionReturn(0); 5440674ae819SStefano Zampini } 5441674ae819SStefano Zampini 544212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 5443674ae819SStefano Zampini { 5444674ae819SStefano Zampini PetscErrorCode ierr; 5445674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 544658da7f69SStefano Zampini PetscScalar *array; 544712edc857SStefano Zampini Vec from,to; 5448674ae819SStefano Zampini 5449674ae819SStefano Zampini PetscFunctionBegin; 545012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 545112edc857SStefano Zampini from = pcbddc->coarse_vec; 545212edc857SStefano Zampini to = pcbddc->vec1_P; 545312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 545412edc857SStefano Zampini Vec tvec; 545558da7f69SStefano Zampini 545658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 545758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 545812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 545958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 546058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 546158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 546212edc857SStefano Zampini } 546312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 546412edc857SStefano Zampini from = pcbddc->vec1_P; 546512edc857SStefano Zampini to = pcbddc->coarse_vec; 546612edc857SStefano Zampini } 546712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 5468674ae819SStefano Zampini PetscFunctionReturn(0); 5469674ae819SStefano Zampini } 5470674ae819SStefano Zampini 547112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 5472674ae819SStefano Zampini { 5473674ae819SStefano Zampini PetscErrorCode ierr; 5474674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 547558da7f69SStefano Zampini PetscScalar *array; 547612edc857SStefano Zampini Vec from,to; 5477674ae819SStefano Zampini 5478674ae819SStefano Zampini PetscFunctionBegin; 547912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 548012edc857SStefano Zampini from = pcbddc->coarse_vec; 548112edc857SStefano Zampini to = pcbddc->vec1_P; 548212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 548312edc857SStefano Zampini from = pcbddc->vec1_P; 548412edc857SStefano Zampini to = pcbddc->coarse_vec; 548512edc857SStefano Zampini } 548612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 548712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 548812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 548912edc857SStefano Zampini Vec tvec; 549058da7f69SStefano Zampini 549112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 549258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 549358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 549458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 549558da7f69SStefano Zampini } 549658da7f69SStefano Zampini } else { 549758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 549858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 549912edc857SStefano Zampini } 550012edc857SStefano Zampini } 5501674ae819SStefano Zampini PetscFunctionReturn(0); 5502674ae819SStefano Zampini } 5503674ae819SStefano Zampini 5504984c4197SStefano Zampini /* uncomment for testing purposes */ 5505984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 5506674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 5507674ae819SStefano Zampini { 5508674ae819SStefano Zampini PetscErrorCode ierr; 5509674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 5510674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5511674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5512984c4197SStefano Zampini /* one and zero */ 5513984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 5514984c4197SStefano Zampini /* space to store constraints and their local indices */ 55159162d606SStefano Zampini PetscScalar *constraints_data; 55169162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 55179162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 55189162d606SStefano Zampini PetscInt *constraints_n; 5519984c4197SStefano Zampini /* iterators */ 5520b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 5521984c4197SStefano Zampini /* BLAS integers */ 5522e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 5523e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 5524c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 5525727cdba6SStefano Zampini /* reuse */ 55260e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 55270e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 5528984c4197SStefano Zampini /* change of basis */ 5529b3d85658SStefano Zampini PetscBool qr_needed; 55309162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 5531984c4197SStefano Zampini /* auxiliary stuff */ 553264efe560SStefano Zampini PetscInt *nnz,*is_indices; 55338a0068c3SStefano Zampini PetscInt ncc; 5534984c4197SStefano Zampini /* some quantities */ 553545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 5536a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 5537984c4197SStefano Zampini 5538674ae819SStefano Zampini PetscFunctionBegin; 55398e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 55408e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 55418e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 554216909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 5543088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 5544088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 55450e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 55460e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 55470e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 55480e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 55490e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 5550088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5551cf5a6209SStefano Zampini 5552cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 55539162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 5554cf5a6209SStefano Zampini MatNullSpace nearnullsp; 5555cf5a6209SStefano Zampini const Vec *nearnullvecs; 5556cf5a6209SStefano Zampini Vec *localnearnullsp; 5557cf5a6209SStefano Zampini PetscScalar *array; 5558cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 5559cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 5560674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 5561b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 5562674ae819SStefano Zampini PetscScalar *work; 5563674ae819SStefano Zampini PetscReal *singular_vals; 5564674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5565674ae819SStefano Zampini PetscReal *rwork; 5566674ae819SStefano Zampini #endif 5567674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5568674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 5569674ae819SStefano Zampini #else 5570964fefecSStefano Zampini PetscBLASInt dummy_int=1; 5571964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 5572674ae819SStefano Zampini #endif 5573674ae819SStefano Zampini 5574674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 5575d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 5576e4d548c7SStefano Zampini /* print some info */ 55775c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 5578e4d548c7SStefano Zampini PetscInt nv; 5579e4d548c7SStefano Zampini 5580c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5581e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 5582e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5583e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5584e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 5585e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 5586e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 5587e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5588e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5589e4d548c7SStefano Zampini } 5590e4d548c7SStefano Zampini 5591d06fc5fdSStefano Zampini /* free unneeded index sets */ 5592d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 5593d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 5594674ae819SStefano Zampini } 5595d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 5596d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5597d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5598d06fc5fdSStefano Zampini } 5599d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5600d06fc5fdSStefano Zampini n_ISForEdges = 0; 5601d06fc5fdSStefano Zampini } 5602d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 5603d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5604d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5605d06fc5fdSStefano Zampini } 5606d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5607d06fc5fdSStefano Zampini n_ISForFaces = 0; 5608d06fc5fdSStefano Zampini } 560970022509SStefano Zampini 5610674ae819SStefano Zampini /* check if near null space is attached to global mat */ 5611674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 5612674ae819SStefano Zampini if (nearnullsp) { 5613674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 5614f4ddd8eeSStefano Zampini /* remove any stored info */ 5615f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 5616f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5617f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 5618f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 5619f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 5620473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5621f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 5622f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 5623f4ddd8eeSStefano Zampini } 5624984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 5625984c4197SStefano Zampini nnsp_size = 0; 5626674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 5627674ae819SStefano Zampini } 5628984c4197SStefano Zampini /* get max number of constraints on a single cc */ 5629984c4197SStefano Zampini max_constraints = nnsp_size; 5630984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 5631984c4197SStefano Zampini 5632674ae819SStefano Zampini /* 5633674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 56349162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 56359162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 56369162d606SStefano Zampini There can be multiple constraints per connected component 5637674ae819SStefano Zampini */ 5638674ae819SStefano Zampini n_vertices = 0; 5639674ae819SStefano Zampini if (ISForVertices) { 5640674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 5641674ae819SStefano Zampini } 56429162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 56439162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 56449162d606SStefano Zampini 56459162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 56469162d606SStefano Zampini total_counts *= max_constraints; 5647674ae819SStefano Zampini total_counts += n_vertices; 56484641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 56499162d606SStefano Zampini 5650674ae819SStefano Zampini total_counts = 0; 5651674ae819SStefano Zampini max_size_of_constraint = 0; 5652674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 56539162d606SStefano Zampini IS used_is; 5654674ae819SStefano Zampini if (i<n_ISForEdges) { 56559162d606SStefano Zampini used_is = ISForEdges[i]; 5656674ae819SStefano Zampini } else { 56579162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 5658674ae819SStefano Zampini } 56599162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 5660674ae819SStefano Zampini total_counts += j; 5661674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 5662674ae819SStefano Zampini } 56639162d606SStefano 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); 56649162d606SStefano Zampini 5665984c4197SStefano Zampini /* get local part of global near null space vectors */ 5666785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 5667984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5668984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 5669e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5670e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5671984c4197SStefano Zampini } 5672674ae819SStefano Zampini 5673242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 5674242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 5675a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 5676242a89d7SStefano Zampini 5677984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 5678a773dcb8SStefano Zampini if (!skip_lapack) { 5679674ae819SStefano Zampini PetscScalar temp_work; 5680911cabfeSStefano Zampini 5681674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5682984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 5683785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 5684785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 5685785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 5686674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5687785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 5688674ae819SStefano Zampini #endif 5689674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5690c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 5691c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 5692674ae819SStefano Zampini lwork = -1; 5693674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5694674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5695c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 5696674ae819SStefano Zampini #else 5697c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 5698674ae819SStefano Zampini #endif 5699674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5700984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 5701674ae819SStefano Zampini #else /* on missing GESVD */ 5702674ae819SStefano Zampini /* SVD */ 5703674ae819SStefano Zampini PetscInt max_n,min_n; 5704674ae819SStefano Zampini max_n = max_size_of_constraint; 5705984c4197SStefano Zampini min_n = max_constraints; 5706984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 5707674ae819SStefano Zampini min_n = max_size_of_constraint; 5708984c4197SStefano Zampini max_n = max_constraints; 5709674ae819SStefano Zampini } 5710785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 5711674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5712785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 5713674ae819SStefano Zampini #endif 5714674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5715674ae819SStefano Zampini lwork = -1; 5716e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 5717e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 5718b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 5719674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5720674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 57219162d606SStefano 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)); 5722674ae819SStefano Zampini #else 57239162d606SStefano 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)); 5724674ae819SStefano Zampini #endif 5725674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5726984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 5727984c4197SStefano Zampini #endif /* on missing GESVD */ 5728674ae819SStefano Zampini /* Allocate optimal workspace */ 5729674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 5730854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 5731674ae819SStefano Zampini } 5732674ae819SStefano Zampini /* Now we can loop on constraining sets */ 5733674ae819SStefano Zampini total_counts = 0; 57349162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 57359162d606SStefano Zampini constraints_data_ptr[0] = 0; 5736674ae819SStefano Zampini /* vertices */ 57379162d606SStefano Zampini if (n_vertices) { 5738674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 57399162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 5740674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 57419162d606SStefano Zampini constraints_n[total_counts] = 1; 57429162d606SStefano Zampini constraints_data[total_counts] = 1.0; 57439162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 57449162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 5745674ae819SStefano Zampini total_counts++; 5746674ae819SStefano Zampini } 5747674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5748674ae819SStefano Zampini n_vertices = total_counts; 5749674ae819SStefano Zampini } 5750984c4197SStefano Zampini 5751674ae819SStefano Zampini /* edges and faces */ 57529162d606SStefano Zampini total_counts_cc = total_counts; 5753911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 57549162d606SStefano Zampini IS used_is; 57559162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 57569162d606SStefano Zampini 5757911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 57589162d606SStefano Zampini used_is = ISForEdges[ncc]; 5759984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 5760674ae819SStefano Zampini } else { 57619162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 5762984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 5763674ae819SStefano Zampini } 5764674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 57659162d606SStefano Zampini 57669162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 57679162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5768984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 5769984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 5770674ae819SStefano Zampini if (nnsp_has_cnst) { 57715b08dc53SStefano Zampini PetscScalar quad_value; 57729162d606SStefano Zampini 57739162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 57749162d606SStefano Zampini idxs_copied = PETSC_TRUE; 57759162d606SStefano Zampini 5776a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 5777674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 5778a773dcb8SStefano Zampini } else { 5779a773dcb8SStefano Zampini quad_value = 1.0; 5780a773dcb8SStefano Zampini } 5781674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 57829162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 5783674ae819SStefano Zampini } 57849162d606SStefano Zampini temp_constraints++; 5785674ae819SStefano Zampini total_counts++; 5786674ae819SStefano Zampini } 5787674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 5788984c4197SStefano Zampini PetscReal real_value; 57899162d606SStefano Zampini PetscScalar *ptr_to_data; 57909162d606SStefano Zampini 5791984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 57929162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 5793674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 57949162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 5795674ae819SStefano Zampini } 5796984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 5797984c4197SStefano Zampini /* check if array is null on the connected component */ 5798e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 57999162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 58005b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 5801674ae819SStefano Zampini temp_constraints++; 5802674ae819SStefano Zampini total_counts++; 58039162d606SStefano Zampini if (!idxs_copied) { 58049162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 58059162d606SStefano Zampini idxs_copied = PETSC_TRUE; 5806674ae819SStefano Zampini } 5807674ae819SStefano Zampini } 58089162d606SStefano Zampini } 58099162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 581045a1bb75SStefano Zampini valid_constraints = temp_constraints; 5811eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 5812a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 58139162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 58149162d606SStefano Zampini 58159162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5816a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 58179162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 5818a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 58199162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 5820a773dcb8SStefano Zampini } else { /* perform SVD */ 5821984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 58229162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5823674ae819SStefano Zampini 5824674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5825984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 5826984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 5827984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 5828984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 5829984c4197SStefano Zampini from that computed using LAPACKgesvd 5830984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 5831984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 5832984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 5833674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 5834e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5835984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5836674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 5837674ae819SStefano Zampini for (k=0;k<j+1;k++) { 58389162d606SStefano 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)); 5839674ae819SStefano Zampini } 5840674ae819SStefano Zampini } 5841e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 5842e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5843e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 5844674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5845c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 5846674ae819SStefano Zampini #else 5847c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 5848674ae819SStefano Zampini #endif 5849674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5850984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 5851984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 5852674ae819SStefano Zampini j = 0; 5853984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 5854674ae819SStefano Zampini total_counts = total_counts-j; 585545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 5856e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 5857c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5858c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5859c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 5860c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5861c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 5862c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5863674ae819SStefano Zampini if (j<temp_constraints) { 5864984c4197SStefano Zampini PetscInt ii; 5865984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 5866674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 58679162d606SStefano 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)); 5868674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5869984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 5870674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 58719162d606SStefano 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]; 5872674ae819SStefano Zampini } 5873674ae819SStefano Zampini } 5874674ae819SStefano Zampini } 5875674ae819SStefano Zampini #else /* on missing GESVD */ 5876e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5877e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5878b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5879674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5880674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 58819162d606SStefano 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)); 5882674ae819SStefano Zampini #else 58839162d606SStefano 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)); 5884674ae819SStefano Zampini #endif 5885984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 5886674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5887984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 5888e310c8b4SStefano Zampini k = temp_constraints; 5889e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 5890674ae819SStefano Zampini j = 0; 5891e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 589245a1bb75SStefano Zampini valid_constraints = k-j; 5893911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 5894984c4197SStefano Zampini #endif /* on missing GESVD */ 5895674ae819SStefano Zampini } 5896a773dcb8SStefano Zampini } 58979162d606SStefano Zampini /* update pointers information */ 58989162d606SStefano Zampini if (valid_constraints) { 58999162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 59009162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 59019162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 59029162d606SStefano Zampini /* set change_of_basis flag */ 590345a1bb75SStefano Zampini if (boolforchange) { 5904b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 59059162d606SStefano Zampini } 5906b3d85658SStefano Zampini total_counts_cc++; 590745a1bb75SStefano Zampini } 590845a1bb75SStefano Zampini } 5909984c4197SStefano Zampini /* free workspace */ 59108f1c130eSStefano Zampini if (!skip_lapack) { 5911984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 5912984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5913984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 5914984c4197SStefano Zampini #endif 5915984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 5916984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5917984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 5918984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 5919984c4197SStefano Zampini #endif 5920984c4197SStefano Zampini } 5921984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5922984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 5923984c4197SStefano Zampini } 5924984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 5925cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 5926cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5927cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5928cf5a6209SStefano Zampini } 5929cf5a6209SStefano Zampini if (n_ISForFaces) { 5930cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5931cf5a6209SStefano Zampini } 5932cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5933cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5934cf5a6209SStefano Zampini } 5935cf5a6209SStefano Zampini if (n_ISForEdges) { 5936cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5937cf5a6209SStefano Zampini } 5938cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 593908122e43SStefano Zampini } else { 594008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5941984c4197SStefano Zampini 594208122e43SStefano Zampini total_counts = 0; 594308122e43SStefano Zampini n_vertices = 0; 5944d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 5945d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 594608122e43SStefano Zampini } 594708122e43SStefano Zampini max_constraints = 0; 59489162d606SStefano Zampini total_counts_cc = 0; 594908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 595008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 59519162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 595208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 595308122e43SStefano Zampini } 59549162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 59559162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 59569162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 59579162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 595874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 59599162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 59609162d606SStefano Zampini total_counts_cc = 0; 59619162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 59629162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 59639162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 596408122e43SStefano Zampini } 596508122e43SStefano Zampini } 59669162d606SStefano Zampini #if 0 59679162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 59689162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 59699162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 59709162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 59719162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 59729162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 59739162d606SStefano Zampini } 59749162d606SStefano Zampini printf("\n"); 59759162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 59769162d606SStefano Zampini } 59771b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 59788bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 59791b968477SStefano Zampini } 59801b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 59818bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] sub %d, edge %d, n %d\n",PetscGlobalRank,i,(PetscBool)PetscBTLookup(sub_schurs->is_edge,i),pcbddc->adaptive_constraints_n[i+n_vertices]); 59821b968477SStefano Zampini } 598308122e43SStefano Zampini #endif 598408122e43SStefano Zampini 59858bec7fa6SStefano Zampini max_size_of_constraint = 0; 59869162d606SStefano 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]); 59879162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 598808122e43SStefano Zampini /* Change of basis */ 5989b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 599008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 599108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 599208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 5993b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 599408122e43SStefano Zampini } 599508122e43SStefano Zampini } 599608122e43SStefano Zampini } 599708122e43SStefano Zampini } 5998984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 59994f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 600008122e43SStefano Zampini 60019162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 60029162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 6003eee23b56SStefano 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\n",constraints_idxs_ptr[total_counts_cc],i); 6004674ae819SStefano Zampini 6005674ae819SStefano Zampini /* Create constraint matrix */ 6006674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 600716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 6008984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 6009984c4197SStefano Zampini 6010984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 6011a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 6012a717540cSStefano Zampini qr_needed = PETSC_FALSE; 601374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 6014984c4197SStefano Zampini total_primal_vertices=0; 6015b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 60169162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 60179162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 601872b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 60199162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 6020b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 602164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 60229162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 60239162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 6024a717540cSStefano Zampini } 6025b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 602691af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 6027a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 6028a717540cSStefano Zampini qr_needed = PETSC_TRUE; 6029a717540cSStefano Zampini } 6030fa434743SStefano Zampini } else { 6031b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 6032fa434743SStefano Zampini } 6033a717540cSStefano Zampini } 6034b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 6035b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 6036674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 603770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 60384f1b2e48SStefano 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); 60390e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 60400e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 6041984c4197SStefano Zampini 6042984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 604374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 6044785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 6045984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 604674d5cdf7SStefano Zampini 6047984c4197SStefano Zampini j = total_primal_vertices; 604874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 6049b3d85658SStefano Zampini cum = total_primal_vertices; 60509162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 60514641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 6052b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 6053b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 6054b3d85658SStefano Zampini cum++; 60559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 605674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 605774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 605874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 605974d5cdf7SStefano Zampini } 60609162d606SStefano Zampini j += constraints_n[i]; 6061674ae819SStefano Zampini } 6062674ae819SStefano Zampini } 6063674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 6064674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 6065088faed8SStefano Zampini 6066674ae819SStefano Zampini /* set values in constraint matrix */ 6067984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 60680e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6069674ae819SStefano Zampini } 6070984c4197SStefano Zampini total_counts = total_primal_vertices; 60719162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 60724641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 60739162d606SStefano Zampini PetscInt *cols; 60749162d606SStefano Zampini 60759162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 60769162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 60779162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 60789162d606SStefano Zampini PetscInt row = total_counts+k; 60799162d606SStefano Zampini PetscScalar *vals; 60809162d606SStefano Zampini 60819162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 60829162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 60839162d606SStefano Zampini } 60849162d606SStefano Zampini total_counts += constraints_n[i]; 6085674ae819SStefano Zampini } 6086674ae819SStefano Zampini } 6087674ae819SStefano Zampini /* assembling */ 6088674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6089674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6090088faed8SStefano Zampini 6091984c4197SStefano Zampini /* 60926a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 6093984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 6094f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 6095984c4197SStefano Zampini */ 6096674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 6097674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 6098026de310SStefano Zampini /* dual and primal dofs on a single cc */ 6099984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 6100984c4197SStefano Zampini /* working stuff for GEQRF */ 610181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 6102984c4197SStefano Zampini PetscBLASInt lqr_work; 6103984c4197SStefano Zampini /* working stuff for UNGQR */ 6104984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 6105984c4197SStefano Zampini PetscBLASInt lgqr_work; 6106984c4197SStefano Zampini /* working stuff for TRTRS */ 6107984c4197SStefano Zampini PetscScalar *trs_rhs; 61083f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 6109984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 6110984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 6111984c4197SStefano Zampini PetscScalar *start_vals; 6112984c4197SStefano Zampini /* working stuff for values insertion */ 61134641a718SStefano Zampini PetscBT is_primal; 611464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 6115906d46d4SStefano Zampini /* matrix sizes */ 6116906d46d4SStefano Zampini PetscInt global_size,local_size; 6117906d46d4SStefano Zampini /* temporary change of basis */ 6118906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 6119cf5a6209SStefano Zampini /* extra space for debugging */ 6120cf5a6209SStefano Zampini PetscScalar *dbg_work; 6121984c4197SStefano Zampini 6122906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 6123906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 612416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 6125bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 6126906d46d4SStefano Zampini /* nonzeros for local mat */ 6127bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 61281dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 6129bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 61301dd7afcfSStefano Zampini } else { 61311dd7afcfSStefano Zampini const PetscInt *ii; 61321dd7afcfSStefano Zampini PetscInt n; 61331dd7afcfSStefano Zampini PetscBool flg_row; 61341dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 61351dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 61361dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 61371dd7afcfSStefano Zampini } 61389162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 6139a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 61409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 6141a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 61429162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 6143a717540cSStefano Zampini } else { 61449162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 61459162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 6146a717540cSStefano Zampini } 6147a717540cSStefano Zampini } 6148a717540cSStefano Zampini } 6149906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 6150bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 61511dd7afcfSStefano Zampini /* Set interior change in the matrix */ 61521dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 6153bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 6154906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 6155a717540cSStefano Zampini } 61561dd7afcfSStefano Zampini } else { 61571dd7afcfSStefano Zampini const PetscInt *ii,*jj; 61581dd7afcfSStefano Zampini PetscScalar *aa; 61591dd7afcfSStefano Zampini PetscInt n; 61601dd7afcfSStefano Zampini PetscBool flg_row; 61611dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 61621dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 61631dd7afcfSStefano Zampini for (i=0;i<n;i++) { 61641dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 61651dd7afcfSStefano Zampini } 61661dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 61671dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 61681dd7afcfSStefano Zampini } 6169a717540cSStefano Zampini 6170a717540cSStefano Zampini if (pcbddc->dbg_flag) { 6171a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6172a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6173a717540cSStefano Zampini } 6174a717540cSStefano Zampini 6175a717540cSStefano Zampini 6176a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 6177a717540cSStefano Zampini /* 6178a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 6179a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 6180a717540cSStefano Zampini 6181a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 6182a717540cSStefano Zampini 6183a6b551f4SStefano 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) 6184a6b551f4SStefano Zampini 6185a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 6186a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 6187a717540cSStefano Zampini | ... | 6188a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 6189a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 6190a717540cSStefano Zampini 6191a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 6192a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 6193a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 6194a6b551f4SStefano Zampini 6195a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 6196a717540cSStefano Zampini */ 6197a717540cSStefano Zampini if (qr_needed) { 6198984c4197SStefano Zampini /* space to store Q */ 6199854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 62004e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 62014e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 6202984c4197SStefano Zampini /* first we issue queries for optimal work */ 62033f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 62043f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 62053f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6206984c4197SStefano Zampini lqr_work = -1; 62073f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 6208984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 6209984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 6210785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 6211984c4197SStefano Zampini lgqr_work = -1; 62123f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 62133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 62143f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 62153f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 62163f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 6217c964aadfSJose E. Roman PetscStackCallBLAS("LAPACKorgqr",LAPACKorgqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 6218c964aadfSJose E. Roman if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to ORGQR/UNGQR Lapack routine %d",(int)lierr); 6219984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 6220785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 6221984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 6222785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 6223a717540cSStefano Zampini /* allocating workspace for check */ 6224a717540cSStefano Zampini if (pcbddc->dbg_flag) { 6225cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 6226a717540cSStefano Zampini } 6227a717540cSStefano Zampini } 6228984c4197SStefano Zampini /* array to store whether a node is primal or not */ 62294641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 6230473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 62310e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 6232eee23b56SStefano Zampini if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D\n",total_primal_vertices,i); 623339e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 623439e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 623539e2fb2aSStefano Zampini } 623639e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 6237984c4197SStefano Zampini 6238a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 62399162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 62409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 62414641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 6242984c4197SStefano Zampini /* get constraint info */ 62439162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 6244984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 6245984c4197SStefano Zampini 6246984c4197SStefano Zampini if (pcbddc->dbg_flag) { 62479162d606SStefano 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); 6248674ae819SStefano Zampini } 6249984c4197SStefano Zampini 6250fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 6251a717540cSStefano Zampini 6252a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 6253a717540cSStefano Zampini if (pcbddc->dbg_flag) { 62549162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 6255a717540cSStefano Zampini } 6256984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 62579162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 6258984c4197SStefano Zampini 6259984c4197SStefano Zampini /* compute QR decomposition of constraints */ 62603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 62613f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 62623f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6263674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 62643f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 6265984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 6266674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6267984c4197SStefano Zampini 6268984c4197SStefano Zampini /* explictly compute R^-T */ 6269984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 6270984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 62713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 62723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 62733f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 62743f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 6275984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 62763f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 6277984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 6278984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6279984c4197SStefano Zampini 6280a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 62813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 62823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 62833f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 62843f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6285984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6286c964aadfSJose E. Roman PetscStackCallBLAS("LAPACKorgqr",LAPACKorgqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 6287c964aadfSJose E. Roman if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in ORGQR/UNGQR Lapack routine %d",(int)lierr); 6288984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6289984c4197SStefano Zampini 6290984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 6291984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 6292984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 62933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 62943f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 62953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 62963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 62973f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 62983f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 6299984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 63009162d606SStefano 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)); 6301984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 63029162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 6303984c4197SStefano Zampini 6304984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 63059162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 6306984c4197SStefano Zampini /* insert cols for primal dofs */ 6307984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 6308984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 63099162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6310906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6311984c4197SStefano Zampini } 6312984c4197SStefano Zampini /* insert cols for dual dofs */ 6313984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 63149162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 6315984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 63169162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6317906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6318984c4197SStefano Zampini j++; 6319674ae819SStefano Zampini } 6320674ae819SStefano Zampini } 6321984c4197SStefano Zampini 6322984c4197SStefano Zampini /* check change of basis */ 6323984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6324984c4197SStefano Zampini PetscInt ii,jj; 6325984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 6326c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 6327c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 6328c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 6329c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6330c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 6331c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 6332984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6333cf5a6209SStefano 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)); 6334984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6335984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6336984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6337cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 6338c068d9bbSLisandro 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; 6339674ae819SStefano Zampini } 6340674ae819SStefano Zampini } 6341984c4197SStefano Zampini if (!valid_qr) { 634222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 6343984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6344984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6345cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 6346cf5a6209SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not orthogonal to constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii])); 6347674ae819SStefano Zampini } 6348c068d9bbSLisandro Dalcin if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-(PetscReal)1) > 1.e-12) { 6349cf5a6209SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %d is not unitary w.r.t constraint %d (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii])); 6350984c4197SStefano Zampini } 6351984c4197SStefano Zampini } 6352984c4197SStefano Zampini } 6353674ae819SStefano Zampini } else { 635422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 6355674ae819SStefano Zampini } 6356674ae819SStefano Zampini } 6357a717540cSStefano Zampini } else { /* simple transformation block */ 6358a717540cSStefano Zampini PetscInt row,col; 6359a6b551f4SStefano Zampini PetscScalar val,norm; 6360a6b551f4SStefano Zampini 6361a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 63629162d606SStefano 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)); 6363a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 63649162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 63659162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6366bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 63679162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 6368906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 63699162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 6370a717540cSStefano Zampini } else { 6371a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 63729162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6373a717540cSStefano Zampini if (row != col) { 63749162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 6375a717540cSStefano Zampini } else { 63769162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 6377a717540cSStefano Zampini } 6378906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 6379a717540cSStefano Zampini } 6380a717540cSStefano Zampini } 6381a717540cSStefano Zampini } 638298a51de6SStefano Zampini if (pcbddc->dbg_flag) { 638322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 6384a717540cSStefano Zampini } 6385674ae819SStefano Zampini } 6386984c4197SStefano Zampini } else { 6387984c4197SStefano Zampini if (pcbddc->dbg_flag) { 63889162d606SStefano 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); 6389674ae819SStefano Zampini } 6390674ae819SStefano Zampini } 6391674ae819SStefano Zampini } 6392a717540cSStefano Zampini 6393a717540cSStefano Zampini /* free workspace */ 6394a717540cSStefano Zampini if (qr_needed) { 6395984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6396cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 6397984c4197SStefano Zampini } 6398984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 6399984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 6400984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 6401984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 6402984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 6403674ae819SStefano Zampini } 6404a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 6405906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6406906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6407906d46d4SStefano Zampini 6408906d46d4SStefano Zampini /* assembling of global change of variable */ 640988c03ad3SStefano Zampini if (!pcbddc->fake_change) { 6410bbb9e6c6SStefano Zampini Mat tmat; 641116f15bc4SStefano Zampini PetscInt bs; 641216f15bc4SStefano Zampini 6413906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 6414906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 6415bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 6416bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 6417bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6418bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 641916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 642016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 6421906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 6422bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 6423bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6424bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6425bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6426bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6427e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6428e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6429bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 6430bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 643188c03ad3SStefano Zampini 6432906d46d4SStefano Zampini /* check */ 6433906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 6434906d46d4SStefano Zampini PetscReal error; 6435906d46d4SStefano Zampini Vec x,x_change; 6436906d46d4SStefano Zampini 6437906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 6438906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 6439906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 6440906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 6441e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6442e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6443bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 6444e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6445e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6446906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 6447906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 6448906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 6449637e8532SStefano Zampini if (error > PETSC_SMALL) { 6450637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 6451637e8532SStefano Zampini } 6452906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 6453906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 6454906d46d4SStefano Zampini } 6455b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 6456b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 6457b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6458bf3a8328SStefano Zampini 645913903a91SSatish 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"); 6460b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 6461ac632422SStefano Zampini Mat S_new,tmat; 6462bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 6463bbb9e6c6SStefano Zampini 6464bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 64657dae84e0SHong Zhang ierr = MatCreateSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 6466bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6467bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 6468bf3a8328SStefano Zampini IS is_V; 6469b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 6470b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 6471b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 6472b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 6473b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 6474bf3a8328SStefano Zampini } 6475bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 6476ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6477b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 6478ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6479bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6480bf3a8328SStefano Zampini const PetscScalar *array; 6481bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 6482bf3a8328SStefano Zampini PetscInt i,n_V; 6483bf3a8328SStefano Zampini 6484b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6485b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 6486b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6487b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6488b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 6489b087196eSStefano Zampini for (i=0;i<n_V;i++) { 6490b087196eSStefano Zampini PetscScalar val; 6491b087196eSStefano Zampini PetscInt idx; 6492b087196eSStefano Zampini 6493b087196eSStefano Zampini idx = idxs_V[i]; 6494b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 6495b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 6496b087196eSStefano Zampini } 6497b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6498b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6499bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 6500bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6501bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6502bf3a8328SStefano Zampini } 6503ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 6504ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6505ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 6506ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6507b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 6508ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6509bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6510b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6511bf3a8328SStefano Zampini } 6512ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 6513ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6514ac632422SStefano Zampini } 6515b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 651688c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6517b96c3477SStefano Zampini } 6518c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 6519b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 6520c9db6a07SStefano Zampini PetscInt i; 6521c9db6a07SStefano Zampini 6522c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6523c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 6524c9db6a07SStefano Zampini } 6525c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 6526c9db6a07SStefano Zampini } 6527b96c3477SStefano Zampini } 652816909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 652916909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 653016909a7fSStefano Zampini } else { 6531906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 653216909a7fSStefano Zampini } 65331dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 653427b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 653572b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 653672b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 653772b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 653872b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 653972b8c272SStefano Zampini } 65401dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 654127b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 6542b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 6543b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 6544906d46d4SStefano Zampini } else { 65451dd7afcfSStefano Zampini Mat benign_global = NULL; 654627b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 65471dd7afcfSStefano Zampini Mat tmat; 65481dd7afcfSStefano Zampini 65491dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 65501dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 65511dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 65521dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 65531dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 65541dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 65551dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 65561dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 65571dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 65581dd7afcfSStefano Zampini if (pcbddc->benign_change) { 65591dd7afcfSStefano Zampini Mat M; 65601dd7afcfSStefano Zampini 65611dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 65621dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 65631dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 65641dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 6565906d46d4SStefano Zampini } else { 65661dd7afcfSStefano Zampini Mat eye; 65671dd7afcfSStefano Zampini PetscScalar *array; 65681dd7afcfSStefano Zampini 65691dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 65701dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 65711dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 65721dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 6573906d46d4SStefano Zampini } 65741dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 65751dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 65761dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 65771dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 65781dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 65791dd7afcfSStefano Zampini } 65801dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 65811dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 65821dd7afcfSStefano Zampini } 65831dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 65841dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 65851dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 658627b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 65871dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 65881dd7afcfSStefano Zampini } 65891dd7afcfSStefano Zampini } 659016909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 659116909a7fSStefano Zampini IS is_global; 659216909a7fSStefano Zampini const PetscInt *gidxs; 659316909a7fSStefano Zampini 659416909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 659516909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 659616909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 65977dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 659816909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 659916909a7fSStefano Zampini } 66001dd7afcfSStefano Zampini } 66011dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 66021dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 6603b9b85e73SStefano Zampini } 6604a717540cSStefano Zampini 660572b8c272SStefano Zampini if (!pcbddc->fake_change) { 66064f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 66074f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 66084f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 66094f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 6610019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 6611019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 6612019a44ceSStefano Zampini pcbddc->local_primal_size++; 6613019a44ceSStefano Zampini } 6614019a44ceSStefano Zampini 6615019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 6616727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 6617727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 66189f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 6619c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 66200e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 66219f47a83aSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 6622727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 6623727cdba6SStefano Zampini } 66240e6343abSStefano Zampini } 6625727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 6626b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 662772b8c272SStefano Zampini } 662872b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 6629727cdba6SStefano Zampini 6630a717540cSStefano Zampini /* flush dbg viewer */ 6631b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 6632b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6633b8ffe317SStefano Zampini } 6634a717540cSStefano Zampini 6635e310c8b4SStefano Zampini /* free workspace */ 6636a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 66374641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 663808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 66399162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 66409162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 664108122e43SStefano Zampini } else { 66429162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 66439162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 66449162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 664508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 664608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 66479162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 66489162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 664908122e43SStefano Zampini } 6650674ae819SStefano Zampini PetscFunctionReturn(0); 6651674ae819SStefano Zampini } 6652674ae819SStefano Zampini 6653674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 6654674ae819SStefano Zampini { 665571582508SStefano Zampini ISLocalToGlobalMapping map; 6656674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 6657674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 665866da6bd7Sstefano_zampini PetscInt i,N; 665966da6bd7Sstefano_zampini PetscBool rcsr = PETSC_FALSE; 666066da6bd7Sstefano_zampini PetscErrorCode ierr; 6661674ae819SStefano Zampini 6662674ae819SStefano Zampini PetscFunctionBegin; 66638af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 6664b03ebc13SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 66658e61c736SStefano Zampini /* Reset previously computed graph */ 66668e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 6667674ae819SStefano Zampini /* Init local Graph struct */ 66687fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 666971582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 6670be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 6671674ae819SStefano Zampini 66727a0e7b2cSstefano_zampini if (pcbddc->user_primal_vertices_local && !pcbddc->user_primal_vertices) { 66737a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 66747a0e7b2cSstefano_zampini } 6675575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 66769a962809SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %d, expected %d\n",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs); 66779577ea80SStefano Zampini 6678674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 667966da6bd7Sstefano_zampini if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) { 66804d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 66814d379d7bSStefano Zampini PetscInt nvtxs; 6682e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 6683674ae819SStefano Zampini 66842fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 66852fffb893SStefano Zampini if (flg_row) { 66864d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 6687b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 66882fffb893SStefano Zampini } 66892fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 669066da6bd7Sstefano_zampini rcsr = PETSC_TRUE; 6691674ae819SStefano Zampini } 66929b28b941SStefano Zampini if (pcbddc->dbg_flag) { 66939b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6694674ae819SStefano Zampini } 6695674ae819SStefano Zampini 6696674ae819SStefano Zampini /* Setup of Graph */ 66974b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 669814f95afaSStefano 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); 6699674ae819SStefano Zampini 67004f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 67014f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 67024f1b2e48SStefano Zampini PetscInt *local_subs; 67034f1b2e48SStefano Zampini 67044f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 67054f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 67064f1b2e48SStefano Zampini const PetscInt *idxs; 67074f1b2e48SStefano Zampini PetscInt nl,j; 67084f1b2e48SStefano Zampini 67094f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 67104f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 671171582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 67124f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 67134f1b2e48SStefano Zampini } 67144f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 67154f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 67164f1b2e48SStefano Zampini } 67178af8fcf9SStefano Zampini } 67184f1b2e48SStefano Zampini 6719cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 6720674ae819SStefano Zampini /* Graph's connected components analysis */ 6721674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 672271582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 67238af8fcf9SStefano Zampini } 672466da6bd7Sstefano_zampini if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0; 6725674ae819SStefano Zampini PetscFunctionReturn(0); 6726674ae819SStefano Zampini } 6727674ae819SStefano Zampini 67289a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 67299a7d3425SStefano Zampini { 67309a7d3425SStefano Zampini PetscInt i,j; 67319a7d3425SStefano Zampini PetscScalar *alphas; 67329a7d3425SStefano Zampini PetscErrorCode ierr; 67339a7d3425SStefano Zampini 67349a7d3425SStefano Zampini PetscFunctionBegin; 6735785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 67369a7d3425SStefano Zampini for (i=0;i<n;i++) { 67379a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 6738669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 6739669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 6740669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 67419a7d3425SStefano Zampini } 67429a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 67439a7d3425SStefano Zampini PetscFunctionReturn(0); 67449a7d3425SStefano Zampini } 67459a7d3425SStefano Zampini 6746bb360cb4SStefano Zampini PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 6747e7931f94SStefano Zampini { 674857de7509SStefano Zampini Mat A; 6749e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 6750e7931f94SStefano Zampini PetscMPIInt size,rank,color; 675152e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 675252e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 6753bb360cb4SStefano Zampini PetscInt im_active,active_procs,N,n,i,j,threshold = 2; 675457de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 675527b6a85dSStefano Zampini PetscInt xadj_count,*count; 675627b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 675727b6a85dSStefano Zampini PetscSubcomm psubcomm; 675827b6a85dSStefano Zampini MPI_Comm subcomm; 675952e5ac9dSStefano Zampini PetscErrorCode ierr; 6760a57a6d2fSStefano Zampini 6761e7931f94SStefano Zampini PetscFunctionBegin; 676257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 676357de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 6764fbfcfee5SBarry 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); 676557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 676657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 676757de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 676857de7509SStefano Zampini 676957de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 677057de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 677157de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 677257de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 677357de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 6774bb360cb4SStefano Zampini im_active = !!n; 677557de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 677657de7509SStefano Zampini void_procs = size - active_procs; 677757de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 677857de7509SStefano Zampini if (void_procs) { 677957de7509SStefano Zampini PetscInt ncand; 678057de7509SStefano Zampini 678157de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 678257de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 678357de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 678457de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 678557de7509SStefano Zampini if (!procs_candidates[i]) { 678657de7509SStefano Zampini procs_candidates[ncand++] = i; 678757de7509SStefano Zampini } 678857de7509SStefano Zampini } 678957de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 679057de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 679157de7509SStefano Zampini } 679257de7509SStefano Zampini 6793bb360cb4SStefano Zampini /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix 679414f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 6795bb360cb4SStefano Zampini ierr = MatGetSize(mat,&N,NULL);CHKERRQ(ierr); 6796bb360cb4SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) { 679714f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 679814f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 679914f0bfb9SStefano Zampini else dest = rank; 680057de7509SStefano Zampini if (im_active) { 680157de7509SStefano Zampini issize = 1; 680257de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 680314f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 680457de7509SStefano Zampini } else { 680514f0bfb9SStefano Zampini isidx = dest; 680657de7509SStefano Zampini } 680757de7509SStefano Zampini } else { 680857de7509SStefano Zampini issize = 0; 680957de7509SStefano Zampini isidx = -1; 681057de7509SStefano Zampini } 6811bb360cb4SStefano Zampini if (*n_subdomains != 1) *n_subdomains = active_procs; 681257de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 6813daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 681457de7509SStefano Zampini PetscFunctionReturn(0); 681557de7509SStefano Zampini } 6816c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 6817c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 681827b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 6819e7931f94SStefano Zampini 6820e7931f94SStefano Zampini /* Get info on mapping */ 68213bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6822e7931f94SStefano Zampini 6823e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 6824785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 6825e7931f94SStefano Zampini xadj[0] = 0; 6826e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 6827785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 6828785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 6829bb360cb4SStefano Zampini ierr = PetscCalloc1(n,&count);CHKERRQ(ierr); 683027b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 683127b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 683227b6a85dSStefano Zampini count[shared[i][j]] += 1; 6833e7931f94SStefano Zampini 683427b6a85dSStefano Zampini xadj_count = 0; 68352b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 683627b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 683727b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 6838d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 6839d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 6840d023bfaeSStefano Zampini xadj_count++; 684127b6a85dSStefano Zampini break; 684227b6a85dSStefano Zampini } 6843e7931f94SStefano Zampini } 6844e7931f94SStefano Zampini } 6845d023bfaeSStefano Zampini xadj[1] = xadj_count; 684627b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 68473bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6848e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 6849e7931f94SStefano Zampini 68503837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 6851e7931f94SStefano Zampini 685227b6a85dSStefano Zampini /* Restrict work on active processes only */ 685327b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 685427b6a85dSStefano Zampini if (void_procs) { 685527b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 685627b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 685727b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 685827b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 685927b6a85dSStefano Zampini } else { 686027b6a85dSStefano Zampini psubcomm = NULL; 686127b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 686227b6a85dSStefano Zampini } 686327b6a85dSStefano Zampini 686427b6a85dSStefano Zampini v_wgt = NULL; 686527b6a85dSStefano Zampini if (!color) { 6866e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 6867e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 6868e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6869c8587f34SStefano Zampini } else { 687052e5ac9dSStefano Zampini Mat subdomain_adj; 687152e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 687252e5ac9dSStefano Zampini MatPartitioning partitioner; 687327b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 687452e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 687557de7509SStefano Zampini PetscMPIInt size; 6876b0c7d250SStefano Zampini PetscBool aggregate; 6877b0c7d250SStefano Zampini 687827b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 687927b6a85dSStefano Zampini if (void_procs) { 688027b6a85dSStefano Zampini PetscInt prank = rank; 6881785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 688227b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 6883e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 6884e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 6885c8587f34SStefano Zampini } 6886e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 688727b6a85dSStefano Zampini } else { 688827b6a85dSStefano Zampini oldranks = NULL; 688927b6a85dSStefano Zampini } 6890b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 689127b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 6892b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 6893b0c7d250SStefano Zampini PetscMPIInt nrank; 6894b0c7d250SStefano Zampini PetscScalar *vals; 6895b0c7d250SStefano Zampini 689627b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 6897b0c7d250SStefano Zampini lrows = 0; 6898b0c7d250SStefano Zampini if (nrank<redprocs) { 6899b0c7d250SStefano Zampini lrows = size/redprocs; 6900b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 6901b0c7d250SStefano Zampini } 690227b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 6903b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 6904b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6905b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6906b0c7d250SStefano Zampini row = nrank; 6907b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 6908b0c7d250SStefano Zampini cols = adjncy; 6909b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 6910b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 6911b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 6912b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6913b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 691452e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 691552e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 691652e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6917b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 691827b6a85dSStefano Zampini if (use_vwgt) { 691927b6a85dSStefano Zampini Vec v; 692027b6a85dSStefano Zampini const PetscScalar *array; 692127b6a85dSStefano Zampini PetscInt nl; 692227b6a85dSStefano Zampini 692327b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 6924bb360cb4SStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)n,INSERT_VALUES);CHKERRQ(ierr); 692527b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 692627b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 692727b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 692827b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 692927b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 693022db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 693127b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 693227b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 693327b6a85dSStefano Zampini } 6934b0c7d250SStefano Zampini } else { 693527b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 693627b6a85dSStefano Zampini if (use_vwgt) { 693727b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 6938bb360cb4SStefano Zampini v_wgt[0] = n; 693927b6a85dSStefano Zampini } 6940b0c7d250SStefano Zampini } 694122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 6942e7931f94SStefano Zampini 6943e7931f94SStefano Zampini /* Partition */ 694427b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 6945e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 694627b6a85dSStefano Zampini if (v_wgt) { 6947e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 6948c8587f34SStefano Zampini } 694957de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 695057de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 6951e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 6952e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 695322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 6954e7931f94SStefano Zampini 695552e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 69566583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 695752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 695852e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 695957de7509SStefano Zampini if (!aggregate) { 696057de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 696127b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 696227b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 696327b6a85dSStefano Zampini #endif 696457de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 696527b6a85dSStefano Zampini } else if (oldranks) { 6966b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 696727b6a85dSStefano Zampini } else { 696827b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 696957de7509SStefano Zampini } 697028143c3dSStefano Zampini } else { 69717fb8a5e4SKarl Rupp PetscInt idx = 0; 6972b0c7d250SStefano Zampini PetscMPIInt tag; 6973b0c7d250SStefano Zampini MPI_Request *reqs; 6974b0c7d250SStefano Zampini 6975b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 6976b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 6977b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 697827b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 697928143c3dSStefano Zampini } 69807fb8a5e4SKarl Rupp ierr = MPI_Recv(&idx,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 6981b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6982b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 698357de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 698427b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 698527b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 698627b6a85dSStefano Zampini #endif 69877fb8a5e4SKarl Rupp ranks_send_to_idx[0] = procs_candidates[oldranks[idx]]; 698827b6a85dSStefano Zampini } else if (oldranks) { 69897fb8a5e4SKarl Rupp ranks_send_to_idx[0] = oldranks[idx]; 699027b6a85dSStefano Zampini } else { 69917fb8a5e4SKarl Rupp ranks_send_to_idx[0] = idx; 6992e7931f94SStefano Zampini } 699357de7509SStefano Zampini } 699452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6995e7931f94SStefano Zampini /* clean up */ 6996e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 699752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 6998e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 6999e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 7000e7931f94SStefano Zampini } 700127b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 700257de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 7003e7931f94SStefano Zampini 7004e7931f94SStefano Zampini /* assemble parallel IS for sends */ 7005e7931f94SStefano Zampini i = 1; 700627b6a85dSStefano Zampini if (!color) i=0; 700757de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 7008e7931f94SStefano Zampini PetscFunctionReturn(0); 7009e7931f94SStefano Zampini } 7010e7931f94SStefano Zampini 7011e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 7012e7931f94SStefano Zampini 70131e0482f5SStefano 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[]) 7014e7931f94SStefano Zampini { 701570cf5478SStefano Zampini Mat local_mat; 7016e7931f94SStefano Zampini IS is_sends_internal; 70179d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 70181ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 70199d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 7020e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 7021e7931f94SStefano Zampini PetscInt* l2gmap_indices; 7022e7931f94SStefano Zampini const PetscInt* is_indices; 7023e7931f94SStefano Zampini MatType new_local_type; 7024e7931f94SStefano Zampini /* buffers */ 7025e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 702628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 70279d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 7028e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 70291ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 7030e7931f94SStefano Zampini /* MPI */ 703128143c3dSStefano Zampini MPI_Comm comm,comm_n; 703228143c3dSStefano Zampini PetscSubcomm subcomm; 7033e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 703428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 703528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 70361ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 70371ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 70381ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 7039e7931f94SStefano Zampini PetscErrorCode ierr; 7040e7931f94SStefano Zampini 7041e7931f94SStefano Zampini PetscFunctionBegin; 704257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 7043e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 7044fbfcfee5SBarry 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); 704557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 704657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 704757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 704857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 704957de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 70501ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 70511ae86dd6SStefano Zampini if (nvecs) { 70521ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 70531ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 70541ae86dd6SStefano Zampini } 705557de7509SStefano Zampini /* further checks */ 7056e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 7057e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 7058e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 7059e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 7060e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 706157de7509SStefano Zampini if (reuse && *mat_n) { 706270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 706357de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 706470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 706528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 706670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 706770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 706870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 706970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 707070cf5478SStefano Zampini } 7071e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 7072e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 707357de7509SStefano Zampini 7074e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 7075e7931f94SStefano Zampini if (!is_sends) { 707628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 7077bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 7078c8587f34SStefano Zampini } else { 7079e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 7080e7931f94SStefano Zampini is_sends_internal = is_sends; 7081c8587f34SStefano Zampini } 7082e7931f94SStefano Zampini 7083e7931f94SStefano Zampini /* get comm */ 7084a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 7085e7931f94SStefano Zampini 7086e7931f94SStefano Zampini /* compute number of sends */ 7087e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 7088e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 7089e7931f94SStefano Zampini 7090e7931f94SStefano Zampini /* compute number of receives */ 7091e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 7092785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 7093e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 7094e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 7095e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 7096e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 7097e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 7098e7931f94SStefano Zampini 709928143c3dSStefano Zampini /* restrict comm if requested */ 710028143c3dSStefano Zampini subcomm = 0; 710128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 710228143c3dSStefano Zampini if (restrict_comm) { 7103779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 7104779c1cceSStefano Zampini 710528143c3dSStefano Zampini color = 0; 710653a05cb3SStefano Zampini if (restrict_full) { 710753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 710853a05cb3SStefano Zampini } else { 710953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 711053a05cb3SStefano Zampini } 7111b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 711228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 711328143c3dSStefano Zampini /* check if reuse has been requested */ 711457de7509SStefano Zampini if (reuse) { 711528143c3dSStefano Zampini if (*mat_n) { 711628143c3dSStefano Zampini PetscMPIInt subcommsize2; 711728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 711828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 711928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 712028143c3dSStefano Zampini } else { 712128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 712228143c3dSStefano Zampini } 712328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 7124779c1cceSStefano Zampini PetscMPIInt rank; 7125779c1cceSStefano Zampini 7126779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 712728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 712828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 712928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 7130306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 713128143c3dSStefano Zampini } 713228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 713328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 713428143c3dSStefano Zampini } else { 713528143c3dSStefano Zampini comm_n = comm; 713628143c3dSStefano Zampini } 713728143c3dSStefano Zampini 7138e7931f94SStefano Zampini /* prepare send/receive buffers */ 7139785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 7140e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 7141785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 7142e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 714328143c3dSStefano Zampini if (nis) { 7144854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 714528143c3dSStefano Zampini } 7146e7931f94SStefano Zampini 714728143c3dSStefano Zampini /* Get data from local matrices */ 71486c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 7149e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 7150e7931f94SStefano Zampini /* 7151e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 7152e7931f94SStefano Zampini send_buffer_idxs should contain: 7153e7931f94SStefano Zampini - MatType_PRIVATE type 7154e7931f94SStefano Zampini - PetscInt size_of_l2gmap 7155e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 7156e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 7157e7931f94SStefano Zampini */ 71586c4ed002SBarry Smith else { 7159e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 71603bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 7161854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 7162e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 7163e7931f94SStefano Zampini send_buffer_idxs[1] = i; 71643bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 7165e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 71663bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 7167e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 7168e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 7169e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 7170e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 7171c8587f34SStefano Zampini } 7172c8587f34SStefano Zampini } 7173e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 717428143c3dSStefano Zampini /* additional is (if any) */ 717528143c3dSStefano Zampini if (nis) { 717628143c3dSStefano Zampini PetscMPIInt psum; 717728143c3dSStefano Zampini PetscInt j; 717828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 717928143c3dSStefano Zampini PetscInt plen; 718028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 718128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 718228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 718328143c3dSStefano Zampini } 7184854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 718528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 718628143c3dSStefano Zampini PetscInt plen; 718728143c3dSStefano Zampini const PetscInt *is_array_idxs; 718828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 718928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 719028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 719128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 719228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 719328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 719428143c3dSStefano Zampini } 719528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 719628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 719728143c3dSStefano Zampini } 719828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 719928143c3dSStefano Zampini } 72003b3b1effSJed Brown ierr = MatISRestoreLocalMat(mat,&local_mat);CHKERRQ(ierr); 720128143c3dSStefano Zampini 7202e7931f94SStefano Zampini buf_size_idxs = 0; 7203e7931f94SStefano Zampini buf_size_vals = 0; 720428143c3dSStefano Zampini buf_size_idxs_is = 0; 72051ae86dd6SStefano Zampini buf_size_vecs = 0; 7206e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7207e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 7208e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 720928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 72101ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 7211e7931f94SStefano Zampini } 7212785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 7213785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 721495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 72151ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 7216e7931f94SStefano Zampini 7217e7931f94SStefano Zampini /* get new tags for clean communications */ 7218e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 7219e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 722028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 72211ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 7222e7931f94SStefano Zampini 7223e7931f94SStefano Zampini /* allocate for requests */ 7224785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 7225785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 722695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 72271ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 7228785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 7229785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 723095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 72311ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 7232e7931f94SStefano Zampini 7233e7931f94SStefano Zampini /* communications */ 7234e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 7235e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 723628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 72371ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 7238e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7239e7931f94SStefano Zampini source_dest = onodes[i]; 7240e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 7241e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 7242e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7243e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 724428143c3dSStefano Zampini if (nis) { 724557de7509SStefano Zampini source_dest = onodes_is[i]; 724628143c3dSStefano Zampini ierr = MPI_Irecv(ptr_idxs_is,olengths_idxs_is[i],MPIU_INT,source_dest,tag_idxs_is,comm,&recv_req_idxs_is[i]);CHKERRQ(ierr); 724728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 724828143c3dSStefano Zampini } 72491ae86dd6SStefano Zampini if (nvecs) { 72501ae86dd6SStefano Zampini source_dest = onodes[i]; 72511ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 72521ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 72531ae86dd6SStefano Zampini } 7254e7931f94SStefano Zampini } 7255e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 7256e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 7257e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 7258e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 725928143c3dSStefano Zampini if (nis) { 726028143c3dSStefano Zampini ierr = MPI_Isend(send_buffer_idxs_is,ilengths_idxs_is[source_dest],MPIU_INT,source_dest,tag_idxs_is,comm,&send_req_idxs_is[i]);CHKERRQ(ierr); 726128143c3dSStefano Zampini } 72621ae86dd6SStefano Zampini if (nvecs) { 72631ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72641ae86dd6SStefano Zampini ierr = MPI_Isend(send_buffer_vecs,ilengths_idxs[source_dest]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&send_req_vecs[i]);CHKERRQ(ierr); 72651ae86dd6SStefano Zampini } 7266e7931f94SStefano Zampini } 7267e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 7268e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 7269e7931f94SStefano Zampini 7270e7931f94SStefano Zampini /* assemble new l2g map */ 7271e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7272e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 72739d30be91SStefano Zampini new_local_rows = 0; 7274e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 72759d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 7276e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7277e7931f94SStefano Zampini } 72789d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 7279e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 72809d30be91SStefano Zampini new_local_rows = 0; 7281e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 72829d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 72839d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 7284e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7285e7931f94SStefano Zampini } 72869d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 72879d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 7288e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 7289e7931f94SStefano Zampini 7290e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 7291e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 7292e7931f94SStefano 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) */ 7293e7931f94SStefano Zampini if (n_recvs) { 729428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 7295e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 7296e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7297e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 7298e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 7299e7931f94SStefano Zampini break; 7300e7931f94SStefano Zampini } 7301e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7302e7931f94SStefano Zampini } 7303e7931f94SStefano Zampini switch (new_local_type_private) { 730428143c3dSStefano Zampini case MATDENSE_PRIVATE: 7305e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7306e7931f94SStefano Zampini bs = 1; 7307e7931f94SStefano Zampini break; 7308e7931f94SStefano Zampini case MATAIJ_PRIVATE: 7309e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7310e7931f94SStefano Zampini bs = 1; 7311e7931f94SStefano Zampini break; 7312e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 7313e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 7314e7931f94SStefano Zampini break; 7315e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 7316e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 7317e7931f94SStefano Zampini break; 7318e7931f94SStefano Zampini default: 7319fbfcfee5SBarry Smith SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,PETSC_FUNCTION_NAME); 7320e7931f94SStefano Zampini break; 7321e7931f94SStefano Zampini } 7322ed8ed4edSstefano_zampini } else { /* by default, new_local_type is seqaij */ 7323ed8ed4edSstefano_zampini new_local_type = MATSEQAIJ; 732428143c3dSStefano Zampini bs = 1; 7325e7931f94SStefano Zampini } 7326e7931f94SStefano Zampini 732770cf5478SStefano Zampini /* create MATIS object if needed */ 732857de7509SStefano Zampini if (!reuse) { 7329e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 7330e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 733170cf5478SStefano Zampini } else { 733270cf5478SStefano Zampini /* it also destroys the local matrices */ 733357de7509SStefano Zampini if (*mat_n) { 733470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 733557de7509SStefano Zampini } else { /* this is a fake object */ 733657de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 733757de7509SStefano Zampini } 733870cf5478SStefano Zampini } 733970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 7340e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 73419d30be91SStefano Zampini 73429d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 73439d30be91SStefano Zampini 73449d30be91SStefano Zampini /* Global to local map of received indices */ 73459d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 73469d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 73479d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 73489d30be91SStefano Zampini 73499d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 73509d30be91SStefano Zampini buf_size_idxs = 0; 73519d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 73529d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 73539d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 73549d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 73559d30be91SStefano Zampini } 73569d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 73579d30be91SStefano Zampini 73589d30be91SStefano Zampini /* set preallocation */ 73599d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 73609d30be91SStefano Zampini if (!newisdense) { 73619d30be91SStefano Zampini PetscInt *new_local_nnz=0; 73629d30be91SStefano Zampini 73639d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 73649d30be91SStefano Zampini if (n_recvs) { 73659d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 73669d30be91SStefano Zampini } 73679d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 73689d30be91SStefano Zampini PetscInt j; 73699d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 73709d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 73719d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 73729d30be91SStefano Zampini } 73739d30be91SStefano Zampini } else { 73749d30be91SStefano Zampini /* TODO */ 73759d30be91SStefano Zampini } 73769d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 73779d30be91SStefano Zampini } 73789d30be91SStefano Zampini if (new_local_nnz) { 73799d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 73809d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 73819d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 73829d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 73839d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 73849d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 73859d30be91SStefano Zampini } else { 73869d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 73879d30be91SStefano Zampini } 73889d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 73899d30be91SStefano Zampini } else { 73909d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 73919d30be91SStefano Zampini } 7392e7931f94SStefano Zampini 7393e7931f94SStefano Zampini /* set values */ 7394e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 73959d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 7396e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7397e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 7398e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 73999d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 7400e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7401e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7402e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 740328143c3dSStefano Zampini } else { 740428143c3dSStefano Zampini /* TODO */ 7405e7931f94SStefano Zampini } 7406e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7407e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 7408e7931f94SStefano Zampini } 7409e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7410e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 74113b3b1effSJed Brown ierr = MatISRestoreLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 741270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 741370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 74149d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 7415e7931f94SStefano Zampini 7416dfd14d43SStefano Zampini #if 0 741728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 7418e7931f94SStefano Zampini Vec lvec,rvec; 7419e7931f94SStefano Zampini PetscReal infty_error; 7420e7931f94SStefano Zampini 74212a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 7422e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 7423e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 7424e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 742570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 7426e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 7427e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 7428e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 7429e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 7430e7931f94SStefano Zampini } 743128143c3dSStefano Zampini #endif 7432e7931f94SStefano Zampini 743328143c3dSStefano Zampini /* assemble new additional is (if any) */ 743428143c3dSStefano Zampini if (nis) { 743528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 743628143c3dSStefano Zampini 743728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7438854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 743928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 744028143c3dSStefano Zampini psum = 0; 744128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 744228143c3dSStefano Zampini for (j=0;j<nis;j++) { 744328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 744428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 744528143c3dSStefano Zampini psum += plen; 744628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 744728143c3dSStefano Zampini } 744828143c3dSStefano Zampini } 7449854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 7450854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 745128143c3dSStefano Zampini for (i=1;i<nis;i++) { 745228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 745328143c3dSStefano Zampini } 745428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 745528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 745628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 745728143c3dSStefano Zampini for (j=0;j<nis;j++) { 745828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 745928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 746028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 746128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 746228143c3dSStefano Zampini } 746328143c3dSStefano Zampini } 746428143c3dSStefano Zampini for (i=0;i<nis;i++) { 746528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 746628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 746728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 746828143c3dSStefano Zampini } 746928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 747028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 747128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 747228143c3dSStefano Zampini } 7473e7931f94SStefano Zampini /* free workspace */ 747428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 7475e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7476e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 7477e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7478e7931f94SStefano Zampini if (isdense) { 7479e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 7480e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 74813b3b1effSJed Brown ierr = MatISRestoreLocalMat(mat,&local_mat);CHKERRQ(ierr); 7482e7931f94SStefano Zampini } else { 7483e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 7484e7931f94SStefano Zampini } 748528143c3dSStefano Zampini if (nis) { 748628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 748728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 748828143c3dSStefano Zampini } 74891ae86dd6SStefano Zampini 74901ae86dd6SStefano Zampini if (nvecs) { 74911ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 74921ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 74931ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 74941ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 74951ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 74961ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 74971ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 74981ae86dd6SStefano Zampini /* set values */ 74991ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 75001ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 75011ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 75021ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 75031ae86dd6SStefano Zampini PetscInt j; 75041ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 75051ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 75061ae86dd6SStefano Zampini } 75071ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 75081ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 75091ae86dd6SStefano Zampini } 75101ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 75111ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 75121ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 75131ae86dd6SStefano Zampini } 75141ae86dd6SStefano Zampini 75151ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 75161ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 7517e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 7518e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 75191ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 752028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 7521e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 7522e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 75231ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 752428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 7525e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 7526e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 7527e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 7528e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 7529e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 753028143c3dSStefano Zampini if (nis) { 753128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 753228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 753328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 753428143c3dSStefano Zampini } 753528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 753628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 753728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 753828143c3dSStefano Zampini for (i=0;i<nis;i++) { 753928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 754028143c3dSStefano Zampini } 75411ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 75421ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 75431ae86dd6SStefano Zampini } 754453a05cb3SStefano Zampini *mat_n = NULL; 754528143c3dSStefano Zampini } 7546e7931f94SStefano Zampini PetscFunctionReturn(0); 7547e7931f94SStefano Zampini } 7548a57a6d2fSStefano Zampini 754912edc857SStefano Zampini /* temporary hack into ksp private data structure */ 7550af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 755112edc857SStefano Zampini 7552c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 7553c8587f34SStefano Zampini { 7554c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 7555c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 755620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 75571ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 75581e0482f5SStefano Zampini Mat coarseG,t_coarse_mat_is; 75599881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 756020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 75616e683305SStefano Zampini IS coarse_is,*isarray; 75626e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 756330368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 7564f9eb5b7dSStefano Zampini PC pc_temp; 7565c8587f34SStefano Zampini PCType coarse_pc_type; 7566c8587f34SStefano Zampini KSPType coarse_ksp_type; 7567f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 75687274672aSStefano Zampini PetscBool coarse_reuse; 75691e0482f5SStefano Zampini PetscInt ncoarse,nedcfield; 757068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 757122bc73bbSStefano Zampini PetscScalar *array; 757257de7509SStefano Zampini MatReuse coarse_mat_reuse; 757357de7509SStefano Zampini PetscBool restr, full_restr, have_void; 75747de4f681Sstefano_zampini PetscMPIInt commsize; 75759881197aSStefano Zampini PetscErrorCode ierr; 7576fdc09c96SStefano Zampini 7577c8587f34SStefano Zampini PetscFunctionBegin; 7578c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 757968457ee5SStefano 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 */ 7580fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 75815a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 75827de4f681Sstefano_zampini 75837de4f681Sstefano_zampini pcbddc->new_primal_space = PETSC_TRUE; 7584fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 7585f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 7586f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 7587f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 7588fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 758951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 759051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 7591727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 7592fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7593fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 7594fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7595f4ddd8eeSStefano Zampini } 7596fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 7597fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7598f4ddd8eeSStefano Zampini } 759970cf5478SStefano Zampini /* reset any subassembling information */ 760057de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 760170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 760257de7509SStefano Zampini } 76036e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 7604fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7605f4ddd8eeSStefano Zampini } 760657de7509SStefano Zampini /* assemble coarse matrix */ 760757de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 760857de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 760957de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 761057de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 761118a45a71SStefano Zampini } else { 761257de7509SStefano Zampini coarse_mat = NULL; 761357de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 76146e683305SStefano Zampini } 7615e7931f94SStefano Zampini 7616abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 7617abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 7618abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 7619abbbba34SStefano Zampini 7620abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 762122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 762222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 762322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 762422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 7625e176bc59SStefano 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); 76266e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 76276e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 76286e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7629abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 7630abbbba34SStefano Zampini 763157de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 763257de7509SStefano Zampini im_active = !!(pcis->n); 763357de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 763457de7509SStefano Zampini 763514f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 763657de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 763757de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 76387de4f681Sstefano_zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&commsize);CHKERRQ(ierr); 763957de7509SStefano Zampini coarse_mat_is = NULL; 764057de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 764157de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 76421ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 764357de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 764457de7509SStefano Zampini if (multilevel_requested) { 764557de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 764657de7509SStefano Zampini restr = PETSC_FALSE; 764757de7509SStefano Zampini full_restr = PETSC_FALSE; 764857de7509SStefano Zampini } else { 764957de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 765057de7509SStefano Zampini restr = PETSC_TRUE; 765157de7509SStefano Zampini full_restr = PETSC_TRUE; 765257de7509SStefano Zampini } 76537de4f681Sstefano_zampini if (!pcbddc->coarse_size || commsize == 1) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 765457de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 765557de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 7656a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 7657bb360cb4SStefano Zampini if (multilevel_requested) { 7658bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7659bb360cb4SStefano Zampini } else { 7660bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7661bb360cb4SStefano Zampini } 7662a198735bSStefano Zampini } else { 76637de4f681Sstefano_zampini PetscMPIInt rank; 7664a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 76657de4f681Sstefano_zampini have_void = (active_procs == (PetscInt)commsize) ? PETSC_FALSE : PETSC_TRUE; 7666a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 7667a198735bSStefano Zampini } 766857de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 766957de7509SStefano Zampini PetscInt psum; 767057de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 767157de7509SStefano Zampini else psum = 0; 767257de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 76737de4f681Sstefano_zampini if (ncoarse < commsize) have_void = PETSC_TRUE; 767457de7509SStefano Zampini } 767557de7509SStefano Zampini /* determine if we can go multilevel */ 767657de7509SStefano Zampini if (multilevel_requested) { 767757de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 767857de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 767957de7509SStefano Zampini } 768057de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 768157de7509SStefano Zampini 7682e4d548c7SStefano Zampini /* dump subassembling pattern */ 7683e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 7684e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 7685e4d548c7SStefano Zampini } 7686e4d548c7SStefano Zampini 76876e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 76881e0482f5SStefano Zampini nedcfield = -1; 76891e0482f5SStefano Zampini if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal)) { /* protects from unneded computations */ 76906e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 76916e683305SStefano Zampini const PetscInt *idxs; 76926e683305SStefano Zampini ISLocalToGlobalMapping tmap; 76936e683305SStefano Zampini 76946e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 76950be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 76966e683305SStefano Zampini /* allocate space for temporary storage */ 7697854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 7698854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 76996e683305SStefano Zampini /* allocate for IS array */ 77006e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 77011e0482f5SStefano Zampini if (pcbddc->nedclocal) { 77021e0482f5SStefano Zampini if (pcbddc->nedfield > -1) { 77031e0482f5SStefano Zampini nedcfield = pcbddc->nedfield; 77041e0482f5SStefano Zampini } else { 77051e0482f5SStefano Zampini nedcfield = 0; 77061e0482f5SStefano Zampini if (nisdofs) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen (%d)",nisdofs); 77071e0482f5SStefano Zampini nisdofs = 1; 77081e0482f5SStefano Zampini } 77091e0482f5SStefano Zampini } 77106e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 771127b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 771230368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 7713854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 77146e683305SStefano Zampini /* dofs splitting */ 77156e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 77166e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 77171e0482f5SStefano Zampini if (nedcfield != i) { 77186e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 77196e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 77206e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 77216e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 77221e0482f5SStefano Zampini } else { 77231e0482f5SStefano Zampini ierr = ISGetLocalSize(pcbddc->nedclocal,&tsize);CHKERRQ(ierr); 77241e0482f5SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 77251e0482f5SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 7726eee23b56SStefano Zampini if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping coarse nedelec field! %d != %d\n",tsize,nout); 77271e0482f5SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 77281e0482f5SStefano Zampini } 77296e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 773030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 77316e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 77326e683305SStefano Zampini } 77336e683305SStefano Zampini /* neumann boundaries */ 77346e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 77356e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 77366e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 77376e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 77386e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 77396e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 77406e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 774130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 77426e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 77436e683305SStefano Zampini } 77446e683305SStefano Zampini /* free memory */ 77456e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 77466e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 77476e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 77486e683305SStefano Zampini } else { 77496e683305SStefano Zampini nis = 0; 77506e683305SStefano Zampini nisdofs = 0; 77516e683305SStefano Zampini nisneu = 0; 775230368db7SStefano Zampini nisvert = 0; 77536e683305SStefano Zampini isarray = NULL; 77546e683305SStefano Zampini } 77556e683305SStefano Zampini /* destroy no longer needed map */ 77566e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 77576e683305SStefano Zampini 775857de7509SStefano Zampini /* subassemble */ 775957de7509SStefano Zampini if (multilevel_allowed) { 77601ae86dd6SStefano Zampini Vec vp[1]; 77611ae86dd6SStefano Zampini PetscInt nvecs = 0; 776257de7509SStefano Zampini PetscBool reuse,reuser; 77631ae86dd6SStefano Zampini 776457de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 776557de7509SStefano Zampini else reuse = PETSC_FALSE; 776657de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 77671ae86dd6SStefano Zampini vp[0] = NULL; 77681ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 77691ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 77701ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 77711ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 77721ae86dd6SStefano Zampini nvecs = 1; 77731ae86dd6SStefano Zampini 77741ae86dd6SStefano Zampini if (pcbddc->divudotp) { 7775a198735bSStefano Zampini Mat B,loc_divudotp; 77761ae86dd6SStefano Zampini Vec v,p; 77771ae86dd6SStefano Zampini IS dummy; 77781ae86dd6SStefano Zampini PetscInt np; 77791ae86dd6SStefano Zampini 7780a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 7781a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 77821ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 77837dae84e0SHong Zhang ierr = MatCreateSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 77841ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 77851ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 77861ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 77871ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 77881ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 77891ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 77901ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 77911ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 77921ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 77931ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 77941ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 77951ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 779674e2c79eSStefano Zampini } 77971ae86dd6SStefano Zampini } 77981ae86dd6SStefano Zampini if (reuser) { 77991e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 780074e2c79eSStefano Zampini } else { 78011e0482f5SStefano 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); 78021ae86dd6SStefano Zampini } 78031ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 78041ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 78051ae86dd6SStefano Zampini PetscInt nl; 78061ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 78071ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 78081ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 78091ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 78101ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 78111ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 78121ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 78131ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 7814a198735bSStefano Zampini } else { 7815a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 78161ae86dd6SStefano Zampini } 78171ae86dd6SStefano Zampini } else { 78181e0482f5SStefano 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); 78196e683305SStefano Zampini } 782057de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 782157de7509SStefano Zampini PetscMPIInt size; 7822f913dca9SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size);CHKERRQ(ierr); 782357de7509SStefano Zampini if (!multilevel_allowed) { 782457de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 78256e683305SStefano Zampini } else { 782657de7509SStefano Zampini Mat A; 7827779c1cceSStefano Zampini 782857de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 782957de7509SStefano Zampini if (coarse_mat_is) { 783057de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 783157de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 783257de7509SStefano Zampini coarse_mat = coarse_mat_is; 783357de7509SStefano Zampini } 783457de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 783557de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 783657de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 7837779c1cceSStefano Zampini } 7838779c1cceSStefano Zampini } 783957de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 784057de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 78416e683305SStefano Zampini 78426e683305SStefano Zampini /* create local to global scatters for coarse problem */ 784368457ee5SStefano Zampini if (compute_vecs) { 78446e683305SStefano Zampini PetscInt lrows; 78456e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 784657de7509SStefano Zampini if (coarse_mat) { 784757de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 78486e683305SStefano Zampini } else { 78496e683305SStefano Zampini lrows = 0; 78506e683305SStefano Zampini } 78516e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 78526e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 78536e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 78546e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 78556e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 78566e683305SStefano Zampini } 78576e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 7858c8587f34SStefano Zampini 7859f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 7860f9eb5b7dSStefano Zampini if (multilevel_allowed) { 7861f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 7862f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 7863f9eb5b7dSStefano Zampini } else { 7864f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 7865f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 7866c8587f34SStefano Zampini } 7867c8587f34SStefano Zampini 78686e683305SStefano Zampini /* print some info if requested */ 78696e683305SStefano Zampini if (pcbddc->dbg_flag) { 78706e683305SStefano Zampini if (!multilevel_allowed) { 78716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 78726e683305SStefano Zampini if (multilevel_requested) { 78736e683305SStefano 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); 78746e683305SStefano Zampini } else if (pcbddc->max_levels) { 78756e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 78766e683305SStefano Zampini } 78776e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 78786e683305SStefano Zampini } 78796e683305SStefano Zampini } 78806e683305SStefano Zampini 78811e0482f5SStefano Zampini /* communicate coarse discrete gradient */ 78821e0482f5SStefano Zampini coarseG = NULL; 78831e0482f5SStefano Zampini if (pcbddc->nedcG && multilevel_allowed) { 78841e0482f5SStefano Zampini MPI_Comm ccomm; 78851e0482f5SStefano Zampini if (coarse_mat) { 78861e0482f5SStefano Zampini ccomm = PetscObjectComm((PetscObject)coarse_mat); 78871e0482f5SStefano Zampini } else { 78881e0482f5SStefano Zampini ccomm = MPI_COMM_NULL; 78891e0482f5SStefano Zampini } 78901e0482f5SStefano Zampini ierr = MatMPIAIJRestrict(pcbddc->nedcG,ccomm,&coarseG);CHKERRQ(ierr); 78911e0482f5SStefano Zampini } 78921e0482f5SStefano Zampini 7893f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 789457de7509SStefano Zampini if (coarse_mat) { 78957274672aSStefano Zampini PetscBool isredundant,isnn,isbddc; 78966a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 78977274672aSStefano Zampini 78986e683305SStefano Zampini if (pcbddc->dbg_flag) { 789957de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 79006e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 79016e683305SStefano Zampini } 7902f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 7903312be037SStefano Zampini char prefix[256],str_level[16]; 7904e604994aSStefano Zampini size_t len; 79051e0482f5SStefano Zampini 790657de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 7907422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 7908c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 7909f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 791057de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7911c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 79126e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 7913c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 79141e0482f5SStefano Zampini /* TODO is this logic correct? should check for coarse_mat type */ 7915c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7916e604994aSStefano Zampini /* prefix */ 7917e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 7918e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 7919e604994aSStefano Zampini if (!pcbddc->current_level) { 7920e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 7921e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 7922c8587f34SStefano Zampini } else { 7923e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 7924312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 7925312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 792634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 792735529e7bSStefano Zampini ierr = PetscSNPrintf(str_level,sizeof(str_level),"l%d_",(int)(pcbddc->current_level));CHKERRQ(ierr); 7928e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 7929e604994aSStefano Zampini } 7930e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 79313e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 79323e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 79333e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 79343e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 7935f9eb5b7dSStefano Zampini /* allow user customization */ 7936f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 79373e3c6dadSStefano Zampini } 79383e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 793951bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 79403e3c6dadSStefano Zampini if (nisdofs) { 79413e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 79423e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 79433e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 79443e3c6dadSStefano Zampini } 79453e3c6dadSStefano Zampini } 79463e3c6dadSStefano Zampini if (nisneu) { 79473e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 79483e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 7949312be037SStefano Zampini } 795030368db7SStefano Zampini if (nisvert) { 795130368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 795230368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 795330368db7SStefano Zampini } 79541e0482f5SStefano Zampini if (coarseG) { 79551e0482f5SStefano Zampini ierr = PCBDDCSetDiscreteGradient(pc_temp,coarseG,1,nedcfield,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 79561e0482f5SStefano Zampini } 7957f9eb5b7dSStefano Zampini 7958f9eb5b7dSStefano Zampini /* get some info after set from options */ 7959f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 7960b76f3995Sstefano_zampini /* multilevel can only be requested via -pc_bddc_levels or PCBDDCSetLevels */ 7961b76f3995Sstefano_zampini if (isbddc && !multilevel_allowed) { 7962f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7963f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 7964f9eb5b7dSStefano Zampini } 7965b76f3995Sstefano_zampini /* multilevel cannot be done with coarse PCs different from BDDC or NN */ 79667274672aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 79677274672aSStefano Zampini if (multilevel_requested && multilevel_allowed && !isbddc && !isnn) { 7968b76f3995Sstefano_zampini ierr = PCSetType(pc_temp,PCBDDC);CHKERRQ(ierr); 7969b76f3995Sstefano_zampini isbddc = PETSC_TRUE; 7970b76f3995Sstefano_zampini } 79717274672aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 79724f3a063dSStefano Zampini if (isredundant) { 79734f3a063dSStefano Zampini KSP inner_ksp; 79744f3a063dSStefano Zampini PC inner_pc; 79759326c5c6Sstefano_zampini 79764f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 79774f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 79784f3a063dSStefano Zampini } 7979f9eb5b7dSStefano Zampini 798057de7509SStefano Zampini /* parameters which miss an API */ 79817274672aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 798257de7509SStefano Zampini if (isbddc) { 7983720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 79847274672aSStefano Zampini 7985720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 798657de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 798727b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 798827b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 7989a198735bSStefano Zampini Mat coarsedivudotp_is; 7990a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 7991a198735bSStefano Zampini IS row,col; 7992a198735bSStefano Zampini const PetscInt *gidxs; 7993a198735bSStefano Zampini PetscInt n,st,M,N; 7994a198735bSStefano Zampini 7995a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 7996a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 7997a198735bSStefano Zampini st = st-n; 7998a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 7999a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 8000a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 8001a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 8002a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 8003a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 8004a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 8005a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 8006a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 8007a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 8008a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 8009a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 8010a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 8011a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 8012a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 8013a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 8014a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 8015a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 8016a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 8017a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 80188ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 8019a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 8020720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 802159e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 8022720d30f9SStefano Zampini } 8023d4d8cf7bSStefano Zampini } 80249881197aSStefano Zampini 80253301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 80265a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 80273301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 80283301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 80293301b35fSStefano Zampini } 80303301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 80313301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 80323301b35fSStefano Zampini } 80333301b35fSStefano Zampini if (pc->pmat->spd_set) { 80343301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 80353301b35fSStefano Zampini } 803627b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 803727b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 803827b6a85dSStefano Zampini } 80396e683305SStefano Zampini /* set operators */ 80405f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 80416e683305SStefano Zampini if (pcbddc->dbg_flag) { 80426e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 80436e683305SStefano Zampini } 80446e683305SStefano Zampini } 80451e0482f5SStefano Zampini ierr = MatDestroy(&coarseG);CHKERRQ(ierr); 80466e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 8047b1ecc7b1SStefano Zampini #if 0 8048b9b85e73SStefano Zampini { 8049b9b85e73SStefano Zampini PetscViewer viewer; 8050b9b85e73SStefano Zampini char filename[256]; 8051b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 8052b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 80536a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 8054b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 8055f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 8056b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 8057b9b85e73SStefano Zampini } 8058b9b85e73SStefano Zampini #endif 8059f9eb5b7dSStefano Zampini 806098a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 806198a51de6SStefano Zampini Vec crhs,csol; 806204708bb6SStefano Zampini 8063f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 8064f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 8065f347579bSStefano Zampini if (!csol) { 80662a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 8067f9eb5b7dSStefano Zampini } 8068f347579bSStefano Zampini if (!crhs) { 80692a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 8070f347579bSStefano Zampini } 8071b0f5fe93SStefano Zampini } 80721ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 8073b0f5fe93SStefano Zampini 8074b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 8075b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 8076b0f5fe93SStefano Zampini 8077b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 80784f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 80794f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 80804f1b2e48SStefano Zampini } 8081b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 8082b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 8083b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8084b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8085b0f5fe93SStefano Zampini if (coarse_mat) { 8086b0f5fe93SStefano Zampini Vec nullv; 8087b0f5fe93SStefano Zampini PetscScalar *array,*array2; 8088b0f5fe93SStefano Zampini PetscInt nl; 8089b0f5fe93SStefano Zampini 8090b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 8091b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 8092b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 8093b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 8094b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 8095b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 8096b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 8097b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 8098b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 8099b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 8100b0f5fe93SStefano Zampini } 8101b0f5fe93SStefano Zampini } 8102b0f5fe93SStefano Zampini 8103b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 8104b0f5fe93SStefano Zampini PetscBool ispreonly; 8105b0f5fe93SStefano Zampini 8106b0f5fe93SStefano Zampini if (CoarseNullSpace) { 8107b0f5fe93SStefano Zampini PetscBool isnull; 8108b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 8109bef83e63SStefano Zampini if (isnull) { 8110b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 8111b0f5fe93SStefano Zampini } 8112bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 8113b0f5fe93SStefano Zampini } 8114b0f5fe93SStefano Zampini /* setup coarse ksp */ 8115b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 8116cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 8117cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 81186e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 8119c8587f34SStefano Zampini KSP check_ksp; 81202b510759SStefano Zampini KSPType check_ksp_type; 8121c8587f34SStefano Zampini PC check_pc; 81226e683305SStefano Zampini Vec check_vec,coarse_vec; 81236a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 81242b510759SStefano Zampini PetscInt its; 81256e683305SStefano Zampini PetscBool compute_eigs; 81266e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 81276e683305SStefano Zampini PetscInt neigs; 81288e185a42SStefano Zampini const char *prefix; 8129c8587f34SStefano Zampini 81302b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 81316e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 8132422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 813323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 8134f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 8135e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 8136e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 8137e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 81382b510759SStefano Zampini if (ispreonly) { 81392b510759SStefano Zampini check_ksp_type = KSPPREONLY; 81406e683305SStefano Zampini compute_eigs = PETSC_FALSE; 81412b510759SStefano Zampini } else { 8142cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 81436e683305SStefano Zampini compute_eigs = PETSC_TRUE; 8144c8587f34SStefano Zampini } 8145c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 81466e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 81476e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 81486e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 8149a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 8150a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 8151a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 8152a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 8153c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 8154c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 8155c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 8156c8587f34SStefano Zampini /* create random vec */ 81572701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 8158c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 81596e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 8160c8587f34SStefano Zampini /* solve coarse problem */ 81616e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 8162cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 81636e683305SStefano Zampini if (compute_eigs) { 8164854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 8165854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 81666e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 81671ae86dd6SStefano Zampini if (neigs) { 81686e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 81696e683305SStefano Zampini lambda_min = eigs_r[0]; 81706e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 81712701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 81722701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 8173cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 8174cbcc2c2aSStefano Zampini } 8175c8587f34SStefano Zampini } 8176c8587f34SStefano Zampini } 81771ae86dd6SStefano Zampini } 8178cbcc2c2aSStefano Zampini 8179c8587f34SStefano Zampini /* check coarse problem residual error */ 81806e683305SStefano Zampini if (pcbddc->dbg_flag) { 81816e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 81826e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 81836e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 8184c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 81856e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 81866e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 8187779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 81886e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 81896e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 81906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 81916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 8192b0f5fe93SStefano Zampini if (CoarseNullSpace) { 8193b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 8194b0f5fe93SStefano Zampini } 81956e683305SStefano Zampini if (compute_eigs) { 81966e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 8197b03ebc13SStefano Zampini KSPConvergedReason reason; 8198deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 8199c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 8200b03ebc13SStefano Zampini ierr = KSPGetConvergedReason(check_ksp,&reason);CHKERRQ(ierr); 82016e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 8202b03ebc13SStefano 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); 82036e683305SStefano Zampini for (i=0;i<neigs;i++) { 82046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 8205c8587f34SStefano Zampini } 82066e683305SStefano Zampini } 82076e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 82086e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 82096e683305SStefano Zampini } 8210e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 82112701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 8212c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 82136e683305SStefano Zampini if (compute_eigs) { 82146e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 82156e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 8216c8587f34SStefano Zampini } 82176e683305SStefano Zampini } 82186e683305SStefano Zampini } 8219bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 8220cbcc2c2aSStefano Zampini /* print additional info */ 8221cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 82226e683305SStefano Zampini /* waits until all processes reaches this point */ 82236e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 8224cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 8225cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8226cbcc2c2aSStefano Zampini } 8227cbcc2c2aSStefano Zampini 82282b510759SStefano Zampini /* free memory */ 8229fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 8230c8587f34SStefano Zampini PetscFunctionReturn(0); 8231c8587f34SStefano Zampini } 8232674ae819SStefano Zampini 8233f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 8234f34684f1SStefano Zampini { 8235f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8236f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 8237f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 8238dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 8239dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 824073be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 8241dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 8242f34684f1SStefano Zampini PetscErrorCode ierr; 8243f34684f1SStefano Zampini 8244f34684f1SStefano Zampini PetscFunctionBegin; 8245f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 82466c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 8247dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 82483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 8249dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 8250dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 82516583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 8252dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 8253dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 8254dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 82556c4ed002SBarry 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); 8256dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 8257dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 8258dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 8259dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 8260dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 8261f34684f1SStefano Zampini 8262f34684f1SStefano Zampini /* check numbering */ 8263f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 8264019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 8265dc456d91SStefano Zampini PetscInt i; 8266b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 8267f34684f1SStefano Zampini 8268f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8269f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 8270f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 82711575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8272019a44ceSStefano Zampini /* counter */ 8273019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8274019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 8275019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8276019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8277019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8278019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8279f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 8280f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 8281727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 8282f34684f1SStefano Zampini } 8283f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8284f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8285f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8286e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8287e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8288e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8289e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8290f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8291019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8292f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8293019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 82942c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 829575c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 8296b9b85e73SStefano Zampini set_error = PETSC_TRUE; 82972c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 82982c66d082SStefano 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); 8299f34684f1SStefano Zampini } 8300f34684f1SStefano Zampini } 8301019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8302b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 8303f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8304f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8305f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 8306f34684f1SStefano Zampini } 8307f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8308f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8309e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8310e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8311f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 8312f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 8313b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 8314ca8b9ea9SStefano Zampini PetscInt *gidxs; 8315ca8b9ea9SStefano Zampini 8316ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 83173bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 8318f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 8319f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8320f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 8321f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 83224bc2dc4bSStefano 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); 8323f34684f1SStefano Zampini } 8324f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8325ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 8326f34684f1SStefano Zampini } 8327f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 83281575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8329302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 8330f34684f1SStefano Zampini } 83318bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 8332f34684f1SStefano Zampini /* get back data */ 8333f34684f1SStefano Zampini *coarse_size_n = coarse_size; 8334f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 8335674ae819SStefano Zampini PetscFunctionReturn(0); 8336674ae819SStefano Zampini } 8337674ae819SStefano Zampini 8338a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 8339e456f2a8SStefano Zampini { 8340e456f2a8SStefano Zampini IS localis_t; 8341a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 8342e456f2a8SStefano Zampini PetscScalar *vals; 8343e456f2a8SStefano Zampini PetscErrorCode ierr; 8344e456f2a8SStefano Zampini 8345e456f2a8SStefano Zampini PetscFunctionBegin; 8346a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 8347e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 8348854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 8349e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 8350e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 8351a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 8352a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 83531035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 83543151afb1SStefano Zampini ierr = VecSetOption(gwork,VEC_IGNORE_NEGATIVE_INDICES,PETSC_TRUE);CHKERRQ(ierr); 8355a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 83561035eff8SStefano Zampini } 8357a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 8358e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 8359e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 8360a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 8361a7dc3881SStefano Zampini /* now compute set in local ordering */ 8362a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8363a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8364a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8365a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 8366a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8367ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8368e456f2a8SStefano Zampini lsize++; 8369e456f2a8SStefano Zampini } 8370e456f2a8SStefano Zampini } 8371854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 8372a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8373ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8374e456f2a8SStefano Zampini idxs[lsize++] = i; 8375e456f2a8SStefano Zampini } 8376e456f2a8SStefano Zampini } 8377a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8378a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 8379e456f2a8SStefano Zampini *localis = localis_t; 8380e456f2a8SStefano Zampini PetscFunctionReturn(0); 8381e456f2a8SStefano Zampini } 8382906d46d4SStefano Zampini 838308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 8384b96c3477SStefano Zampini { 8385a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8386b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8387b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 8388a64f4aa4SStefano Zampini Mat S_j; 8389b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 8390b96c3477SStefano Zampini PetscBool free_used_adj; 8391b96c3477SStefano Zampini PetscErrorCode ierr; 8392b96c3477SStefano Zampini 8393b96c3477SStefano Zampini PetscFunctionBegin; 8394b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 8395b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 839608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 8397b96c3477SStefano Zampini used_xadj = NULL; 8398b96c3477SStefano Zampini used_adjncy = NULL; 8399b96c3477SStefano Zampini } else { 840008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 840108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 840208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 840308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 8404b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 8405b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 8406b96c3477SStefano Zampini } else { 84072fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 8408b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 8409b96c3477SStefano Zampini PetscInt nvtxs; 8410b96c3477SStefano Zampini 84112fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 84122fffb893SStefano Zampini if (flg_row) { 8413b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 8414b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 8415b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 8416b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 84172fffb893SStefano Zampini } else { 84182fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 84192fffb893SStefano Zampini used_xadj = NULL; 84202fffb893SStefano Zampini used_adjncy = NULL; 84212fffb893SStefano Zampini } 84222fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 8423b96c3477SStefano Zampini } 8424b96c3477SStefano Zampini } 8425d5574798SStefano Zampini 8426d5574798SStefano Zampini /* setup sub_schurs data */ 8427a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8428df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 8429df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 8430a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 843191af6908SStefano 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); 8432a64f4aa4SStefano Zampini } else { 843372b8c272SStefano Zampini Mat change = NULL; 84349d54b7f4SStefano Zampini Vec scaling = NULL; 8435111315fdSstefano_zampini IS change_primal = NULL, iP; 8436111315fdSstefano_zampini PetscInt benign_n; 8437111315fdSstefano_zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 8438111315fdSstefano_zampini PetscBool isseqaij,need_change = PETSC_FALSE; 8439111315fdSstefano_zampini PetscBool discrete_harmonic = PETSC_FALSE; 8440a3df083aSStefano Zampini 84415feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 84425feab87aSStefano Zampini PetscInt n_vertices; 84435feab87aSStefano Zampini 84445feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 84452034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 84465feab87aSStefano Zampini } 844704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 844804708bb6SStefano Zampini if (!isseqaij) { 844904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 845004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 845104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 845204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 845304708bb6SStefano Zampini } else { 8454511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 845504708bb6SStefano Zampini } 845604708bb6SStefano Zampini } 8457a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 8458a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 8459ca92afb2SStefano Zampini } else { 8460a3df083aSStefano Zampini benign_n = 0; 8461ca92afb2SStefano Zampini } 8462b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 8463b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 8464b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 846572b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 846622db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 8467b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 846822db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 8469b7ab4a40SStefano Zampini } 8470b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 8471b7ab4a40SStefano 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 */ 8472b7ab4a40SStefano Zampini if (need_change) { 847388c03ad3SStefano Zampini PC_IS *pcisf; 847488c03ad3SStefano Zampini PC_BDDC *pcbddcf; 847588c03ad3SStefano Zampini PC pcf; 847688c03ad3SStefano Zampini 8477e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 847888c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 847988c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 848088c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 8481b9be95fcSstefano_zampini 848288c03ad3SStefano Zampini /* hacks */ 848388c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 848472b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 848572b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 848672b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 848772b8c272SStefano Zampini pcisf->n = pcis->n; 848872b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 848988c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 849088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 849188c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 849288c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 849388c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 849488c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 849572b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 849688c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 8497b9be95fcSstefano_zampini 8498b9be95fcSstefano_zampini /* setup constraints so that we can get information on primal vertices and change of basis (in local numbering) */ 849988c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 850072b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 850172b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 850272b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 850372b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 8504b9be95fcSstefano_zampini 850588c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 850672b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 850788c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 850888c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 850988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 851088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 851188c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 851288c03ad3SStefano Zampini pcf->ops->destroy = NULL; 8513b9be95fcSstefano_zampini pcf->ops->reset = NULL; 851488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 851588c03ad3SStefano Zampini } 85169d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 8517111315fdSstefano_zampini 8518111315fdSstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_iP",(PetscObject*)&iP);CHKERRQ(ierr); 8519111315fdSstefano_zampini if (iP) { 8520111315fdSstefano_zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)iP),sub_schurs->prefix,"BDDC sub_schurs options","PC");CHKERRQ(ierr); 8521111315fdSstefano_zampini ierr = PetscOptionsBool("-sub_schurs_discrete_harmonic",NULL,NULL,discrete_harmonic,&discrete_harmonic,NULL);CHKERRQ(ierr); 8522111315fdSstefano_zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 8523111315fdSstefano_zampini } 8524111315fdSstefano_zampini if (discrete_harmonic) { 8525111315fdSstefano_zampini Mat A; 8526111315fdSstefano_zampini ierr = MatDuplicate(pcbddc->local_mat,MAT_COPY_VALUES,&A);CHKERRQ(ierr); 8527111315fdSstefano_zampini ierr = MatZeroRowsColumnsIS(A,iP,1.0,NULL,NULL);CHKERRQ(ierr); 8528111315fdSstefano_zampini ierr = PetscObjectCompose((PetscObject)A,"__KSPFETIDP_iP",(PetscObject)iP);CHKERRQ(ierr); 8529111315fdSstefano_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); 8530111315fdSstefano_zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 8531111315fdSstefano_zampini } else { 853291af6908SStefano 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); 8533111315fdSstefano_zampini } 853472b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 853572b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 8536ca92afb2SStefano Zampini } 8537d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8538b96c3477SStefano Zampini 8539b96c3477SStefano Zampini /* free adjacency */ 8540b96c3477SStefano Zampini if (free_used_adj) { 8541b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 8542b96c3477SStefano Zampini } 8543b96c3477SStefano Zampini PetscFunctionReturn(0); 8544b96c3477SStefano Zampini } 8545b96c3477SStefano Zampini 854608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 8547b96c3477SStefano Zampini { 8548b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8549b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8550b96c3477SStefano Zampini PCBDDCGraph graph; 8551b96c3477SStefano Zampini PetscErrorCode ierr; 8552b96c3477SStefano Zampini 8553b96c3477SStefano Zampini PetscFunctionBegin; 8554b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 855508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 85563301b35fSStefano Zampini IS verticesIS,verticescomm; 85573301b35fSStefano Zampini PetscInt vsize,*idxs; 8558b96c3477SStefano Zampini 8559b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 85603301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 85613301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 85623301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 85633301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 8564c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 8565b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 8566be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 8567441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 85683301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 8569b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 8570b96c3477SStefano Zampini } else { 8571b96c3477SStefano Zampini graph = pcbddc->mat_graph; 8572b96c3477SStefano Zampini } 8573e4d548c7SStefano Zampini /* print some info */ 85745c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 8575e4d548c7SStefano Zampini IS vertices; 8576e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 8577c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 8578e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8579e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 8580e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8581e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 8582e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 8583e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 8584e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 8585e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8586e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8587c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8588e4d548c7SStefano Zampini } 8589b96c3477SStefano Zampini 8590b96c3477SStefano Zampini /* sub_schurs init */ 8591b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 8592b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 8593b334f244SStefano Zampini } 85948b6046baSStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 8595e62b6521Sstefano_zampini pcbddc->sub_schurs->prefix = ((PetscObject)pc)->prefix; 8596a64f4aa4SStefano Zampini 8597b96c3477SStefano Zampini /* free graph struct */ 859808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 8599b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 8600b96c3477SStefano Zampini } 8601b96c3477SStefano Zampini PetscFunctionReturn(0); 8602b96c3477SStefano Zampini } 8603fa34dd3eSStefano Zampini 8604fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 8605fa34dd3eSStefano Zampini { 8606fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8607fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8608fa34dd3eSStefano Zampini PetscErrorCode ierr; 8609fa34dd3eSStefano Zampini 8610fa34dd3eSStefano Zampini PetscFunctionBegin; 8611fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 8612fa34dd3eSStefano Zampini IS zerodiag = NULL; 86134f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 8614fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 86154f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 861675c01103SStefano Zampini PetscReal norm; 8617fa34dd3eSStefano Zampini PetscInt i; 8618fa34dd3eSStefano Zampini 8619fa34dd3eSStefano Zampini /* B0 and B0_B */ 8620fa34dd3eSStefano Zampini if (zerodiag) { 8621fa34dd3eSStefano Zampini IS dummy; 8622fa34dd3eSStefano Zampini 86234f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 86247dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 8625fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 8626fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 8627fa34dd3eSStefano Zampini } 8628fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 8629fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 8630fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 8631fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8632fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8633fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8634fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8635fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 8636fa34dd3eSStefano Zampini /* S_j */ 8637fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8638fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 8639fa34dd3eSStefano Zampini 8640fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 8641fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 8642fa34dd3eSStefano Zampini /* continuous in primal space */ 8643fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 8644fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8645fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8646fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 86474f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 86484f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 8649fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8650fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8651fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8652fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8653fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8654fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8655fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 8656fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 8657fa34dd3eSStefano Zampini 8658fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 8659fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 8660fa34dd3eSStefano Zampini /* local with Schur */ 8661fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 8662fa34dd3eSStefano Zampini if (zerodiag) { 8663fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 86644f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 8665fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8666fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 8667fa34dd3eSStefano Zampini } 8668fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 8669fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8670fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8671fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8672fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8673fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 8674fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8675fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8676fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 8677fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8678fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8679fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8680fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8681fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8682fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 8683fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 8684fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8685fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8686fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8687fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8688fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8689fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8690fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 8691fa34dd3eSStefano Zampini if (zerodiag) { 8692fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 8693fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 86944f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 8695fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8696fa34dd3eSStefano Zampini } 8697fa34dd3eSStefano Zampini /* BDDC */ 8698fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 8699fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 8700fa34dd3eSStefano Zampini 8701fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 8702fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 8703fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 8704fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 87054f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 87064f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 8707fa34dd3eSStefano Zampini } 87084f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 8709fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 8710fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 8711fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 8712fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8713fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 8714fa34dd3eSStefano Zampini } 8715fa34dd3eSStefano Zampini PetscFunctionReturn(0); 8716fa34dd3eSStefano Zampini } 87171e0482f5SStefano Zampini 87181e0482f5SStefano Zampini #include <../src/mat/impls/aij/mpi/mpiaij.h> 87191e0482f5SStefano Zampini PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B) 87201e0482f5SStefano Zampini { 87211e0482f5SStefano Zampini Mat At; 87221e0482f5SStefano Zampini IS rows; 87231e0482f5SStefano Zampini PetscInt rst,ren; 87241e0482f5SStefano Zampini PetscErrorCode ierr; 87251e0482f5SStefano Zampini PetscLayout rmap; 87261e0482f5SStefano Zampini 87271e0482f5SStefano Zampini PetscFunctionBegin; 87281e0482f5SStefano Zampini rst = ren = 0; 87291e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 87301e0482f5SStefano Zampini ierr = PetscLayoutCreate(ccomm,&rmap);CHKERRQ(ierr); 87311e0482f5SStefano Zampini ierr = PetscLayoutSetSize(rmap,A->rmap->N);CHKERRQ(ierr); 87321e0482f5SStefano Zampini ierr = PetscLayoutSetBlockSize(rmap,1);CHKERRQ(ierr); 87331e0482f5SStefano Zampini ierr = PetscLayoutSetUp(rmap);CHKERRQ(ierr); 87341e0482f5SStefano Zampini ierr = PetscLayoutGetRange(rmap,&rst,&ren);CHKERRQ(ierr); 87351e0482f5SStefano Zampini } 8736e07686f2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)A),ren-rst,rst,1,&rows);CHKERRQ(ierr); 87377dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,rows,NULL,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 87381e0482f5SStefano Zampini ierr = ISDestroy(&rows);CHKERRQ(ierr); 87391e0482f5SStefano Zampini 87401e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 87411e0482f5SStefano Zampini Mat_MPIAIJ *a,*b; 87421e0482f5SStefano Zampini IS from,to; 87431e0482f5SStefano Zampini Vec gvec; 87441e0482f5SStefano Zampini PetscInt lsize; 87451e0482f5SStefano Zampini 87461e0482f5SStefano Zampini ierr = MatCreate(ccomm,B);CHKERRQ(ierr); 87471e0482f5SStefano Zampini ierr = MatSetSizes(*B,ren-rst,PETSC_DECIDE,PETSC_DECIDE,At->cmap->N);CHKERRQ(ierr); 87481e0482f5SStefano Zampini ierr = MatSetType(*B,MATAIJ);CHKERRQ(ierr); 87491e0482f5SStefano Zampini ierr = PetscLayoutDestroy(&((*B)->rmap));CHKERRQ(ierr); 87501e0482f5SStefano Zampini ierr = PetscLayoutSetUp((*B)->cmap);CHKERRQ(ierr); 87511e0482f5SStefano Zampini a = (Mat_MPIAIJ*)At->data; 87521e0482f5SStefano Zampini b = (Mat_MPIAIJ*)(*B)->data; 87531e0482f5SStefano Zampini ierr = MPI_Comm_size(ccomm,&b->size);CHKERRQ(ierr); 87541e0482f5SStefano Zampini ierr = MPI_Comm_rank(ccomm,&b->rank);CHKERRQ(ierr); 87551e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->A);CHKERRQ(ierr); 87561e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->B);CHKERRQ(ierr); 87571e0482f5SStefano Zampini b->A = a->A; 87581e0482f5SStefano Zampini b->B = a->B; 87591e0482f5SStefano Zampini 87601e0482f5SStefano Zampini b->donotstash = a->donotstash; 87611e0482f5SStefano Zampini b->roworiented = a->roworiented; 87621e0482f5SStefano Zampini b->rowindices = 0; 87631e0482f5SStefano Zampini b->rowvalues = 0; 87641e0482f5SStefano Zampini b->getrowactive = PETSC_FALSE; 87651e0482f5SStefano Zampini 87661e0482f5SStefano Zampini (*B)->rmap = rmap; 87671e0482f5SStefano Zampini (*B)->factortype = A->factortype; 87681e0482f5SStefano Zampini (*B)->assembled = PETSC_TRUE; 87691e0482f5SStefano Zampini (*B)->insertmode = NOT_SET_VALUES; 87701e0482f5SStefano Zampini (*B)->preallocated = PETSC_TRUE; 87711e0482f5SStefano Zampini 87721e0482f5SStefano Zampini if (a->colmap) { 87731e0482f5SStefano Zampini #if defined(PETSC_USE_CTABLE) 87741e0482f5SStefano Zampini ierr = PetscTableCreateCopy(a->colmap,&b->colmap);CHKERRQ(ierr); 87751e0482f5SStefano Zampini #else 87761e0482f5SStefano Zampini ierr = PetscMalloc1(At->cmap->N,&b->colmap);CHKERRQ(ierr); 87771e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)*B,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 87781e0482f5SStefano Zampini ierr = PetscMemcpy(b->colmap,a->colmap,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 87791e0482f5SStefano Zampini #endif 87801e0482f5SStefano Zampini } else b->colmap = 0; 87811e0482f5SStefano Zampini if (a->garray) { 87821e0482f5SStefano Zampini PetscInt len; 87831e0482f5SStefano Zampini len = a->B->cmap->n; 87841e0482f5SStefano Zampini ierr = PetscMalloc1(len+1,&b->garray);CHKERRQ(ierr); 87851e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)(*B),len*sizeof(PetscInt));CHKERRQ(ierr); 87861e0482f5SStefano Zampini if (len) { ierr = PetscMemcpy(b->garray,a->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 87871e0482f5SStefano Zampini } else b->garray = 0; 87881e0482f5SStefano Zampini 87891e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->lvec);CHKERRQ(ierr); 87901e0482f5SStefano Zampini b->lvec = a->lvec; 87911e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->lvec);CHKERRQ(ierr); 87921e0482f5SStefano Zampini 87931e0482f5SStefano Zampini /* cannot use VecScatterCopy */ 87941e0482f5SStefano Zampini ierr = VecGetLocalSize(b->lvec,&lsize);CHKERRQ(ierr); 87951e0482f5SStefano Zampini ierr = ISCreateGeneral(ccomm,lsize,b->garray,PETSC_USE_POINTER,&from);CHKERRQ(ierr); 87961e0482f5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,lsize,0,1,&to);CHKERRQ(ierr); 87971e0482f5SStefano Zampini ierr = MatCreateVecs(*B,&gvec,NULL);CHKERRQ(ierr); 87981e0482f5SStefano Zampini ierr = VecScatterCreate(gvec,from,b->lvec,to,&b->Mvctx);CHKERRQ(ierr); 87991e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->Mvctx);CHKERRQ(ierr); 88001e0482f5SStefano Zampini ierr = ISDestroy(&from);CHKERRQ(ierr); 88011e0482f5SStefano Zampini ierr = ISDestroy(&to);CHKERRQ(ierr); 88021e0482f5SStefano Zampini ierr = VecDestroy(&gvec);CHKERRQ(ierr); 88031e0482f5SStefano Zampini } 88041e0482f5SStefano Zampini ierr = MatDestroy(&At);CHKERRQ(ierr); 88051e0482f5SStefano Zampini PetscFunctionReturn(0); 88061e0482f5SStefano Zampini } 8807