11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 6674ae819SStefano Zampini 71e0482f5SStefano Zampini static PetscErrorCode MatMPIAIJRestrict(Mat,MPI_Comm,Mat*); 81e0482f5SStefano Zampini 9f498cd09SStefano Zampini /* if range is true, it returns B s.t. span{B} = range(A) 10f498cd09SStefano Zampini if range is false, it returns B s.t. range(B) _|_ range(A) */ 11a13144ffSStefano Zampini #undef __FUNCT__ 12f498cd09SStefano Zampini #define __FUNCT__ "MatDenseOrthogonalRangeOrComplement" 13f498cd09SStefano Zampini PetscErrorCode MatDenseOrthogonalRangeOrComplement(Mat A, PetscBool range, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B) 14a13144ffSStefano Zampini { 15a13144ffSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 16a13144ffSStefano Zampini PetscScalar *uwork,*data,*U, ds = 0.; 17a13144ffSStefano Zampini PetscReal *sing; 18a13144ffSStefano Zampini PetscBLASInt bM,bN,lwork,lierr,di = 1; 19a13144ffSStefano Zampini PetscInt ulw,i,nr,nc,n; 20a13144ffSStefano Zampini PetscErrorCode ierr; 21a13144ffSStefano Zampini 22a13144ffSStefano Zampini PetscFunctionBegin; 23a13144ffSStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 24a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"LAPACK _GESVD not available"); 25614dbb09SStefano Zampini #else 26a13144ffSStefano Zampini ierr = MatGetSize(A,&nr,&nc);CHKERRQ(ierr); 27a13144ffSStefano Zampini if (!nr || !nc) PetscFunctionReturn(0); 28a13144ffSStefano Zampini 29a13144ffSStefano Zampini /* workspace */ 30a13144ffSStefano Zampini if (!work) { 31a13144ffSStefano Zampini ulw = PetscMax(PetscMax(1,5*PetscMin(nr,nc)),3*PetscMin(nr,nc)+PetscMax(nr,nc)); 32f913dca9SStefano Zampini ierr = PetscMalloc1(ulw,&uwork);CHKERRQ(ierr); 33a13144ffSStefano Zampini } else { 34a13144ffSStefano Zampini ulw = lw; 35a13144ffSStefano Zampini uwork = work; 36a13144ffSStefano Zampini } 37a13144ffSStefano Zampini n = PetscMin(nr,nc); 38a13144ffSStefano Zampini if (!rwork) { 39a13144ffSStefano Zampini ierr = PetscMalloc1(n,&sing);CHKERRQ(ierr); 40a13144ffSStefano Zampini } else { 41a13144ffSStefano Zampini sing = rwork; 42a13144ffSStefano Zampini } 43a13144ffSStefano Zampini 44a13144ffSStefano Zampini /* SVD */ 45a13144ffSStefano Zampini ierr = PetscMalloc1(nr*nr,&U);CHKERRQ(ierr); 46a13144ffSStefano Zampini ierr = PetscBLASIntCast(nr,&bM);CHKERRQ(ierr); 47a13144ffSStefano Zampini ierr = PetscBLASIntCast(nc,&bN);CHKERRQ(ierr); 48a13144ffSStefano Zampini ierr = PetscBLASIntCast(ulw,&lwork);CHKERRQ(ierr); 49a13144ffSStefano Zampini ierr = MatDenseGetArray(A,&data);CHKERRQ(ierr); 50a13144ffSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 51a13144ffSStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,&lierr)); 52a13144ffSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 53a13144ffSStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 54a13144ffSStefano Zampini ierr = MatDenseRestoreArray(A,&data);CHKERRQ(ierr); 55a13144ffSStefano Zampini for (i=0;i<n;i++) if (sing[i] < PETSC_SMALL) break; 56a13144ffSStefano Zampini if (!rwork) { 57a13144ffSStefano Zampini ierr = PetscFree(sing);CHKERRQ(ierr); 58a13144ffSStefano Zampini } 59a13144ffSStefano Zampini if (!work) { 60a13144ffSStefano Zampini ierr = PetscFree(uwork);CHKERRQ(ierr); 61a13144ffSStefano Zampini } 62a13144ffSStefano Zampini /* create B */ 63f498cd09SStefano Zampini if (!range) { 64a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,nr-i,NULL,B);CHKERRQ(ierr); 65a13144ffSStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 66a13144ffSStefano Zampini ierr = PetscMemcpy(data,U+nr*i,(nr-i)*nr*sizeof(PetscScalar));CHKERRQ(ierr); 67f498cd09SStefano Zampini } else { 68f498cd09SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,i,NULL,B);CHKERRQ(ierr); 69f498cd09SStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 70f498cd09SStefano Zampini ierr = PetscMemcpy(data,U,i*nr*sizeof(PetscScalar));CHKERRQ(ierr); 71f498cd09SStefano Zampini } 72a13144ffSStefano Zampini ierr = MatDenseRestoreArray(*B,&data);CHKERRQ(ierr); 73a13144ffSStefano Zampini ierr = PetscFree(U);CHKERRQ(ierr); 74614dbb09SStefano Zampini #endif 75614dbb09SStefano Zampini #else /* PETSC_USE_COMPLEX */ 76a13144ffSStefano Zampini PetscFunctionBegin; 77a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for complexes"); 78a13144ffSStefano Zampini #endif 79a13144ffSStefano Zampini PetscFunctionReturn(0); 80a13144ffSStefano Zampini } 81a13144ffSStefano Zampini 821e0482f5SStefano Zampini /* TODO REMOVE */ 831e0482f5SStefano Zampini #if defined(PRINT_GDET) 841e0482f5SStefano Zampini static int inc = 0; 851e0482f5SStefano Zampini static int lev = 0; 861e0482f5SStefano Zampini #endif 871e0482f5SStefano Zampini 88a13144ffSStefano Zampini #undef __FUNCT__ 89a13144ffSStefano Zampini #define __FUNCT__ "PCBDDCComputeNedelecChangeEdge" 901e0482f5SStefano Zampini PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, IS corners, Mat* Gins, Mat* GKins, PetscScalar cvals[2], PetscScalar *work, PetscReal *rwork) 91a13144ffSStefano Zampini { 92a13144ffSStefano Zampini PetscErrorCode ierr; 93a13144ffSStefano Zampini Mat GE,GEd; 94a13144ffSStefano Zampini PetscInt rsize,csize,esize; 95a13144ffSStefano Zampini PetscScalar *ptr; 96a13144ffSStefano Zampini 97a13144ffSStefano Zampini PetscFunctionBegin; 98a13144ffSStefano Zampini ierr = ISGetSize(edge,&esize);CHKERRQ(ierr); 99c3c0e390SStefano Zampini if (!esize) PetscFunctionReturn(0); 100a13144ffSStefano Zampini ierr = ISGetSize(extrow,&rsize);CHKERRQ(ierr); 101a13144ffSStefano Zampini ierr = ISGetSize(extcol,&csize);CHKERRQ(ierr); 102a13144ffSStefano Zampini 103a13144ffSStefano Zampini /* gradients */ 104a13144ffSStefano Zampini ptr = work + 5*esize; 105a13144ffSStefano Zampini ierr = MatGetSubMatrix(lG,extrow,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 106a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,rsize,csize,ptr,Gins);CHKERRQ(ierr); 107a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,Gins);CHKERRQ(ierr); 108a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 109a13144ffSStefano Zampini 110a13144ffSStefano Zampini /* constants */ 111a13144ffSStefano Zampini ptr += rsize*csize; 112a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,esize,csize,ptr,&GEd);CHKERRQ(ierr); 113a13144ffSStefano Zampini ierr = MatGetSubMatrix(lG,edge,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 114a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,&GEd);CHKERRQ(ierr); 115a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 116f498cd09SStefano Zampini ierr = MatDenseOrthogonalRangeOrComplement(GEd,PETSC_FALSE,5*esize,work,rwork,GKins);CHKERRQ(ierr); 117a13144ffSStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1181e0482f5SStefano Zampini 1191e0482f5SStefano Zampini if (corners) { 1201e0482f5SStefano Zampini Mat GEc; 1211e0482f5SStefano Zampini PetscScalar *vals,v; 1221e0482f5SStefano Zampini 1231e0482f5SStefano Zampini ierr = MatGetSubMatrix(lG,edge,corners,MAT_INITIAL_MATRIX,&GEc);CHKERRQ(ierr); 1241e0482f5SStefano Zampini ierr = MatTransposeMatMult(GEc,*GKins,MAT_INITIAL_MATRIX,1.0,&GEd);CHKERRQ(ierr); 1251e0482f5SStefano Zampini ierr = MatDenseGetArray(GEd,&vals);CHKERRQ(ierr); 126637e8532SStefano Zampini /* v = PetscAbsScalar(vals[0]) */; 127637e8532SStefano Zampini v = 1.; 1281e0482f5SStefano Zampini cvals[0] = vals[0]/v; 1291e0482f5SStefano Zampini cvals[1] = vals[1]/v; 1301e0482f5SStefano Zampini ierr = MatDenseRestoreArray(GEd,&vals);CHKERRQ(ierr); 1311e0482f5SStefano Zampini ierr = MatScale(*GKins,1./v);CHKERRQ(ierr); 1321e0482f5SStefano Zampini #if defined(PRINT_GDET) 1331e0482f5SStefano Zampini { 1341e0482f5SStefano Zampini PetscViewer viewer; 1351e0482f5SStefano Zampini char filename[256]; 1361e0482f5SStefano Zampini sprintf(filename,"Gdet_l%d_r%d_cc%d.m",lev,PetscGlobalRank,inc++); 1371e0482f5SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1381e0482f5SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1391e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEc,"GEc");CHKERRQ(ierr); 1401e0482f5SStefano Zampini ierr = MatView(GEc,viewer);CHKERRQ(ierr); 1411e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)(*GKins),"GK");CHKERRQ(ierr); 1421e0482f5SStefano Zampini ierr = MatView(*GKins,viewer);CHKERRQ(ierr); 1431e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEd,"Gproj");CHKERRQ(ierr); 1441e0482f5SStefano Zampini ierr = MatView(GEd,viewer);CHKERRQ(ierr); 1451e0482f5SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1461e0482f5SStefano Zampini } 1471e0482f5SStefano Zampini #endif 1481e0482f5SStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1491e0482f5SStefano Zampini ierr = MatDestroy(&GEc);CHKERRQ(ierr); 1501e0482f5SStefano Zampini } 1511e0482f5SStefano Zampini 152a13144ffSStefano Zampini PetscFunctionReturn(0); 153a13144ffSStefano Zampini } 154a13144ffSStefano Zampini 155a13144ffSStefano Zampini #undef __FUNCT__ 156a13144ffSStefano Zampini #define __FUNCT__ "PCBDDCNedelecSupport" 157a13144ffSStefano Zampini PetscErrorCode PCBDDCNedelecSupport(PC pc) 158a13144ffSStefano Zampini { 159a13144ffSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 160a13144ffSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1610569b399SStefano Zampini Mat G,T,conn,lG,lGt,lGis,lGall,lGe,lGinit; 162eee23b56SStefano Zampini Vec tvec; 163a13144ffSStefano Zampini PetscSF sfv; 1641e0482f5SStefano Zampini ISLocalToGlobalMapping el2g,vl2g,fl2g,al2g; 165a13144ffSStefano Zampini MPI_Comm comm; 166c2151214SStefano Zampini IS lned,primals,allprimals,nedfieldlocal; 167c2151214SStefano Zampini IS *eedges,*extrows,*extcols,*alleedges; 1687d871cd7SStefano Zampini PetscBT btv,bte,btvc,btb,btbd,btvcand,btvi,btee,bter; 169a13144ffSStefano Zampini PetscScalar *vals,*work; 170a13144ffSStefano Zampini PetscReal *rwork; 171a13144ffSStefano Zampini const PetscInt *idxs,*ii,*jj,*iit,*jjt; 1721e0482f5SStefano Zampini PetscInt ne,nv,Lv,order,n,field; 173a13144ffSStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 174eee23b56SStefano Zampini PetscInt i,j,extmem,cum,maxsize,nee; 175b03ebc13SStefano Zampini PetscInt *extrow,*extrowcum,*marks,*vmarks,*gidxs; 176a13144ffSStefano Zampini PetscInt *sfvleaves,*sfvroots; 177b03ebc13SStefano Zampini PetscInt *corners,*cedges; 178637e8532SStefano Zampini PetscInt *ecount,**eneighs,*vcount,**vneighs; 179b03ebc13SStefano Zampini #if defined(PETSC_USE_DEBUG) 180b03ebc13SStefano Zampini PetscInt *emarks; 181b03ebc13SStefano Zampini #endif 182213b8bfaSStefano Zampini PetscBool print,eerr,done,lrc[2],conforming,global,singular,setprimal; 183a13144ffSStefano Zampini PetscErrorCode ierr; 184a13144ffSStefano Zampini 185a13144ffSStefano Zampini PetscFunctionBegin; 186213b8bfaSStefano Zampini /* If the discrete gradient is defined for a subset of dofs and global is true, 187213b8bfaSStefano Zampini it assumes G is given in global ordering for all the dofs. 188213b8bfaSStefano Zampini Otherwise, the ordering is global for the Nedelec field */ 189213b8bfaSStefano Zampini order = pcbddc->nedorder; 190213b8bfaSStefano Zampini conforming = pcbddc->conforming; 191213b8bfaSStefano Zampini field = pcbddc->nedfield; 192213b8bfaSStefano Zampini global = pcbddc->nedglobal; 193213b8bfaSStefano Zampini setprimal = PETSC_FALSE; 194a13144ffSStefano Zampini print = PETSC_FALSE; 195213b8bfaSStefano Zampini singular = PETSC_FALSE; 196a13144ffSStefano Zampini 197213b8bfaSStefano Zampini /* Command line customization */ 198213b8bfaSStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC Nedelec options","PC");CHKERRQ(ierr); 199213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_field_primal","All edge dofs set as primals: Toselli's algorithm C",NULL,setprimal,&setprimal,NULL);CHKERRQ(ierr); 200213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_singular","Infer nullspace from discrete gradient",NULL,singular,&singular,NULL);CHKERRQ(ierr); 201213b8bfaSStefano Zampini ierr = PetscOptionsInt ("-pc_bddc_nedelec_order","Test variable order code (to be removed)",NULL,order,&order,NULL);CHKERRQ(ierr); 202213b8bfaSStefano Zampini /* print debug info TODO: to be removed */ 203213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_print","Print debug info",NULL,print,&print,NULL);CHKERRQ(ierr); 204213b8bfaSStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 205213b8bfaSStefano Zampini 206213b8bfaSStefano Zampini /* Return if there are no edges in the decomposition and the problem is not singular */ 2071e0482f5SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&al2g,NULL);CHKERRQ(ierr); 2081e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(al2g,&n);CHKERRQ(ierr); 209213b8bfaSStefano Zampini ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 210213b8bfaSStefano Zampini if (!singular) { 211a13144ffSStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 212a13144ffSStefano Zampini lrc[0] = PETSC_FALSE; 213c2151214SStefano Zampini for (i=0;i<n;i++) { 214a13144ffSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 215a13144ffSStefano Zampini lrc[0] = PETSC_TRUE; 216a13144ffSStefano Zampini break; 217a13144ffSStefano Zampini } 218a13144ffSStefano Zampini } 219a13144ffSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 220a13144ffSStefano Zampini ierr = MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 221a13144ffSStefano Zampini if (!lrc[1]) PetscFunctionReturn(0); 222213b8bfaSStefano Zampini } 223a13144ffSStefano Zampini 224213b8bfaSStefano Zampini /* Get Nedelec field */ 225213b8bfaSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 226c2151214SStefano 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); 227213b8bfaSStefano Zampini if (pcbddc->n_ISForDofsLocal && field >= 0) { 228c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);CHKERRQ(ierr); 229c2151214SStefano Zampini nedfieldlocal = pcbddc->ISForDofsLocal[field]; 230c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 231213b8bfaSStefano Zampini } else if (!pcbddc->n_ISForDofsLocal && field != PETSC_DECIDE) { 232213b8bfaSStefano Zampini ne = n; 233213b8bfaSStefano Zampini nedfieldlocal = NULL; 234213b8bfaSStefano Zampini global = PETSC_TRUE; 235213b8bfaSStefano Zampini } else if (field == PETSC_DECIDE) { 236213b8bfaSStefano Zampini PetscInt rst,ren,*idx; 237213b8bfaSStefano Zampini 238213b8bfaSStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 239213b8bfaSStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 240213b8bfaSStefano Zampini ierr = MatGetOwnershipRange(pcbddc->discretegradient,&rst,&ren);CHKERRQ(ierr); 241213b8bfaSStefano Zampini for (i=rst;i<ren;i++) { 242213b8bfaSStefano Zampini PetscInt nc; 243213b8bfaSStefano Zampini 244213b8bfaSStefano Zampini ierr = MatGetRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 245213b8bfaSStefano Zampini if (nc > 1) matis->sf_rootdata[i-rst] = 1; 246213b8bfaSStefano Zampini ierr = MatRestoreRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 247213b8bfaSStefano Zampini } 248213b8bfaSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 249213b8bfaSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 250213b8bfaSStefano Zampini ierr = PetscMalloc1(n,&idx);CHKERRQ(ierr); 251213b8bfaSStefano Zampini for (i=0,ne=0;i<n;i++) if (matis->sf_leafdata[i]) idx[ne++] = i; 252213b8bfaSStefano Zampini ierr = ISCreateGeneral(comm,ne,idx,PETSC_OWN_POINTER,&nedfieldlocal);CHKERRQ(ierr); 253213b8bfaSStefano Zampini } else { 254213b8bfaSStefano Zampini SETERRQ(comm,PETSC_ERR_USER,"When multiple fields are present, the Nedelec field has to be specified"); 255213b8bfaSStefano Zampini } 256213b8bfaSStefano Zampini 257213b8bfaSStefano Zampini /* Sanity checks */ 258213b8bfaSStefano Zampini if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time"); 259213b8bfaSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis"); 260213b8bfaSStefano 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); 261213b8bfaSStefano Zampini 262213b8bfaSStefano Zampini /* Just set primal dofs and return */ 2631e0482f5SStefano Zampini if (setprimal) { 264eee23b56SStefano Zampini IS enedfieldlocal; 265eee23b56SStefano Zampini PetscInt *eidxs; 266eee23b56SStefano Zampini 267eee23b56SStefano Zampini ierr = PetscMalloc1(ne,&eidxs);CHKERRQ(ierr); 268eee23b56SStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 269213b8bfaSStefano Zampini if (nedfieldlocal) { 270213b8bfaSStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 271eee23b56SStefano Zampini for (i=0,cum=0;i<ne;i++) { 272eee23b56SStefano Zampini if (PetscRealPart(vals[idxs[i]]) > 2.) { 273eee23b56SStefano Zampini eidxs[cum++] = idxs[i]; 274eee23b56SStefano Zampini } 275eee23b56SStefano Zampini } 276eee23b56SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 277213b8bfaSStefano Zampini } else { 278213b8bfaSStefano Zampini for (i=0,cum=0;i<ne;i++) { 279213b8bfaSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 280213b8bfaSStefano Zampini eidxs[cum++] = i; 281213b8bfaSStefano Zampini } 282213b8bfaSStefano Zampini } 283213b8bfaSStefano Zampini } 284213b8bfaSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 285eee23b56SStefano Zampini ierr = ISCreateGeneral(comm,cum,eidxs,PETSC_COPY_VALUES,&enedfieldlocal);CHKERRQ(ierr); 286eee23b56SStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,enedfieldlocal);CHKERRQ(ierr); 287eee23b56SStefano Zampini ierr = PetscFree(eidxs);CHKERRQ(ierr); 2881e0482f5SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 289eee23b56SStefano Zampini ierr = ISDestroy(&enedfieldlocal);CHKERRQ(ierr); 2901e0482f5SStefano Zampini PetscFunctionReturn(0); 2911e0482f5SStefano Zampini } 292a13144ffSStefano Zampini 293213b8bfaSStefano Zampini /* Compute some l2g maps */ 294213b8bfaSStefano Zampini if (nedfieldlocal) { 295c2151214SStefano Zampini IS is; 296c2151214SStefano Zampini 297c2151214SStefano Zampini /* need to map from the local Nedelec field to local numbering */ 298c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);CHKERRQ(ierr); 2991e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering for the whole dofs*/ 3001e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(al2g,nedfieldlocal,&is);CHKERRQ(ierr); 3011e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is,&al2g);CHKERRQ(ierr); 3021e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering (for Nedelec only) */ 3031e0482f5SStefano Zampini if (global) { 3041e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3051e0482f5SStefano Zampini el2g = al2g; 3061e0482f5SStefano Zampini } else { 3071e0482f5SStefano Zampini IS gis; 3081e0482f5SStefano Zampini 3091e0482f5SStefano Zampini ierr = ISRenumber(is,NULL,NULL,&gis);CHKERRQ(ierr); 3101e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(gis,&el2g);CHKERRQ(ierr); 3111e0482f5SStefano Zampini ierr = ISDestroy(&gis);CHKERRQ(ierr); 3121e0482f5SStefano Zampini } 313c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 314c2151214SStefano Zampini } else { 3151e0482f5SStefano Zampini /* restore default */ 3161e0482f5SStefano Zampini pcbddc->nedfield = -1; 3171e0482f5SStefano Zampini /* one ref for the destruction of al2g, one for el2g */ 3181e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3191e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3201e0482f5SStefano Zampini el2g = al2g; 321c2151214SStefano Zampini fl2g = NULL; 322c2151214SStefano Zampini } 323a13144ffSStefano Zampini 324213b8bfaSStefano Zampini /* Start communication to drop connections for interior edges (for cc analysis only) */ 325c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 326c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 327c2151214SStefano Zampini if (nedfieldlocal) { 328c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 329c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1; 330c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 331c2151214SStefano Zampini } else { 332c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1; 333c2151214SStefano Zampini } 334c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 335c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 336213b8bfaSStefano Zampini 337213b8bfaSStefano Zampini if (!singular) { /* drop connections with interior edges to avoid unneeded communications and memory movements */ 338213b8bfaSStefano Zampini ierr = MatDuplicate(pcbddc->discretegradient,MAT_COPY_VALUES,&G);CHKERRQ(ierr); 339213b8bfaSStefano Zampini ierr = MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 3401e0482f5SStefano Zampini if (global) { 3411e0482f5SStefano Zampini PetscInt rst; 3421e0482f5SStefano Zampini 343c2151214SStefano Zampini ierr = MatGetOwnershipRange(G,&rst,NULL);CHKERRQ(ierr); 344c2151214SStefano Zampini for (i=0,cum=0;i<pc->pmat->rmap->n;i++) { 345c2151214SStefano Zampini if (matis->sf_rootdata[i] < 2) { 346c2151214SStefano Zampini matis->sf_rootdata[cum++] = i + rst; 347c2151214SStefano Zampini } 348c2151214SStefano Zampini } 349a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);CHKERRQ(ierr); 350c2151214SStefano Zampini ierr = MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);CHKERRQ(ierr); 3511e0482f5SStefano Zampini } else { 3521e0482f5SStefano Zampini PetscInt *tbz; 3531e0482f5SStefano Zampini 3541e0482f5SStefano Zampini ierr = PetscMalloc1(ne,&tbz);CHKERRQ(ierr); 3551e0482f5SStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 3561e0482f5SStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 3571e0482f5SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3581e0482f5SStefano Zampini for (i=0,cum=0;i<ne;i++) 3591e0482f5SStefano Zampini if (matis->sf_leafdata[idxs[i]] == 1) 3601e0482f5SStefano Zampini tbz[cum++] = i; 3611e0482f5SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3621e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApply(el2g,cum,tbz,tbz);CHKERRQ(ierr); 3631e0482f5SStefano Zampini ierr = MatZeroRows(G,cum,tbz,0.,NULL,NULL);CHKERRQ(ierr); 3641e0482f5SStefano Zampini ierr = PetscFree(tbz);CHKERRQ(ierr); 3651e0482f5SStefano Zampini } 366213b8bfaSStefano Zampini } else { /* we need the entire G to infer the nullspace */ 367213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->discretegradient);CHKERRQ(ierr); 368213b8bfaSStefano Zampini G = pcbddc->discretegradient; 369213b8bfaSStefano Zampini } 370a13144ffSStefano Zampini 371a13144ffSStefano Zampini /* Extract subdomain relevant rows of G */ 372a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(el2g,&idxs);CHKERRQ(ierr); 373a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);CHKERRQ(ierr); 374a13144ffSStefano Zampini ierr = MatGetSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);CHKERRQ(ierr); 375a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);CHKERRQ(ierr); 376a13144ffSStefano Zampini ierr = ISDestroy(&lned);CHKERRQ(ierr); 377a13144ffSStefano Zampini ierr = MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);CHKERRQ(ierr); 378a13144ffSStefano Zampini ierr = MatDestroy(&lGall);CHKERRQ(ierr); 379a13144ffSStefano Zampini ierr = MatISGetLocalMat(lGis,&lG);CHKERRQ(ierr); 380a13144ffSStefano Zampini 381213b8bfaSStefano Zampini /* SF for nodal dofs communications */ 382c2151214SStefano Zampini ierr = MatGetLocalSize(G,NULL,&Lv);CHKERRQ(ierr); 383a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);CHKERRQ(ierr); 384a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)vl2g);CHKERRQ(ierr); 385a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(vl2g,&nv);CHKERRQ(ierr); 386a13144ffSStefano Zampini ierr = PetscSFCreate(comm,&sfv);CHKERRQ(ierr); 387a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(vl2g,&idxs);CHKERRQ(ierr); 388a13144ffSStefano Zampini ierr = PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);CHKERRQ(ierr); 389a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);CHKERRQ(ierr); 390213b8bfaSStefano Zampini i = singular ? 2 : 1; 391213b8bfaSStefano Zampini ierr = PetscMalloc2(i*nv,&sfvleaves,i*Lv,&sfvroots);CHKERRQ(ierr); 392a13144ffSStefano Zampini 3931e0482f5SStefano Zampini /* Destroy temporary G created in MATIS format and modified G */ 394213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)lG);CHKERRQ(ierr); 395a13144ffSStefano Zampini ierr = MatDestroy(&lGis);CHKERRQ(ierr); 396213b8bfaSStefano Zampini ierr = MatDestroy(&G);CHKERRQ(ierr); 397a13144ffSStefano Zampini 398213b8bfaSStefano Zampini if (print) { 399213b8bfaSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"initial_lG");CHKERRQ(ierr); 400213b8bfaSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 401213b8bfaSStefano Zampini } 402213b8bfaSStefano Zampini 403213b8bfaSStefano Zampini /* Save lG for values insertion in change of basis */ 4040569b399SStefano Zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);CHKERRQ(ierr); 4050569b399SStefano Zampini 406a13144ffSStefano Zampini /* Analyze the edge-nodes connections (duplicate lG) */ 4074e64d54eSstefano_zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGe);CHKERRQ(ierr); 4084e64d54eSstefano_zampini ierr = MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 409a13144ffSStefano Zampini ierr = PetscBTCreate(nv,&btv);CHKERRQ(ierr); 410a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&bte);CHKERRQ(ierr); 411a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&btb);CHKERRQ(ierr); 4127d871cd7SStefano Zampini ierr = PetscBTCreate(ne,&btbd);CHKERRQ(ierr); 413c2151214SStefano Zampini ierr = PetscBTCreate(nv,&btvcand);CHKERRQ(ierr); 414a13144ffSStefano Zampini /* need to import the boundary specification to ensure the 415a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 416a13144ffSStefano Zampini if (pcbddc->DirichletBoundariesLocal) { 417c2151214SStefano Zampini IS is; 418c2151214SStefano Zampini 419c2151214SStefano Zampini if (fl2g) { 420c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);CHKERRQ(ierr); 421c2151214SStefano Zampini } else { 422c2151214SStefano Zampini is = pcbddc->DirichletBoundariesLocal; 423c2151214SStefano Zampini } 424c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 425c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 426a13144ffSStefano Zampini for (i=0;i<cum;i++) { 427a13144ffSStefano Zampini if (idxs[i] >= 0) { 428a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 4297d871cd7SStefano Zampini ierr = PetscBTSet(btbd,idxs[i]);CHKERRQ(ierr); 430a13144ffSStefano Zampini } 431a13144ffSStefano Zampini } 432c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 433c2151214SStefano Zampini if (fl2g) { 434c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 435c2151214SStefano Zampini } 436a13144ffSStefano Zampini } 437a13144ffSStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 438c2151214SStefano Zampini IS is; 439c2151214SStefano Zampini 440c2151214SStefano Zampini if (fl2g) { 441c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);CHKERRQ(ierr); 442c2151214SStefano Zampini } else { 443c2151214SStefano Zampini is = pcbddc->NeumannBoundariesLocal; 444c2151214SStefano Zampini } 445c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 446c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 447a13144ffSStefano Zampini for (i=0;i<cum;i++) { 448a13144ffSStefano Zampini if (idxs[i] >= 0) { 449a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 450a13144ffSStefano Zampini } 451a13144ffSStefano Zampini } 452c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 453c2151214SStefano Zampini if (fl2g) { 454c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 455a13144ffSStefano Zampini } 456c2151214SStefano Zampini } 457c2151214SStefano Zampini 458213b8bfaSStefano Zampini /* Count neighs per dof */ 459637e8532SStefano Zampini ierr = PetscCalloc1(ne,&ecount);CHKERRQ(ierr); 460637e8532SStefano Zampini ierr = PetscMalloc1(ne,&eneighs);CHKERRQ(ierr); 461637e8532SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 462637e8532SStefano Zampini for (i=1,cum=0;i<n_neigh;i++) { 463637e8532SStefano Zampini cum += n_shared[i]; 464637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 465637e8532SStefano Zampini ecount[shared[i][j]]++; 466637e8532SStefano Zampini } 467637e8532SStefano Zampini } 468637e8532SStefano Zampini if (ne) { 469637e8532SStefano Zampini ierr = PetscMalloc1(cum,&eneighs[0]);CHKERRQ(ierr); 470637e8532SStefano Zampini } 471637e8532SStefano Zampini for (i=1;i<ne;i++) eneighs[i] = eneighs[i-1] + ecount[i-1]; 472637e8532SStefano Zampini ierr = PetscMemzero(ecount,ne*sizeof(PetscInt));CHKERRQ(ierr); 473637e8532SStefano Zampini for (i=1;i<n_neigh;i++) { 474637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 475637e8532SStefano Zampini PetscInt k = shared[i][j]; 476637e8532SStefano Zampini eneighs[k][ecount[k]] = neigh[i]; 477637e8532SStefano Zampini ecount[k]++; 478637e8532SStefano Zampini } 479637e8532SStefano Zampini } 480637e8532SStefano Zampini for (i=0;i<ne;i++) { 481637e8532SStefano Zampini ierr = PetscSortRemoveDupsInt(&ecount[i],eneighs[i]);CHKERRQ(ierr); 482637e8532SStefano Zampini } 483637e8532SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 484637e8532SStefano Zampini ierr = PetscCalloc1(nv,&vcount);CHKERRQ(ierr); 485637e8532SStefano Zampini ierr = PetscMalloc1(nv,&vneighs);CHKERRQ(ierr); 486637e8532SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 487637e8532SStefano Zampini for (i=1,cum=0;i<n_neigh;i++) { 488637e8532SStefano Zampini cum += n_shared[i]; 489637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 490637e8532SStefano Zampini vcount[shared[i][j]]++; 491637e8532SStefano Zampini } 492637e8532SStefano Zampini } 493637e8532SStefano Zampini if (nv) { 494637e8532SStefano Zampini ierr = PetscMalloc1(cum,&vneighs[0]);CHKERRQ(ierr); 495637e8532SStefano Zampini } 496637e8532SStefano Zampini for (i=1;i<nv;i++) vneighs[i] = vneighs[i-1] + vcount[i-1]; 497637e8532SStefano Zampini ierr = PetscMemzero(vcount,nv*sizeof(PetscInt));CHKERRQ(ierr); 498637e8532SStefano Zampini for (i=1;i<n_neigh;i++) { 499637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 500637e8532SStefano Zampini PetscInt k = shared[i][j]; 501637e8532SStefano Zampini vneighs[k][vcount[k]] = neigh[i]; 502637e8532SStefano Zampini vcount[k]++; 503637e8532SStefano Zampini } 504637e8532SStefano Zampini } 505637e8532SStefano Zampini for (i=0;i<nv;i++) { 506637e8532SStefano Zampini ierr = PetscSortRemoveDupsInt(&vcount[i],vneighs[i]);CHKERRQ(ierr); 507637e8532SStefano Zampini } 508637e8532SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 509637e8532SStefano Zampini 5107d871cd7SStefano Zampini /* need to remove coarse faces' dofs and coarse edges' dirichlet dofs 5117d871cd7SStefano Zampini for proper detection of coarse edges' endpoints */ 51262b0c6f7SStefano Zampini ierr = PetscBTCreate(ne,&btee);CHKERRQ(ierr); 51362b0c6f7SStefano Zampini for (i=0;i<ne;i++) { 5147d871cd7SStefano Zampini if ((ecount[i] > 1 && !PetscBTLookup(btbd,i)) || (ecount[i] == 1 && PetscBTLookup(btb,i))) { 51562b0c6f7SStefano Zampini ierr = PetscBTSet(btee,i);CHKERRQ(ierr); 51662b0c6f7SStefano Zampini } 51762b0c6f7SStefano Zampini } 518637e8532SStefano Zampini ierr = PetscMalloc1(ne,&marks);CHKERRQ(ierr); 51962b0c6f7SStefano Zampini if (!conforming) { 52062b0c6f7SStefano Zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 52162b0c6f7SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 52262b0c6f7SStefano Zampini } 5234e64d54eSstefano_zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 524dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGe,&vals);CHKERRQ(ierr); 52562b0c6f7SStefano Zampini cum = 0; 526a13144ffSStefano Zampini for (i=0;i<ne;i++) { 527dec27d64SStefano Zampini /* eliminate rows corresponding to edge dofs belonging to coarse faces */ 52862b0c6f7SStefano Zampini if (!PetscBTLookup(btee,i)) { 529a13144ffSStefano Zampini marks[cum++] = i; 530dec27d64SStefano Zampini continue; 531dec27d64SStefano Zampini } 532dec27d64SStefano Zampini /* set badly connected edge dofs as primal */ 53362b0c6f7SStefano Zampini if (!conforming) { 53462b0c6f7SStefano Zampini if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */ 535a13144ffSStefano Zampini marks[cum++] = i; 536a13144ffSStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 537a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 538a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 539a13144ffSStefano Zampini } 54062b0c6f7SStefano Zampini } else { 54162b0c6f7SStefano Zampini /* every edge dofs should be connected trough a certain number of nodal dofs 54262b0c6f7SStefano Zampini to other edge dofs belonging to coarse edges 54362b0c6f7SStefano Zampini - at most 2 endpoints 54462b0c6f7SStefano Zampini - order-1 interior nodal dofs 54562b0c6f7SStefano Zampini - no undefined nodal dofs (nconn < order) 54662b0c6f7SStefano Zampini */ 54762b0c6f7SStefano Zampini PetscInt ends = 0,ints = 0, undef = 0; 54862b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 54962b0c6f7SStefano Zampini PetscInt v = jj[j],k; 55062b0c6f7SStefano Zampini PetscInt nconn = iit[v+1]-iit[v]; 55162b0c6f7SStefano Zampini for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--; 55262b0c6f7SStefano Zampini if (nconn > order) ends++; 55362b0c6f7SStefano Zampini else if (nconn == order) ints++; 55462b0c6f7SStefano Zampini else undef++; 55562b0c6f7SStefano Zampini } 55662b0c6f7SStefano Zampini if (undef || ends > 2 || ints != order -1) { 55762b0c6f7SStefano Zampini marks[cum++] = i; 55862b0c6f7SStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 55962b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 56062b0c6f7SStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 56162b0c6f7SStefano Zampini } 56262b0c6f7SStefano Zampini } 56362b0c6f7SStefano Zampini } 564a13144ffSStefano Zampini } 565dec27d64SStefano Zampini /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */ 566dec27d64SStefano Zampini if (!order && ii[i+1] != ii[i]) { 567dec27d64SStefano Zampini PetscScalar val = 1./(ii[i+1]-ii[i]-1); 568dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vals[j] = val; 569a13144ffSStefano Zampini } 570dec27d64SStefano Zampini } 57162b0c6f7SStefano Zampini ierr = PetscBTDestroy(&btee);CHKERRQ(ierr); 572dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGe,&vals);CHKERRQ(ierr); 5734e64d54eSstefano_zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 57462b0c6f7SStefano Zampini if (!conforming) { 57562b0c6f7SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 57662b0c6f7SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 57762b0c6f7SStefano Zampini } 5784e64d54eSstefano_zampini ierr = MatZeroRows(lGe,cum,marks,0.,NULL,NULL);CHKERRQ(ierr); 579637e8532SStefano Zampini 580b03ebc13SStefano Zampini /* identify splitpoints and corner candidates */ 5814e64d54eSstefano_zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 582a13144ffSStefano Zampini if (print) { 5834e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");CHKERRQ(ierr); 5844e64d54eSstefano_zampini ierr = MatView(lGe,NULL);CHKERRQ(ierr); 5854e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");CHKERRQ(ierr); 586a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 587a13144ffSStefano Zampini } 588a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 589dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGt,&vals);CHKERRQ(ierr); 590a13144ffSStefano Zampini for (i=0;i<nv;i++) { 591637e8532SStefano Zampini PetscInt ord = order, test = ii[i+1]-ii[i], vc = vcount[i]; 5927d871cd7SStefano Zampini PetscBool sneighs = PETSC_TRUE, bdir = PETSC_FALSE; 593b03ebc13SStefano Zampini if (!order) { /* variable order */ 594dec27d64SStefano Zampini PetscReal vorder = 0.; 595dec27d64SStefano Zampini 596dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]); 597dec27d64SStefano Zampini test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON); 598dec27d64SStefano Zampini if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%d)",vorder,test); 599dec27d64SStefano Zampini ord = 1; 600dec27d64SStefano Zampini } 601a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 602dec27d64SStefano 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); 603a13144ffSStefano Zampini #endif 604637e8532SStefano Zampini for (j=ii[i];j<ii[i+1] && sneighs;j++) { 6057d871cd7SStefano Zampini if (PetscBTLookup(btbd,jj[j])) { 6067d871cd7SStefano Zampini bdir = PETSC_TRUE; 6077d871cd7SStefano Zampini break; 6087d871cd7SStefano Zampini } 609637e8532SStefano Zampini if (vc != ecount[jj[j]]) { 610637e8532SStefano Zampini sneighs = PETSC_FALSE; 611637e8532SStefano Zampini } else { 612637e8532SStefano Zampini PetscInt k,*vn = vneighs[i], *en = eneighs[jj[j]]; 613637e8532SStefano Zampini for (k=0;k<vc;k++) { 614637e8532SStefano Zampini if (vn[k] != en[k]) { 615637e8532SStefano Zampini sneighs = PETSC_FALSE; 616637e8532SStefano Zampini break; 617637e8532SStefano Zampini } 618637e8532SStefano Zampini } 619637e8532SStefano Zampini } 620637e8532SStefano Zampini } 6217d871cd7SStefano Zampini if (!sneighs || test >= 3*ord || bdir) { /* splitpoints */ 6227d871cd7SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %d (%d %d %d)\n",i,!sneighs,test >= 3*ord,bdir); 623a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 624dec27d64SStefano Zampini } else if (test == ord) { 625b03ebc13SStefano Zampini if (order == 1 || (!order && ii[i+1]-ii[i] == 1)) { 626a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %d\n",i); 627a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 628a13144ffSStefano Zampini } else { 629a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %d\n",i); 630a13144ffSStefano Zampini ierr = PetscBTSet(btvcand,i);CHKERRQ(ierr); 631a13144ffSStefano Zampini } 632a13144ffSStefano Zampini } 633a13144ffSStefano Zampini } 634637e8532SStefano Zampini ierr = PetscFree(ecount);CHKERRQ(ierr); 635637e8532SStefano Zampini ierr = PetscFree(vcount);CHKERRQ(ierr); 636637e8532SStefano Zampini if (ne) { 637637e8532SStefano Zampini ierr = PetscFree(eneighs[0]);CHKERRQ(ierr); 638637e8532SStefano Zampini } 639637e8532SStefano Zampini if (nv) { 640637e8532SStefano Zampini ierr = PetscFree(vneighs[0]);CHKERRQ(ierr); 641637e8532SStefano Zampini } 642637e8532SStefano Zampini ierr = PetscFree(eneighs);CHKERRQ(ierr); 643637e8532SStefano Zampini ierr = PetscFree(vneighs);CHKERRQ(ierr); 6447d871cd7SStefano Zampini ierr = PetscBTDestroy(&btbd);CHKERRQ(ierr); 645b03ebc13SStefano Zampini 646b03ebc13SStefano Zampini /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */ 647b03ebc13SStefano Zampini if (order != 1) { 648b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"INSPECTING CANDIDATES\n"); 649b03ebc13SStefano Zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 650b03ebc13SStefano Zampini for (i=0;i<nv;i++) { 651b03ebc13SStefano Zampini if (PetscBTLookup(btvcand,i)) { 652b03ebc13SStefano Zampini PetscBool found = PETSC_FALSE; 653b03ebc13SStefano Zampini for (j=ii[i];j<ii[i+1] && !found;j++) { 654b03ebc13SStefano Zampini PetscInt k,e = jj[j]; 655b03ebc13SStefano Zampini if (PetscBTLookup(bte,e)) continue; 656b03ebc13SStefano Zampini for (k=iit[e];k<iit[e+1];k++) { 657b03ebc13SStefano Zampini PetscInt v = jjt[k]; 658b03ebc13SStefano Zampini if (v != i && PetscBTLookup(btvcand,v)) { 659b03ebc13SStefano Zampini found = PETSC_TRUE; 660b03ebc13SStefano Zampini break; 661b03ebc13SStefano Zampini } 662b03ebc13SStefano Zampini } 663b03ebc13SStefano Zampini } 664b03ebc13SStefano Zampini if (!found) { 665b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d CLEARED\n",i); 666b03ebc13SStefano Zampini ierr = PetscBTClear(btvcand,i);CHKERRQ(ierr); 667b03ebc13SStefano Zampini } else { 668b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d ACCEPTED\n",i); 669b03ebc13SStefano Zampini } 670b03ebc13SStefano Zampini } 671b03ebc13SStefano Zampini } 672b03ebc13SStefano Zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 673b03ebc13SStefano Zampini } 674dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGt,&vals);CHKERRQ(ierr); 675a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 676b03ebc13SStefano Zampini ierr = MatDestroy(&lGe);CHKERRQ(ierr); 677a13144ffSStefano Zampini 678a13144ffSStefano Zampini /* Get the local G^T explicitly */ 6790569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 680a13144ffSStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 6814e64d54eSstefano_zampini ierr = MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 682a13144ffSStefano Zampini 6834e64d54eSstefano_zampini /* Mark interior nodal dofs */ 684a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 6854e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvi);CHKERRQ(ierr); 686a13144ffSStefano Zampini for (i=1;i<n_neigh;i++) { 687a13144ffSStefano Zampini for (j=0;j<n_shared[i];j++) { 6884e64d54eSstefano_zampini ierr = PetscBTSet(btvi,shared[i][j]);CHKERRQ(ierr); 689a13144ffSStefano Zampini } 690a13144ffSStefano Zampini } 691a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 692a13144ffSStefano Zampini 693a13144ffSStefano Zampini /* communicate corners and splitpoints */ 694a13144ffSStefano Zampini ierr = PetscMalloc1(nv,&vmarks);CHKERRQ(ierr); 695a13144ffSStefano Zampini ierr = PetscMemzero(sfvleaves,nv*sizeof(PetscInt));CHKERRQ(ierr); 696a13144ffSStefano Zampini ierr = PetscMemzero(sfvroots,Lv*sizeof(PetscInt));CHKERRQ(ierr); 697a13144ffSStefano Zampini for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1; 698a13144ffSStefano Zampini 699a13144ffSStefano Zampini if (print) { 700a13144ffSStefano Zampini IS tbz; 701a13144ffSStefano Zampini 702a13144ffSStefano Zampini cum = 0; 703a13144ffSStefano Zampini for (i=0;i<nv;i++) 704a13144ffSStefano Zampini if (sfvleaves[i]) 705a13144ffSStefano Zampini vmarks[cum++] = i; 706a13144ffSStefano Zampini 707a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 708a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");CHKERRQ(ierr); 709a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 710a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 711a13144ffSStefano Zampini } 712a13144ffSStefano Zampini 713a13144ffSStefano Zampini ierr = PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 714a13144ffSStefano Zampini ierr = PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 715a13144ffSStefano Zampini ierr = PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 716a13144ffSStefano Zampini ierr = PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 717a13144ffSStefano Zampini 7184e64d54eSstefano_zampini /* Zero rows of lGt corresponding to identified corners 7194e64d54eSstefano_zampini and interior nodal dofs */ 720a13144ffSStefano Zampini cum = 0; 721a13144ffSStefano Zampini for (i=0;i<nv;i++) { 722a13144ffSStefano Zampini if (sfvleaves[i]) { 723a13144ffSStefano Zampini vmarks[cum++] = i; 724a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 725a13144ffSStefano Zampini } 7264e64d54eSstefano_zampini if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i; 727a13144ffSStefano Zampini } 7284e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvi);CHKERRQ(ierr); 729a13144ffSStefano Zampini if (print) { 730a13144ffSStefano Zampini IS tbz; 731a13144ffSStefano Zampini 732a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 7334e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");CHKERRQ(ierr); 734a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 735a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 736a13144ffSStefano Zampini } 737a13144ffSStefano Zampini ierr = MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);CHKERRQ(ierr); 738a13144ffSStefano Zampini ierr = PetscFree(vmarks);CHKERRQ(ierr); 739a13144ffSStefano Zampini ierr = PetscSFDestroy(&sfv);CHKERRQ(ierr); 740a13144ffSStefano Zampini ierr = PetscFree2(sfvleaves,sfvroots);CHKERRQ(ierr); 741a13144ffSStefano Zampini 742a13144ffSStefano Zampini /* Recompute G */ 743a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 744a13144ffSStefano Zampini ierr = MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);CHKERRQ(ierr); 745a13144ffSStefano Zampini if (print) { 746a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"used_lG");CHKERRQ(ierr); 747a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 748a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lGt,"used_lGt");CHKERRQ(ierr); 749a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 750a13144ffSStefano Zampini } 751a13144ffSStefano Zampini 752a13144ffSStefano Zampini /* Get primal dofs (if any) */ 753a13144ffSStefano Zampini cum = 0; 754a13144ffSStefano Zampini for (i=0;i<ne;i++) { 755a13144ffSStefano Zampini if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i; 756a13144ffSStefano Zampini } 757c2151214SStefano Zampini if (fl2g) { 758c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);CHKERRQ(ierr); 759c2151214SStefano Zampini } 760a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 761a13144ffSStefano Zampini if (print) { 762a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");CHKERRQ(ierr); 763a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 764a13144ffSStefano Zampini } 765a13144ffSStefano Zampini ierr = PetscBTDestroy(&bte);CHKERRQ(ierr); 766c2151214SStefano Zampini /* TODO: what if the user passed in some of them ? */ 767a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 768a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 769a13144ffSStefano Zampini 770a13144ffSStefano Zampini /* Compute edge connectivity */ 771a13144ffSStefano Zampini ierr = PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");CHKERRQ(ierr); 772a13144ffSStefano Zampini ierr = MatMatMultSymbolic(lG,lGt,PETSC_DEFAULT,&conn);CHKERRQ(ierr); 773a13144ffSStefano Zampini ierr = MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 774c2151214SStefano Zampini if (fl2g) { 775c2151214SStefano Zampini PetscBT btf; 776c2151214SStefano Zampini PetscInt *iia,*jja,*iiu,*jju; 777c2151214SStefano Zampini PetscBool rest = PETSC_FALSE,free = PETSC_FALSE; 778c2151214SStefano Zampini 779c2151214SStefano Zampini /* create CSR for all local dofs */ 780c2151214SStefano Zampini ierr = PetscMalloc1(n+1,&iia);CHKERRQ(ierr); 781c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */ 782c2151214SStefano 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); 783c2151214SStefano Zampini iiu = pcbddc->mat_graph->xadj; 784c2151214SStefano Zampini jju = pcbddc->mat_graph->adjncy; 785c2151214SStefano Zampini } else if (pcbddc->use_local_adj) { 786c2151214SStefano Zampini rest = PETSC_TRUE; 787c2151214SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 788c2151214SStefano Zampini } else { 789c2151214SStefano Zampini free = PETSC_TRUE; 790c2151214SStefano Zampini ierr = PetscMalloc2(n+1,&iiu,n,&jju);CHKERRQ(ierr); 791c2151214SStefano Zampini iiu[0] = 0; 792c2151214SStefano Zampini for (i=0;i<n;i++) { 793c2151214SStefano Zampini iiu[i+1] = i+1; 794c2151214SStefano Zampini jju[i] = -1; 795d904f53bSStefano Zampini } 796c2151214SStefano Zampini } 797c2151214SStefano Zampini 798c2151214SStefano Zampini /* import sizes of CSR */ 799c2151214SStefano Zampini iia[0] = 0; 800c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i]; 801c2151214SStefano Zampini 802c2151214SStefano Zampini /* overwrite entries corresponding to the Nedelec field */ 803c2151214SStefano Zampini ierr = PetscBTCreate(n,&btf);CHKERRQ(ierr); 804c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 805c2151214SStefano Zampini for (i=0;i<ne;i++) { 806c2151214SStefano Zampini ierr = PetscBTSet(btf,idxs[i]);CHKERRQ(ierr); 807c2151214SStefano Zampini iia[idxs[i]+1] = ii[i+1]-ii[i]; 808c2151214SStefano Zampini } 809c2151214SStefano Zampini 810c2151214SStefano Zampini /* iia in CSR */ 811c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] += iia[i]; 812c2151214SStefano Zampini 813c2151214SStefano Zampini /* jja in CSR */ 814c2151214SStefano Zampini ierr = PetscMalloc1(iia[n],&jja);CHKERRQ(ierr); 815c2151214SStefano Zampini for (i=0;i<n;i++) 816c2151214SStefano Zampini if (!PetscBTLookup(btf,i)) 817c2151214SStefano Zampini for (j=0;j<iiu[i+1]-iiu[i];j++) 818c2151214SStefano Zampini jja[iia[i]+j] = jju[iiu[i]+j]; 819c2151214SStefano Zampini 820c2151214SStefano Zampini /* map edge dofs connectivity */ 8211e0482f5SStefano Zampini if (jj) { 822c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);CHKERRQ(ierr); 823c2151214SStefano Zampini for (i=0;i<ne;i++) { 824c2151214SStefano Zampini PetscInt e = idxs[i]; 825c2151214SStefano Zampini for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j]; 826c2151214SStefano Zampini } 8271e0482f5SStefano Zampini } 828c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 829c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);CHKERRQ(ierr); 830c2151214SStefano Zampini if (rest) { 831c2151214SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 832c2151214SStefano Zampini } 833c2151214SStefano Zampini if (free) { 834c2151214SStefano Zampini ierr = PetscFree2(iiu,jju);CHKERRQ(ierr); 835c2151214SStefano Zampini } 836c2151214SStefano Zampini ierr = PetscBTDestroy(&btf);CHKERRQ(ierr); 837c2151214SStefano Zampini } else { 838c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);CHKERRQ(ierr); 839c2151214SStefano Zampini } 840c2151214SStefano Zampini 841a13144ffSStefano Zampini /* Analyze interface for edge dofs */ 842a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 843213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 844a13144ffSStefano Zampini 845a13144ffSStefano Zampini /* Get coarse edges in the edge space */ 846c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 847a13144ffSStefano Zampini ierr = MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 848a13144ffSStefano Zampini 849c2151214SStefano Zampini if (fl2g) { 850c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 851c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 852c2151214SStefano Zampini for (i=0;i<nee;i++) { 853c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 854c2151214SStefano Zampini } 855c2151214SStefano Zampini } else { 856c2151214SStefano Zampini eedges = alleedges; 857c2151214SStefano Zampini primals = allprimals; 858c2151214SStefano Zampini } 859c2151214SStefano Zampini 860a13144ffSStefano Zampini /* Mark fine edge dofs with their coarse edge id */ 861a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 862c2151214SStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 863c2151214SStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 864c2151214SStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = nee+1; 865c2151214SStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 866c2151214SStefano Zampini if (print) { 867c2151214SStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");CHKERRQ(ierr); 868c2151214SStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 869c2151214SStefano Zampini } 870c2151214SStefano Zampini 871c2151214SStefano Zampini maxsize = 0; 872a13144ffSStefano Zampini for (i=0;i<nee;i++) { 873a13144ffSStefano Zampini PetscInt size,mark = i+1; 874a13144ffSStefano Zampini 875a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 876a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 877a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 878a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 879a13144ffSStefano Zampini maxsize = PetscMax(maxsize,size); 880a13144ffSStefano Zampini } 881a13144ffSStefano Zampini 882a13144ffSStefano Zampini /* Find coarse edge endpoints */ 883a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 884a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 885a13144ffSStefano Zampini for (i=0;i<nee;i++) { 886a13144ffSStefano Zampini PetscInt mark = i+1,size; 887a13144ffSStefano Zampini 888a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 8891e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 8901e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 891a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 892a13144ffSStefano Zampini if (print) { 893a13144ffSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %d\n",i); 894a13144ffSStefano Zampini ISView(eedges[i],NULL); 895a13144ffSStefano Zampini } 896a13144ffSStefano Zampini for (j=0;j<size;j++) { 897a13144ffSStefano Zampini PetscInt k, ee = idxs[j]; 898a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," idx %d\n",ee); 899a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 900a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," inspect %d\n",jj[k]); 901a13144ffSStefano Zampini if (PetscBTLookup(btv,jj[k])) { 902a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %d\n",jj[k]); 903a13144ffSStefano Zampini } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */ 904a13144ffSStefano Zampini PetscInt k2; 905a13144ffSStefano Zampini PetscBool corner = PETSC_FALSE; 906a13144ffSStefano Zampini for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) { 907c2151214SStefano 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])); 908c2151214SStefano Zampini /* it's a corner if either is connected with an edge dof belonging to a different cc or 909c2151214SStefano Zampini if the edge dof lie on the natural part of the boundary */ 910c2151214SStefano Zampini if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) { 911a13144ffSStefano Zampini corner = PETSC_TRUE; 912a13144ffSStefano Zampini break; 913a13144ffSStefano Zampini } 914a13144ffSStefano Zampini } 915a13144ffSStefano Zampini if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */ 916a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found %d\n",jj[k]); 917a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 918a13144ffSStefano Zampini } else { 919a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n"); 920a13144ffSStefano Zampini } 921a13144ffSStefano Zampini } 922a13144ffSStefano Zampini } 923a13144ffSStefano Zampini } 924a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 925a13144ffSStefano Zampini } 926a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 927a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 928c2151214SStefano Zampini ierr = PetscBTDestroy(&btb);CHKERRQ(ierr); 929a13144ffSStefano Zampini 930a13144ffSStefano Zampini /* Reset marked primal dofs */ 931a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 932a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 933a13144ffSStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = 0; 934a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 935a13144ffSStefano Zampini 9360569b399SStefano Zampini /* Now use the initial lG */ 9370569b399SStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 9380569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 9390569b399SStefano Zampini lG = lGinit; 9400569b399SStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 9410569b399SStefano Zampini 942a13144ffSStefano Zampini /* Compute extended cols indices */ 943b03ebc13SStefano Zampini ierr = PetscBTCreate(nv,&btvc);CHKERRQ(ierr); 944b03ebc13SStefano Zampini ierr = PetscBTCreate(nee,&bter);CHKERRQ(ierr); 945a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 946a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lG,&i);CHKERRQ(ierr); 947a13144ffSStefano Zampini i *= maxsize; 948b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 949a13144ffSStefano Zampini ierr = PetscMalloc2(i,&extrow,i,&gidxs);CHKERRQ(ierr); 950a13144ffSStefano Zampini eerr = PETSC_FALSE; 951a13144ffSStefano Zampini for (i=0;i<nee;i++) { 952b03ebc13SStefano Zampini PetscInt size,found = 0; 953a13144ffSStefano Zampini 954a13144ffSStefano Zampini cum = 0; 955a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 9561e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 9571e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 958a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 959b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 960a13144ffSStefano Zampini for (j=0;j<size;j++) { 961a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 962b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 963b03ebc13SStefano Zampini PetscInt vv = jj[k]; 964b03ebc13SStefano Zampini if (!PetscBTLookup(btv,vv)) extrow[cum++] = vv; 965b03ebc13SStefano Zampini else if (!PetscBTLookupSet(btvc,vv)) found++; 966b03ebc13SStefano Zampini } 967a13144ffSStefano Zampini } 968a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 969a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 970a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 971a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 972a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 973a13144ffSStefano Zampini /* it may happen that endpoints are not defined at this point 974a13144ffSStefano Zampini if it is the case, mark this edge for a second pass */ 975b03ebc13SStefano Zampini if (cum != size -1 || found != 2) { 976b03ebc13SStefano Zampini ierr = PetscBTSet(bter,i);CHKERRQ(ierr); 977a13144ffSStefano Zampini if (print) { 978a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge");CHKERRQ(ierr); 979a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 980a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol");CHKERRQ(ierr); 981a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 982a13144ffSStefano Zampini } 983a13144ffSStefano Zampini eerr = PETSC_TRUE; 984a13144ffSStefano Zampini } 985a13144ffSStefano Zampini } 9864e64d54eSstefano_zampini /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */ 987a13144ffSStefano Zampini ierr = MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 988a13144ffSStefano Zampini if (done) { 989a13144ffSStefano Zampini PetscInt *newprimals; 990a13144ffSStefano Zampini 991a13144ffSStefano Zampini ierr = PetscMalloc1(ne,&newprimals);CHKERRQ(ierr); 992a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 993a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 994a13144ffSStefano Zampini ierr = PetscMemcpy(newprimals,idxs,cum*sizeof(PetscInt));CHKERRQ(ierr); 995a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 9960569b399SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 997b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"DOING SECOND PASS (eerr %d)\n",eerr); 998a13144ffSStefano Zampini for (i=0;i<nee;i++) { 999b03ebc13SStefano Zampini PetscBool has_candidates = PETSC_FALSE; 1000b03ebc13SStefano Zampini if (PetscBTLookup(bter,i)) { 1001a13144ffSStefano Zampini PetscInt size,mark = i+1; 1002a13144ffSStefano Zampini 1003a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 1004a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1005c2151214SStefano Zampini /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */ 1006a13144ffSStefano Zampini for (j=0;j<size;j++) { 1007a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 1008b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"Inspecting edge dof %d [%d %d)\n",ee,ii[ee],ii[ee+1]); 1009a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 1010a13144ffSStefano Zampini /* set all candidates located on the edge as corners */ 1011a13144ffSStefano Zampini if (PetscBTLookup(btvcand,jj[k])) { 1012a13144ffSStefano Zampini PetscInt k2,vv = jj[k]; 1013b03ebc13SStefano Zampini has_candidates = PETSC_TRUE; 1014b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Candidate set to vertex %d\n",vv); 1015a13144ffSStefano Zampini ierr = PetscBTSet(btv,vv);CHKERRQ(ierr); 1016a13144ffSStefano Zampini /* set all edge dofs connected to candidate as primals */ 1017a13144ffSStefano Zampini for (k2=iit[vv];k2<iit[vv+1];k2++) { 1018a13144ffSStefano Zampini if (marks[jjt[k2]] == mark) { 1019a13144ffSStefano Zampini PetscInt k3,ee2 = jjt[k2]; 1020b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected edge dof set to primal %d\n",ee2); 1021a13144ffSStefano Zampini newprimals[cum++] = ee2; 1022a13144ffSStefano Zampini /* finally set the new corners */ 1023a13144ffSStefano Zampini for (k3=ii[ee2];k3<ii[ee2+1];k3++) { 1024b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected nodal dof set to vertex %d\n",jj[k3]); 1025a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k3]);CHKERRQ(ierr); 1026a13144ffSStefano Zampini } 1027a13144ffSStefano Zampini } 1028a13144ffSStefano Zampini } 1029b03ebc13SStefano Zampini } else { 1030b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Not a candidate vertex %d\n",jj[k]); 1031a13144ffSStefano Zampini } 1032a13144ffSStefano Zampini } 1033a13144ffSStefano Zampini } 1034b03ebc13SStefano Zampini if (!has_candidates) { /* circular edge */ 1035b03ebc13SStefano Zampini PetscInt k, ee = idxs[0],*tmarks; 1036b03ebc13SStefano Zampini 1037b03ebc13SStefano Zampini ierr = PetscCalloc1(ne,&tmarks);CHKERRQ(ierr); 1038b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Circular edge %d\n",i); 1039b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 1040b03ebc13SStefano Zampini PetscInt k2; 1041b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Set to corner %d\n",jj[k]); 1042b03ebc13SStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 1043b03ebc13SStefano Zampini for (k2=iit[jj[k]];k2<iit[jj[k]+1];k2++) tmarks[jjt[k2]]++; 1044b03ebc13SStefano Zampini } 1045b03ebc13SStefano Zampini for (j=0;j<size;j++) { 1046b03ebc13SStefano Zampini if (tmarks[idxs[j]] > 1) { 1047b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Edge dof set to primal %d\n",idxs[j]); 1048b03ebc13SStefano Zampini newprimals[cum++] = idxs[j]; 1049b03ebc13SStefano Zampini } 1050b03ebc13SStefano Zampini } 1051b03ebc13SStefano Zampini ierr = PetscFree(tmarks);CHKERRQ(ierr); 1052b03ebc13SStefano Zampini } 1053a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1054a13144ffSStefano Zampini } 1055a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 1056a13144ffSStefano Zampini } 1057b03ebc13SStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 10580569b399SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 1059a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,newprimals);CHKERRQ(ierr); 1060c2151214SStefano Zampini if (fl2g) { 1061c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);CHKERRQ(ierr); 1062c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1063c2151214SStefano Zampini for (i=0;i<nee;i++) { 1064c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1065c2151214SStefano Zampini } 1066c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1067c2151214SStefano Zampini } 1068c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1069a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 1070a13144ffSStefano Zampini ierr = PetscFree(newprimals);CHKERRQ(ierr); 1071a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 1072a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1073a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 1074213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 1075c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1076c2151214SStefano Zampini if (fl2g) { 1077c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 1078c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 1079c2151214SStefano Zampini for (i=0;i<nee;i++) { 1080c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 1081c2151214SStefano Zampini } 1082c2151214SStefano Zampini } else { 1083c2151214SStefano Zampini eedges = alleedges; 1084c2151214SStefano Zampini primals = allprimals; 1085c2151214SStefano Zampini } 1086b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 1087a13144ffSStefano Zampini 1088a13144ffSStefano Zampini /* Mark again */ 1089a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 1090a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1091a13144ffSStefano Zampini PetscInt size,mark = i+1; 1092a13144ffSStefano Zampini 1093a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 1094a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1095a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 1096a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1097a13144ffSStefano Zampini } 1098a13144ffSStefano Zampini if (print) { 1099a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");CHKERRQ(ierr); 1100a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 1101a13144ffSStefano Zampini } 1102a13144ffSStefano Zampini 1103a13144ffSStefano Zampini /* Recompute extended cols */ 1104a13144ffSStefano Zampini eerr = PETSC_FALSE; 1105a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1106a13144ffSStefano Zampini PetscInt size; 1107a13144ffSStefano Zampini 1108a13144ffSStefano Zampini cum = 0; 1109a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 11101e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 11111e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 1112a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1113a13144ffSStefano Zampini for (j=0;j<size;j++) { 1114a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 11151e0482f5SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) if (!PetscBTLookup(btv,jj[k])) extrow[cum++] = jj[k]; 1116a13144ffSStefano Zampini } 1117a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1118a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 1119a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 1120a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 1121a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 1122a13144ffSStefano Zampini if (cum != size -1) { 1123a13144ffSStefano Zampini if (print) { 1124a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");CHKERRQ(ierr); 1125a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 1126a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");CHKERRQ(ierr); 1127a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 1128a13144ffSStefano Zampini } 1129a13144ffSStefano Zampini eerr = PETSC_TRUE; 1130a13144ffSStefano Zampini } 1131a13144ffSStefano Zampini } 1132a13144ffSStefano Zampini } 1133a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1134a13144ffSStefano Zampini ierr = PetscFree2(extrow,gidxs);CHKERRQ(ierr); 1135b03ebc13SStefano Zampini ierr = PetscBTDestroy(&bter);CHKERRQ(ierr); 11367d871cd7SStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 1137a13144ffSStefano Zampini /* an error should not occur at this point */ 1138a13144ffSStefano Zampini if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS"); 1139a13144ffSStefano Zampini 11404e64d54eSstefano_zampini /* Check the number of endpoints */ 11410569b399SStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1142b03ebc13SStefano Zampini ierr = PetscMalloc1(2*nee,&corners);CHKERRQ(ierr); 1143b03ebc13SStefano Zampini ierr = PetscMalloc1(nee,&cedges);CHKERRQ(ierr); 11444e64d54eSstefano_zampini for (i=0;i<nee;i++) { 1145b03ebc13SStefano Zampini PetscInt size, found = 0, gc[2]; 11464e64d54eSstefano_zampini 1147b03ebc13SStefano Zampini /* init with defaults */ 1148b03ebc13SStefano Zampini cedges[i] = corners[i*2] = corners[i*2+1] = -1; 11494e64d54eSstefano_zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 11501e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 11511e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 11524e64d54eSstefano_zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1153b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 11544e64d54eSstefano_zampini for (j=0;j<size;j++) { 11554e64d54eSstefano_zampini PetscInt k,ee = idxs[j]; 11564e64d54eSstefano_zampini for (k=ii[ee];k<ii[ee+1];k++) { 11574e64d54eSstefano_zampini PetscInt vv = jj[k]; 11584e64d54eSstefano_zampini if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) { 1159b03ebc13SStefano Zampini if (found == 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found more then two corners for edge %d\n",i); 1160b03ebc13SStefano Zampini corners[i*2+found++] = vv; 11614e64d54eSstefano_zampini } 11624e64d54eSstefano_zampini } 11634e64d54eSstefano_zampini } 1164b03ebc13SStefano Zampini if (found != 2) { 1165b03ebc13SStefano Zampini PetscInt e; 1166b03ebc13SStefano Zampini if (fl2g) { 1167b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,1,idxs,&e);CHKERRQ(ierr); 1168b03ebc13SStefano Zampini } else { 1169b03ebc13SStefano Zampini e = idxs[0]; 1170b03ebc13SStefano Zampini } 1171b03ebc13SStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %d corners for edge %d (astart %d, estart %d)\n",found,i,e,idxs[0]); 1172b03ebc13SStefano Zampini } 1173eee23b56SStefano Zampini 1174eee23b56SStefano Zampini /* get primal dof index on this coarse edge */ 1175b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,2,corners+2*i,gc);CHKERRQ(ierr); 1176b03ebc13SStefano Zampini if (gc[0] > gc[1]) { 1177b03ebc13SStefano Zampini PetscInt swap = corners[2*i]; 1178b03ebc13SStefano Zampini corners[2*i] = corners[2*i+1]; 1179b03ebc13SStefano Zampini corners[2*i+1] = swap; 1180b03ebc13SStefano Zampini } 1181eee23b56SStefano Zampini cedges[i] = idxs[size-1]; 11824e64d54eSstefano_zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1183b03ebc13SStefano 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]); 11844e64d54eSstefano_zampini } 11850569b399SStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 11864e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvc);CHKERRQ(ierr); 11874e64d54eSstefano_zampini 1188a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 1189a13144ffSStefano Zampini /* Inspects columns of lG (rows of lGt) and make sure the change of basis will 1190a13144ffSStefano Zampini not interfere with neighbouring coarse edges */ 1191a13144ffSStefano Zampini ierr = PetscMalloc1(nee+1,&emarks);CHKERRQ(ierr); 1192a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1193a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1194a13144ffSStefano Zampini PetscInt emax = 0,eemax = 0; 1195a13144ffSStefano Zampini 1196a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1197a13144ffSStefano Zampini ierr = PetscMemzero(emarks,(nee+1)*sizeof(PetscInt));CHKERRQ(ierr); 1198a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++; 1199a13144ffSStefano Zampini for (j=1;j<nee+1;j++) { 1200a13144ffSStefano Zampini if (emax < emarks[j]) { 1201a13144ffSStefano Zampini emax = emarks[j]; 1202a13144ffSStefano Zampini eemax = j; 1203a13144ffSStefano Zampini } 1204a13144ffSStefano Zampini } 1205a13144ffSStefano Zampini /* not relevant for edges */ 1206a13144ffSStefano Zampini if (!eemax) continue; 1207a13144ffSStefano Zampini 1208a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 1209a13144ffSStefano Zampini if (marks[jj[j]] && marks[jj[j]] != eemax) { 1210c2151214SStefano 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]); 1211a13144ffSStefano Zampini } 1212a13144ffSStefano Zampini } 1213a13144ffSStefano Zampini } 1214a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 1215a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1216a13144ffSStefano Zampini #endif 1217a13144ffSStefano Zampini 1218a13144ffSStefano Zampini /* Compute extended rows indices for edge blocks of the change of basis */ 1219a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1220a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);CHKERRQ(ierr); 1221a13144ffSStefano Zampini extmem *= maxsize; 1222a13144ffSStefano Zampini ierr = PetscMalloc1(extmem*nee,&extrow);CHKERRQ(ierr); 1223a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extrows);CHKERRQ(ierr); 1224a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&extrowcum);CHKERRQ(ierr); 1225a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1226a13144ffSStefano Zampini PetscInt mark = 0,size,start; 1227213b8bfaSStefano Zampini 1228a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1229a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) 1230a13144ffSStefano Zampini if (marks[jj[j]] && !mark) 1231a13144ffSStefano Zampini mark = marks[jj[j]]; 1232a13144ffSStefano Zampini 1233a13144ffSStefano Zampini /* not relevant */ 1234a13144ffSStefano Zampini if (!mark) continue; 1235a13144ffSStefano Zampini 1236a13144ffSStefano Zampini /* import extended row */ 1237a13144ffSStefano Zampini mark--; 1238a13144ffSStefano Zampini start = mark*extmem+extrowcum[mark]; 1239a13144ffSStefano Zampini size = ii[i+1]-ii[i]; 1240a13144ffSStefano Zampini if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %d > %d",extrowcum[mark] + size,extmem); 1241a13144ffSStefano Zampini ierr = PetscMemcpy(extrow+start,jj+ii[i],size*sizeof(PetscInt));CHKERRQ(ierr); 1242a13144ffSStefano Zampini extrowcum[mark] += size; 1243a13144ffSStefano Zampini } 1244a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1245213b8bfaSStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 1246213b8bfaSStefano Zampini ierr = PetscFree(marks);CHKERRQ(ierr); 1247213b8bfaSStefano Zampini 1248213b8bfaSStefano Zampini /* Compress extrows */ 1249a13144ffSStefano Zampini cum = 0; 1250a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1251a13144ffSStefano Zampini PetscInt size = extrowcum[i],*start = extrow + i*extmem; 1252a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&size,start);CHKERRQ(ierr); 1253a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);CHKERRQ(ierr); 1254a13144ffSStefano Zampini cum = PetscMax(cum,size); 1255a13144ffSStefano Zampini } 1256a13144ffSStefano Zampini ierr = PetscFree(extrowcum);CHKERRQ(ierr); 1257a13144ffSStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 1258a13144ffSStefano Zampini ierr = PetscBTDestroy(&btvcand);CHKERRQ(ierr); 1259a13144ffSStefano Zampini 1260a13144ffSStefano Zampini /* Workspace for lapack inner calls and VecSetValues */ 1261a13144ffSStefano Zampini ierr = PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);CHKERRQ(ierr); 1262a13144ffSStefano Zampini 1263a13144ffSStefano Zampini /* Create change of basis matrix (preallocation can be improved) */ 1264a13144ffSStefano Zampini ierr = MatCreate(comm,&T);CHKERRQ(ierr); 1265c2151214SStefano Zampini ierr = MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n, 1266c2151214SStefano Zampini pc->pmat->rmap->N,pc->pmat->rmap->N);CHKERRQ(ierr); 1267a13144ffSStefano Zampini ierr = MatSetType(T,MATAIJ);CHKERRQ(ierr); 1268a13144ffSStefano Zampini ierr = MatSeqAIJSetPreallocation(T,10,NULL);CHKERRQ(ierr); 1269a13144ffSStefano Zampini ierr = MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);CHKERRQ(ierr); 12701e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(T,al2g,al2g);CHKERRQ(ierr); 1271a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1272a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 1273213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&al2g);CHKERRQ(ierr); 1274a13144ffSStefano Zampini 1275a13144ffSStefano Zampini /* Defaults to identity */ 1276c2151214SStefano Zampini ierr = MatCreateVecs(pc->pmat,&tvec,NULL);CHKERRQ(ierr); 1277a13144ffSStefano Zampini ierr = VecSet(tvec,1.0);CHKERRQ(ierr); 1278a13144ffSStefano Zampini ierr = MatDiagonalSet(T,tvec,INSERT_VALUES);CHKERRQ(ierr); 1279a13144ffSStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 1280a13144ffSStefano Zampini 12811e0482f5SStefano Zampini /* Create discrete gradient for the coarser level if needed */ 12821e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 12831e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 12841e0482f5SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 12851e0482f5SStefano Zampini ISLocalToGlobalMapping cel2g,cvl2g; 12861e0482f5SStefano Zampini IS wis,gwis; 12871e0482f5SStefano Zampini PetscInt cnv,cne; 12881e0482f5SStefano Zampini 12891e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,nee,cedges,PETSC_COPY_VALUES,&wis);CHKERRQ(ierr); 12901e0482f5SStefano Zampini if (fl2g) { 12911e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(fl2g,wis,&pcbddc->nedclocal);CHKERRQ(ierr); 12921e0482f5SStefano Zampini } else { 12931e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)wis);CHKERRQ(ierr); 12941e0482f5SStefano Zampini pcbddc->nedclocal = wis; 12951e0482f5SStefano Zampini } 12961e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(el2g,wis,&gwis);CHKERRQ(ierr); 12971e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12981e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cne,&wis);CHKERRQ(ierr); 12991e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cel2g);CHKERRQ(ierr); 13001e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13011e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 13021e0482f5SStefano Zampini 13031e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,2*nee,corners,PETSC_USE_POINTER,&wis);CHKERRQ(ierr); 13041e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(vl2g,wis,&gwis);CHKERRQ(ierr); 13051e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13061e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cnv,&wis);CHKERRQ(ierr); 13071e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cvl2g);CHKERRQ(ierr); 13081e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13091e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 13101e0482f5SStefano Zampini 13111e0482f5SStefano Zampini ierr = MatCreate(comm,&pcbddc->nedcG);CHKERRQ(ierr); 13121e0482f5SStefano Zampini ierr = MatSetSizes(pcbddc->nedcG,PETSC_DECIDE,PETSC_DECIDE,cne,cnv);CHKERRQ(ierr); 13131e0482f5SStefano Zampini ierr = MatSetType(pcbddc->nedcG,MATAIJ);CHKERRQ(ierr); 13141e0482f5SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->nedcG,2,NULL);CHKERRQ(ierr); 13151e0482f5SStefano Zampini ierr = MatMPIAIJSetPreallocation(pcbddc->nedcG,2,NULL,2,NULL);CHKERRQ(ierr); 13161e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->nedcG,cel2g,cvl2g);CHKERRQ(ierr); 13171e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cel2g);CHKERRQ(ierr); 13181e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cvl2g);CHKERRQ(ierr); 13191e0482f5SStefano Zampini } 1320213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&vl2g);CHKERRQ(ierr); 13211e0482f5SStefano Zampini 13221e0482f5SStefano Zampini #if defined(PRINT_GDET) 13231e0482f5SStefano Zampini inc = 0; 13241e0482f5SStefano Zampini lev = pcbddc->current_level; 13251e0482f5SStefano Zampini #endif 1326213b8bfaSStefano Zampini 1327213b8bfaSStefano Zampini /* Insert values in the change of basis matrix */ 1328a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1329a13144ffSStefano Zampini Mat Gins = NULL, GKins = NULL; 13301e0482f5SStefano Zampini IS cornersis = NULL; 13311e0482f5SStefano Zampini PetscScalar cvals[2]; 1332a13144ffSStefano Zampini 13331e0482f5SStefano Zampini if (pcbddc->nedcG) { 13341e0482f5SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,2,corners+2*i,PETSC_USE_POINTER,&cornersis);CHKERRQ(ierr); 13351e0482f5SStefano Zampini } 13361e0482f5SStefano Zampini ierr = PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],cornersis,&Gins,&GKins,cvals,work,rwork);CHKERRQ(ierr); 1337a13144ffSStefano Zampini if (Gins && GKins) { 1338a13144ffSStefano Zampini PetscScalar *data; 1339a13144ffSStefano Zampini const PetscInt *rows,*cols; 1340a13144ffSStefano Zampini PetscInt nrh,nch,nrc,ncc; 1341a13144ffSStefano Zampini 1342a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&cols);CHKERRQ(ierr); 1343a13144ffSStefano Zampini /* H1 */ 1344a13144ffSStefano Zampini ierr = ISGetIndices(extrows[i],&rows);CHKERRQ(ierr); 1345a13144ffSStefano Zampini ierr = MatGetSize(Gins,&nrh,&nch);CHKERRQ(ierr); 1346a13144ffSStefano Zampini ierr = MatDenseGetArray(Gins,&data);CHKERRQ(ierr); 1347a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);CHKERRQ(ierr); 1348a13144ffSStefano Zampini ierr = MatDenseRestoreArray(Gins,&data);CHKERRQ(ierr); 1349a13144ffSStefano Zampini ierr = ISRestoreIndices(extrows[i],&rows);CHKERRQ(ierr); 1350a13144ffSStefano Zampini /* complement */ 1351a13144ffSStefano Zampini ierr = MatGetSize(GKins,&nrc,&ncc);CHKERRQ(ierr); 13521e0482f5SStefano Zampini if (!ncc) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Constant function has not been generated for coarse edge %d",i); 1353213b8bfaSStefano 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); 1354213b8bfaSStefano 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); 1355a13144ffSStefano Zampini ierr = MatDenseGetArray(GKins,&data);CHKERRQ(ierr); 1356a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);CHKERRQ(ierr); 1357a13144ffSStefano Zampini ierr = MatDenseRestoreArray(GKins,&data);CHKERRQ(ierr); 13581e0482f5SStefano Zampini 13591e0482f5SStefano Zampini /* coarse discrete gradient */ 13601e0482f5SStefano Zampini if (pcbddc->nedcG) { 13611e0482f5SStefano Zampini PetscInt cols[2]; 13621e0482f5SStefano Zampini 13631e0482f5SStefano Zampini cols[0] = 2*i; 13641e0482f5SStefano Zampini cols[1] = 2*i+1; 13651e0482f5SStefano Zampini ierr = MatSetValuesLocal(pcbddc->nedcG,1,&i,2,cols,cvals,INSERT_VALUES);CHKERRQ(ierr); 13661e0482f5SStefano Zampini } 1367a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&cols);CHKERRQ(ierr); 1368a13144ffSStefano Zampini } 1369a13144ffSStefano Zampini ierr = ISDestroy(&extrows[i]);CHKERRQ(ierr); 1370a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 13711e0482f5SStefano Zampini ierr = ISDestroy(&cornersis);CHKERRQ(ierr); 1372a13144ffSStefano Zampini ierr = MatDestroy(&Gins);CHKERRQ(ierr); 1373a13144ffSStefano Zampini ierr = MatDestroy(&GKins);CHKERRQ(ierr); 1374a13144ffSStefano Zampini } 1375213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&el2g);CHKERRQ(ierr); 1376a13144ffSStefano Zampini 1377a13144ffSStefano Zampini /* Start assembling */ 1378a13144ffSStefano Zampini ierr = MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13791e0482f5SStefano Zampini if (pcbddc->nedcG) { 13801e0482f5SStefano Zampini ierr = MatAssemblyBegin(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13811e0482f5SStefano Zampini } 1382a13144ffSStefano Zampini 1383a13144ffSStefano Zampini /* Free */ 1384c2151214SStefano Zampini if (fl2g) { 1385c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1386c2151214SStefano Zampini for (i=0;i<nee;i++) { 1387c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1388c2151214SStefano Zampini } 1389c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1390c2151214SStefano Zampini } 1391eee23b56SStefano Zampini 1392eee23b56SStefano Zampini /* hack mat_graph with primal dofs on the coarse edges */ 1393eee23b56SStefano Zampini { 1394eee23b56SStefano Zampini PCBDDCGraph graph = pcbddc->mat_graph; 1395eee23b56SStefano Zampini PetscInt *oqueue = graph->queue; 1396eee23b56SStefano Zampini PetscInt *ocptr = graph->cptr; 1397eee23b56SStefano Zampini PetscInt ncc,*idxs; 1398eee23b56SStefano Zampini 1399eee23b56SStefano Zampini /* find first primal edge */ 1400eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1401eee23b56SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1402eee23b56SStefano Zampini } else { 1403eee23b56SStefano Zampini if (fl2g) { 1404eee23b56SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,nee,cedges,cedges);CHKERRQ(ierr); 1405eee23b56SStefano Zampini } 1406eee23b56SStefano Zampini idxs = cedges; 1407eee23b56SStefano Zampini } 1408eee23b56SStefano Zampini cum = 0; 1409eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1410eee23b56SStefano Zampini 1411eee23b56SStefano Zampini /* adapt connected components */ 1412eee23b56SStefano Zampini ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,ocptr[graph->ncc],&graph->queue);CHKERRQ(ierr); 1413eee23b56SStefano Zampini graph->cptr[0] = 0; 1414eee23b56SStefano Zampini for (i=0,ncc=0;i<graph->ncc;i++) { 1415eee23b56SStefano Zampini PetscInt lc = ocptr[i+1]-ocptr[i]; 1416eee23b56SStefano Zampini if (cum != nee && oqueue[ocptr[i+1]-1] == cedges[cum]) { /* this cc has a primal dof */ 1417eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc]+1; 1418eee23b56SStefano Zampini graph->queue[graph->cptr[ncc]] = cedges[cum]; 1419eee23b56SStefano Zampini ncc++; 1420eee23b56SStefano Zampini lc--; 1421eee23b56SStefano Zampini cum++; 1422eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1423eee23b56SStefano Zampini } 1424eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc] + lc; 1425eee23b56SStefano Zampini for (j=0;j<lc;j++) graph->queue[graph->cptr[ncc]+j] = oqueue[ocptr[i]+j]; 1426eee23b56SStefano Zampini ncc++; 1427eee23b56SStefano Zampini } 1428eee23b56SStefano Zampini graph->ncc = ncc; 1429eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1430eee23b56SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1431eee23b56SStefano Zampini } 1432eee23b56SStefano Zampini ierr = PetscFree2(ocptr,oqueue);CHKERRQ(ierr); 1433eee23b56SStefano Zampini } 1434213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&fl2g);CHKERRQ(ierr); 1435c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1436c2151214SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1437213b8bfaSStefano Zampini ierr = MatDestroy(&conn);CHKERRQ(ierr); 1438eee23b56SStefano Zampini 1439c2151214SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 1440a13144ffSStefano Zampini ierr = PetscFree(extrow);CHKERRQ(ierr); 1441a13144ffSStefano Zampini ierr = PetscFree2(work,rwork);CHKERRQ(ierr); 1442b03ebc13SStefano Zampini ierr = PetscFree(corners);CHKERRQ(ierr); 1443b03ebc13SStefano Zampini ierr = PetscFree(cedges);CHKERRQ(ierr); 1444a13144ffSStefano Zampini ierr = PetscFree(extrows);CHKERRQ(ierr); 1445a13144ffSStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 1446a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 1447a13144ffSStefano Zampini 1448a13144ffSStefano Zampini /* Complete assembling */ 1449a13144ffSStefano Zampini ierr = MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14501e0482f5SStefano Zampini if (pcbddc->nedcG) { 14511e0482f5SStefano Zampini ierr = MatAssemblyEnd(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14521e0482f5SStefano Zampini #if 0 14531e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->nedcG,"coarse_G");CHKERRQ(ierr); 14541e0482f5SStefano Zampini ierr = MatView(pcbddc->nedcG,NULL);CHKERRQ(ierr); 14551e0482f5SStefano Zampini #endif 14561e0482f5SStefano Zampini } 1457a13144ffSStefano Zampini 1458a13144ffSStefano Zampini /* set change of basis */ 1459213b8bfaSStefano Zampini ierr = PCBDDCSetChangeOfBasisMat(pc,T,singular);CHKERRQ(ierr); 1460a13144ffSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1461a13144ffSStefano Zampini 1462a13144ffSStefano Zampini PetscFunctionReturn(0); 1463a13144ffSStefano Zampini } 1464a13144ffSStefano Zampini 1465d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 1466d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 1467d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 1468d8203eabSStefano Zampini #undef __FUNCT__ 1469d8203eabSStefano Zampini #define __FUNCT__ "PCBDDCNullSpaceCreate" 1470d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 1471d8203eabSStefano Zampini { 1472d8203eabSStefano Zampini PetscErrorCode ierr; 1473d8203eabSStefano Zampini PetscInt i; 1474d8203eabSStefano Zampini 1475d8203eabSStefano Zampini PetscFunctionBegin; 1476d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 1477d8203eabSStefano Zampini PetscInt first,last; 1478d8203eabSStefano Zampini 1479d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 148086fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 1481d8203eabSStefano Zampini if (i>=first && i < last) { 1482d8203eabSStefano Zampini PetscScalar *data; 1483d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1484d8203eabSStefano Zampini if (!has_const) { 1485d8203eabSStefano Zampini data[i-first] = 1.; 1486d8203eabSStefano Zampini } else { 148786fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 148886fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 1489d8203eabSStefano Zampini } 1490d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1491d8203eabSStefano Zampini } 1492d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1493d8203eabSStefano Zampini } 1494d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 1495d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 1496d8203eabSStefano Zampini PetscInt first,last; 1497d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 1498d8203eabSStefano Zampini if (i>=first && i < last) { 1499d8203eabSStefano Zampini PetscScalar *data; 1500d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1501d8203eabSStefano Zampini if (!has_const) { 1502d8203eabSStefano Zampini data[i-first] = 0.; 1503d8203eabSStefano Zampini } else { 150486fa73c5SStefano Zampini data[2*i-first] = 0.; 150586fa73c5SStefano Zampini data[2*i-first+1] = 0.; 1506d8203eabSStefano Zampini } 1507d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1508d8203eabSStefano Zampini } 1509d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1510d8203eabSStefano Zampini } 1511d8203eabSStefano Zampini PetscFunctionReturn(0); 1512d8203eabSStefano Zampini } 1513d8203eabSStefano Zampini 1514669cc0f4SStefano Zampini #undef __FUNCT__ 1515669cc0f4SStefano Zampini #define __FUNCT__ "PCBDDCComputeNoNetFlux" 15168ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 1517669cc0f4SStefano Zampini { 1518a198735bSStefano Zampini Mat loc_divudotp; 1519fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 15208ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 1521669cc0f4SStefano Zampini IS *faces,*edges; 1522669cc0f4SStefano Zampini PetscScalar *vals; 1523669cc0f4SStefano Zampini const PetscScalar *array; 1524669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 15251ae86dd6SStefano Zampini PetscMPIInt rank; 1526a198735bSStefano Zampini PetscErrorCode ierr; 1527669cc0f4SStefano Zampini 1528669cc0f4SStefano Zampini PetscFunctionBegin; 1529669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1530669cc0f4SStefano Zampini if (graph->twodim) { 1531669cc0f4SStefano Zampini lmaxneighs = 2; 1532669cc0f4SStefano Zampini } else { 1533669cc0f4SStefano Zampini lmaxneighs = 1; 1534669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1535669cc0f4SStefano Zampini const PetscInt *idxs; 1536669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1537669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 1538669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1539669cc0f4SStefano Zampini } 1540669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 1541669cc0f4SStefano Zampini } 1542669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1543669cc0f4SStefano Zampini maxsize = 0; 1544669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1545669cc0f4SStefano Zampini PetscInt nn; 1546669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1547669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1548669cc0f4SStefano Zampini } 1549669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1550669cc0f4SStefano Zampini PetscInt nn; 1551669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1552669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1553669cc0f4SStefano Zampini } 1554669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 1555669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 1556669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 15578ae0ca82SStefano Zampini if (!transpose) { 15588ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 15598ae0ca82SStefano Zampini } else { 15608ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 15618ae0ca82SStefano Zampini } 1562669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 15631ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 1564d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 1565669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 15668ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 1567669cc0f4SStefano Zampini } 1568d8203eabSStefano Zampini 1569669cc0f4SStefano Zampini /* compute local quad vec */ 1570a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 15718ae0ca82SStefano Zampini if (!transpose) { 1572a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 15738ae0ca82SStefano Zampini } else { 15748ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 15758ae0ca82SStefano Zampini } 1576669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 15778ae0ca82SStefano Zampini if (!transpose) { 1578a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 15798ae0ca82SStefano Zampini } else { 15808ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 15818ae0ca82SStefano Zampini } 1582fa23a32eSStefano Zampini if (vl2l) { 1583fa23a32eSStefano Zampini ierr = VecGetSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1584fa23a32eSStefano Zampini } else { 1585fa23a32eSStefano Zampini vins = v; 1586fa23a32eSStefano Zampini } 1587fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 1588669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 15899a962809SStefano Zampini 15901ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 15911ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 1592669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1593669cc0f4SStefano Zampini const PetscInt *idxs; 1594669cc0f4SStefano Zampini PetscInt idx,nn,j; 1595669cc0f4SStefano Zampini 1596669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 1597669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1598669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 15991ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1600669cc0f4SStefano Zampini idx = -(idx+1); 1601669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1602669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 1603669cc0f4SStefano Zampini } 1604669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1605669cc0f4SStefano Zampini const PetscInt *idxs; 1606669cc0f4SStefano Zampini PetscInt idx,nn,j; 1607669cc0f4SStefano Zampini 1608669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1609669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1610669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 16111ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1612669cc0f4SStefano Zampini idx = -(idx+1); 1613669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1614669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1615669cc0f4SStefano Zampini } 1616c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1617fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 1618fa23a32eSStefano Zampini if (vl2l) { 1619fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1620fa23a32eSStefano Zampini } 1621669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1622669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1623669cc0f4SStefano Zampini 1624669cc0f4SStefano Zampini /* assemble near null space */ 1625669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1626669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 1627669cc0f4SStefano Zampini } 1628669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1629669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 1630669cc0f4SStefano Zampini } 1631669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 1632669cc0f4SStefano Zampini PetscFunctionReturn(0); 1633669cc0f4SStefano Zampini } 1634669cc0f4SStefano Zampini 1635669cc0f4SStefano Zampini 1636a3df083aSStefano Zampini #undef __FUNCT__ 16371f4df5f7SStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalTopologyInfo" 16381f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 16391f4df5f7SStefano Zampini { 16401f4df5f7SStefano Zampini PetscErrorCode ierr; 16411f4df5f7SStefano Zampini Vec local,global; 16421f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16431f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 16441f4df5f7SStefano Zampini 16451f4df5f7SStefano Zampini PetscFunctionBegin; 16461f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 16471f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 16481f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 16491f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 16501f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 16511f4df5f7SStefano Zampini PetscInt i; 16521f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16531f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 16541f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16551f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 16561f4df5f7SStefano Zampini } 16571f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 16581f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 16591f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 16601f4df5f7SStefano Zampini } 16611f4df5f7SStefano Zampini } else { 1662986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 16631f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1664986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 1665986cdee1SStefano Zampini if (i > 1) { 1666986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 16671f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16681f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16691f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16701f4df5f7SStefano Zampini } 16711f4df5f7SStefano Zampini } 16721f4df5f7SStefano Zampini } 1673986cdee1SStefano Zampini } 16741f4df5f7SStefano Zampini 16751f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 16761f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16771f4df5f7SStefano Zampini } 16781f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 16791f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16801f4df5f7SStefano Zampini } 16811f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 16821f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 16831f4df5f7SStefano Zampini } 16841f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 16851f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 16861f4df5f7SStefano Zampini PetscFunctionReturn(0); 16871f4df5f7SStefano Zampini } 16881f4df5f7SStefano Zampini 16891f4df5f7SStefano Zampini #undef __FUNCT__ 16903e589ea0SStefano Zampini #define __FUNCT__ "PCBDDCBenignRemoveInterior" 16913e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 16923e589ea0SStefano Zampini { 16933e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 16943e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 16953e589ea0SStefano Zampini PetscErrorCode ierr; 16963e589ea0SStefano Zampini 16973e589ea0SStefano Zampini PetscFunctionBegin; 16983e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 16993e589ea0SStefano Zampini PetscFunctionReturn(0); 17003e589ea0SStefano Zampini } 17013e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 17023e589ea0SStefano Zampini Vec swap; 17033e589ea0SStefano Zampini 17043e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 17053e589ea0SStefano Zampini swap = pcbddc->work_change; 17063e589ea0SStefano Zampini pcbddc->work_change = r; 17073e589ea0SStefano Zampini r = swap; 17083e589ea0SStefano Zampini } 17093e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 17103e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 17113e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 17123e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 17133e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 17143e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 17153e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1716f913dca9SStefano Zampini pcbddc->work_change = r; 17173e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 17183e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 17193e589ea0SStefano Zampini } 17203e589ea0SStefano Zampini PetscFunctionReturn(0); 17213e589ea0SStefano Zampini } 17223e589ea0SStefano Zampini 17233e589ea0SStefano Zampini #undef __FUNCT__ 1724a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 1725a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1726a3df083aSStefano Zampini { 1727a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1728a3df083aSStefano Zampini PetscErrorCode ierr; 1729a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1730a3df083aSStefano Zampini 1731a3df083aSStefano Zampini PetscFunctionBegin; 1732a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1733a3df083aSStefano Zampini if (transpose) { 1734a3df083aSStefano Zampini apply_right = ctx->apply_left; 1735a3df083aSStefano Zampini apply_left = ctx->apply_right; 1736a3df083aSStefano Zampini } else { 1737a3df083aSStefano Zampini apply_right = ctx->apply_right; 1738a3df083aSStefano Zampini apply_left = ctx->apply_left; 1739a3df083aSStefano Zampini } 1740a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1741a3df083aSStefano Zampini if (apply_right) { 1742a3df083aSStefano Zampini const PetscScalar *ax; 1743a3df083aSStefano Zampini PetscInt nl,i; 1744a3df083aSStefano Zampini 1745a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1746a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1747a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 1748a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1749a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1750a3df083aSStefano Zampini PetscScalar sum,val; 1751a3df083aSStefano Zampini const PetscInt *idxs; 1752a3df083aSStefano Zampini PetscInt nz,j; 1753a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1754a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1755a3df083aSStefano Zampini sum = 0.; 1756a3df083aSStefano Zampini if (ctx->apply_p0) { 1757a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1758a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1759a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1760a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1761a3df083aSStefano Zampini } 1762a3df083aSStefano Zampini } else { 1763a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1764a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1765a3df083aSStefano Zampini } 1766a3df083aSStefano Zampini } 1767a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1768a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1769a3df083aSStefano Zampini } 1770a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1771a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1772a3df083aSStefano Zampini } 1773a3df083aSStefano Zampini if (transpose) { 1774a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1775a3df083aSStefano Zampini } else { 1776a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1777a3df083aSStefano Zampini } 1778a3df083aSStefano Zampini if (reset_x) { 1779a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1780a3df083aSStefano Zampini } 1781a3df083aSStefano Zampini if (apply_left) { 1782a3df083aSStefano Zampini PetscScalar *ay; 1783a3df083aSStefano Zampini PetscInt i; 1784a3df083aSStefano Zampini 1785a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1786a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1787a3df083aSStefano Zampini PetscScalar sum,val; 1788a3df083aSStefano Zampini const PetscInt *idxs; 1789a3df083aSStefano Zampini PetscInt nz,j; 1790a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1791a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1792a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 1793a3df083aSStefano Zampini if (ctx->apply_p0) { 1794a3df083aSStefano Zampini sum = 0.; 1795a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1796a3df083aSStefano Zampini sum += ay[idxs[j]]; 1797a3df083aSStefano Zampini ay[idxs[j]] += val; 1798a3df083aSStefano Zampini } 1799a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 1800a3df083aSStefano Zampini } else { 1801a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1802a3df083aSStefano Zampini ay[idxs[j]] += val; 1803a3df083aSStefano Zampini } 1804a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 1805a3df083aSStefano Zampini } 1806a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1807a3df083aSStefano Zampini } 1808a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 1809a3df083aSStefano Zampini } 1810a3df083aSStefano Zampini PetscFunctionReturn(0); 1811a3df083aSStefano Zampini } 1812a3df083aSStefano Zampini 1813a3df083aSStefano Zampini #undef __FUNCT__ 1814a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 1815a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 1816a3df083aSStefano Zampini { 1817a3df083aSStefano Zampini PetscErrorCode ierr; 1818a3df083aSStefano Zampini 1819a3df083aSStefano Zampini PetscFunctionBegin; 1820a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 1821a3df083aSStefano Zampini PetscFunctionReturn(0); 1822a3df083aSStefano Zampini } 1823a3df083aSStefano Zampini 1824a3df083aSStefano Zampini #undef __FUNCT__ 1825a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 1826a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 1827a3df083aSStefano Zampini { 1828a3df083aSStefano Zampini PetscErrorCode ierr; 1829a3df083aSStefano Zampini 1830a3df083aSStefano Zampini PetscFunctionBegin; 1831a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 1832a3df083aSStefano Zampini PetscFunctionReturn(0); 1833a3df083aSStefano Zampini } 1834a3df083aSStefano Zampini 1835a3df083aSStefano Zampini #undef __FUNCT__ 1836a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 1837a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 1838a3df083aSStefano Zampini { 1839a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1840a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1841a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1842a3df083aSStefano Zampini PetscErrorCode ierr; 1843a3df083aSStefano Zampini 1844a3df083aSStefano Zampini PetscFunctionBegin; 1845a3df083aSStefano Zampini if (!restore) { 18461dd7afcfSStefano Zampini Mat A_IB,A_BI; 1847a3df083aSStefano Zampini PetscScalar *work; 1848b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 1849a3df083aSStefano Zampini 18509a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 18519a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 1852a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 1853a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 1854a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1855a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 1856a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 1857a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 1858a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 1859a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 1860a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 1861a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 1862a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 1863a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 1864059032f7SStefano Zampini if (reuse) { 1865a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 18661dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 1867059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 1868059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 1869059032f7SStefano Zampini PetscInt i; 1870059032f7SStefano Zampini 1871059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 1872059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1873059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1874059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1875059032f7SStefano Zampini } 1876059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 18771dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 1878059032f7SStefano Zampini } 1879a3df083aSStefano Zampini ctx->A = pcis->A_IB; 1880a3df083aSStefano Zampini ctx->work = work; 1881a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 1882a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1883a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1884a3df083aSStefano Zampini pcis->A_IB = A_IB; 1885a3df083aSStefano Zampini 1886a3df083aSStefano Zampini /* A_BI as A_IB^T */ 1887a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 1888a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 1889a3df083aSStefano Zampini pcis->A_BI = A_BI; 1890a3df083aSStefano Zampini } else { 18911dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 18921dd7afcfSStefano Zampini PetscFunctionReturn(0); 18931dd7afcfSStefano Zampini } 1894a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 1895a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 1896a3df083aSStefano Zampini pcis->A_IB = ctx->A; 18971dd7afcfSStefano Zampini ctx->A = NULL; 18981dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 18991dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 19001dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 19011dd7afcfSStefano Zampini if (ctx->free) { 1902059032f7SStefano Zampini PetscInt i; 19031dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1904059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1905059032f7SStefano Zampini } 1906059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1907059032f7SStefano Zampini } 1908a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 1909a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 1910a3df083aSStefano Zampini } 1911a3df083aSStefano Zampini PetscFunctionReturn(0); 1912a3df083aSStefano Zampini } 1913a3df083aSStefano Zampini 1914a3df083aSStefano Zampini /* used just in bddc debug mode */ 1915a3df083aSStefano Zampini #undef __FUNCT__ 1916a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 1917a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 1918a3df083aSStefano Zampini { 1919a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1920a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1921a3df083aSStefano Zampini Mat An; 1922a3df083aSStefano Zampini PetscErrorCode ierr; 1923a3df083aSStefano Zampini 1924a3df083aSStefano Zampini PetscFunctionBegin; 1925a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 1926a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 1927a3df083aSStefano Zampini if (is1) { 1928a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 1929a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 1930a3df083aSStefano Zampini } else { 1931a3df083aSStefano Zampini *B = An; 1932a3df083aSStefano Zampini } 1933a3df083aSStefano Zampini PetscFunctionReturn(0); 1934a3df083aSStefano Zampini } 1935a3df083aSStefano Zampini 19361cf9b237SStefano Zampini /* TODO: add reuse flag */ 19371cf9b237SStefano Zampini #undef __FUNCT__ 19381cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 19391cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 19401cf9b237SStefano Zampini { 19411cf9b237SStefano Zampini Mat Bt; 19421cf9b237SStefano Zampini PetscScalar *a,*bdata; 19431cf9b237SStefano Zampini const PetscInt *ii,*ij; 19441cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 19451cf9b237SStefano Zampini PetscBool flg_row; 19461cf9b237SStefano Zampini PetscErrorCode ierr; 19471cf9b237SStefano Zampini 19481cf9b237SStefano Zampini PetscFunctionBegin; 19491cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 19501cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 19511cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 19521cf9b237SStefano Zampini nnz = n; 19531cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 19541cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 19551cf9b237SStefano Zampini } 19561cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 19571cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 19581cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 19591cf9b237SStefano Zampini nnz = 0; 19601cf9b237SStefano Zampini bii[0] = 0; 19611cf9b237SStefano Zampini for (i=0;i<n;i++) { 19621cf9b237SStefano Zampini PetscInt j; 19631cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 19641cf9b237SStefano Zampini PetscScalar entry = a[j]; 19651cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 19661cf9b237SStefano Zampini bij[nnz] = ij[j]; 19671cf9b237SStefano Zampini bdata[nnz] = entry; 19681cf9b237SStefano Zampini nnz++; 19691cf9b237SStefano Zampini } 19701cf9b237SStefano Zampini } 19711cf9b237SStefano Zampini bii[i+1] = nnz; 19721cf9b237SStefano Zampini } 19731cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 19741cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 19751cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 19761cf9b237SStefano Zampini { 19771cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 19781cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 19791cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 19801cf9b237SStefano Zampini } 19811cf9b237SStefano Zampini *B = Bt; 19821cf9b237SStefano Zampini PetscFunctionReturn(0); 19831cf9b237SStefano Zampini } 19841cf9b237SStefano Zampini 1985674ae819SStefano Zampini #undef __FUNCT__ 19864f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 19874f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 19884f1b2e48SStefano Zampini { 19894f1b2e48SStefano Zampini Mat B; 19904f1b2e48SStefano Zampini IS is_dummy,*cc_n; 19914f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 19924f1b2e48SStefano Zampini PCBDDCGraph graph; 19934f1b2e48SStefano Zampini PetscInt i,n; 19944f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 19954f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 19964f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 19974f1b2e48SStefano Zampini PetscErrorCode ierr; 19984f1b2e48SStefano Zampini 19994f1b2e48SStefano Zampini PetscFunctionBegin; 200063c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 200163c961adSStefano Zampini *ncc = 0; 200263c961adSStefano Zampini *cc = NULL; 200363c961adSStefano Zampini PetscFunctionReturn(0); 200463c961adSStefano Zampini } 20054f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 20064f1b2e48SStefano Zampini if (!isseqaij && filter) { 20071cf9b237SStefano Zampini PetscBool isseqdense; 20081cf9b237SStefano Zampini 20091cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 20101cf9b237SStefano Zampini if (!isseqdense) { 20114f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 20121cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 20131cf9b237SStefano Zampini PetscScalar *array; 20141cf9b237SStefano Zampini PetscReal chop=1.e-6; 20151cf9b237SStefano Zampini 20161cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 20171cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 20181cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 20191cf9b237SStefano Zampini for (i=0;i<n;i++) { 20201cf9b237SStefano Zampini PetscInt j; 20211cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 20221cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 20231cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 20241cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 20251cf9b237SStefano Zampini } 20261cf9b237SStefano Zampini } 20271cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 20289d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 20291cf9b237SStefano Zampini } 20304f1b2e48SStefano Zampini } else { 20314f1b2e48SStefano Zampini B = A; 20324f1b2e48SStefano Zampini } 20334f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 20344f1b2e48SStefano Zampini 20354f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 20364f1b2e48SStefano Zampini if (filter) { 20374f1b2e48SStefano Zampini PetscScalar *data; 20384f1b2e48SStefano Zampini PetscInt j,cum; 20394f1b2e48SStefano Zampini 20404f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 20414f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 20424f1b2e48SStefano Zampini cum = 0; 20434f1b2e48SStefano Zampini for (i=0;i<n;i++) { 20444f1b2e48SStefano Zampini PetscInt t; 20454f1b2e48SStefano Zampini 20464f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 20474f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 20484f1b2e48SStefano Zampini continue; 20494f1b2e48SStefano Zampini } 20504f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 20514f1b2e48SStefano Zampini } 20524f1b2e48SStefano Zampini t = xadj_filtered[i]; 20534f1b2e48SStefano Zampini xadj_filtered[i] = cum; 20544f1b2e48SStefano Zampini cum += t; 20554f1b2e48SStefano Zampini } 20564f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 20574f1b2e48SStefano Zampini } else { 20584f1b2e48SStefano Zampini xadj_filtered = NULL; 20594f1b2e48SStefano Zampini adjncy_filtered = NULL; 20604f1b2e48SStefano Zampini } 20614f1b2e48SStefano Zampini 20624f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 20634f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 20644f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 20654f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 20664f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 2067be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 20684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 20694f1b2e48SStefano Zampini if (xadj_filtered) { 20704f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 20714f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 20724f1b2e48SStefano Zampini } else { 20734f1b2e48SStefano Zampini graph->xadj = xadj; 20744f1b2e48SStefano Zampini graph->adjncy = adjncy; 20754f1b2e48SStefano Zampini } 20764f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 20774f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 20784f1b2e48SStefano Zampini /* partial clean up */ 20794f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 20804f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 20811cf9b237SStefano Zampini if (A != B) { 20824f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 20834f1b2e48SStefano Zampini } 20844f1b2e48SStefano Zampini 20854f1b2e48SStefano Zampini /* get back data */ 20861cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 20871cf9b237SStefano Zampini if (cc) { 20884f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 20894f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 20904f1b2e48SStefano 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); 20914f1b2e48SStefano Zampini } 20924f1b2e48SStefano Zampini *cc = cc_n; 20931cf9b237SStefano Zampini } 20944f1b2e48SStefano Zampini /* clean up graph */ 20954f1b2e48SStefano Zampini graph->xadj = 0; 20964f1b2e48SStefano Zampini graph->adjncy = 0; 20974f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 20984f1b2e48SStefano Zampini PetscFunctionReturn(0); 20994f1b2e48SStefano Zampini } 21004f1b2e48SStefano Zampini 21014f1b2e48SStefano Zampini #undef __FUNCT__ 21025408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 21035408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 21045408967cSStefano Zampini { 21055408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 21065408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2107dee84bffSStefano Zampini IS dirIS = NULL; 21084f1b2e48SStefano Zampini PetscInt i; 21095408967cSStefano Zampini PetscErrorCode ierr; 21105408967cSStefano Zampini 21115408967cSStefano Zampini PetscFunctionBegin; 2112dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 21135408967cSStefano Zampini if (zerodiag) { 21145408967cSStefano Zampini Mat A; 21155408967cSStefano Zampini Vec vec3_N; 21165408967cSStefano Zampini PetscScalar *vals; 21175408967cSStefano Zampini const PetscInt *idxs; 2118d12d3064SStefano Zampini PetscInt nz,*count; 21195408967cSStefano Zampini 21205408967cSStefano Zampini /* p0 */ 21215408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 21225408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 21235408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 21245408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 21254f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 21265408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21275408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 21285408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 21295408967cSStefano Zampini /* v_I */ 21305408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 21315408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 21325408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21335408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 21345408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 21355408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 21365408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21375408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 21385408967cSStefano Zampini if (dirIS) { 21395408967cSStefano Zampini PetscInt n; 21405408967cSStefano Zampini 21415408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 21425408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 21435408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 21445408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21455408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 21465408967cSStefano Zampini } 21475408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 21485408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 21495408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 21505408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 2151669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 21525408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 21535408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 21549a962809SStefano 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])); 21555408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 21565408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 2157d12d3064SStefano Zampini 2158d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 2159d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 2160d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2161d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 2162d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2163d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 21649a962809SStefano 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]); 2165d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 2166d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 21675408967cSStefano Zampini } 2168dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 21695408967cSStefano Zampini 21705408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 21715408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 21724f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 21735408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 21744f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 21755408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 2176f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2177f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 2178f2a566d8SStefano Zampini if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %d instead of %g\n",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i);CHKERRQ(ierr); 2179f2a566d8SStefano Zampini } 21805408967cSStefano Zampini PetscFunctionReturn(0); 21815408967cSStefano Zampini } 21825408967cSStefano Zampini 21835408967cSStefano Zampini #undef __FUNCT__ 2184339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 2185339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 2186339f8db1SStefano Zampini { 2187339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 21884f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 2189b0f5fe93SStefano Zampini PetscInt nz,n; 21901f4df5f7SStefano Zampini PetscInt *interior_dofs,n_interior_dofs; 21914f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 2192339f8db1SStefano Zampini PetscErrorCode ierr; 2193339f8db1SStefano Zampini 2194339f8db1SStefano Zampini PetscFunctionBegin; 21959f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 21969f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 2197a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 2198a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 2199a3df083aSStefano Zampini } 2200a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 2201a3df083aSStefano Zampini pcbddc->benign_n = 0; 22021ae86dd6SStefano Zampini /* if a local info on dofs is present, assumes that the last field represents "pressures" 22034f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 22044f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 22054f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 22061ae86dd6SStefano Zampini since the local Schur complements are already SPD 22074f1b2e48SStefano Zampini */ 22084f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 22094f1b2e48SStefano Zampini have_null = PETSC_TRUE; 221040fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 22114f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 22124f1b2e48SStefano Zampini 22134f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 22144f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 22154f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 22164f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 2217ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 221840fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 221940fa8d13SStefano Zampini if (!sorted) { 222040fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 222140fa8d13SStefano Zampini } 222240fa8d13SStefano Zampini } else { 222340fa8d13SStefano Zampini pressures = NULL; 222440fa8d13SStefano Zampini } 222597d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 222697d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 222727b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 222897d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 2229339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 2230339f8db1SStefano Zampini if (!sorted) { 2231339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 2232339f8db1SStefano Zampini } 2233339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 22344f1b2e48SStefano Zampini if (!nz) { 22354f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 22364f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 223740fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 223840fa8d13SStefano Zampini } 22394f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 22404f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 22414f1b2e48SStefano Zampini zerodiag_subs = NULL; 22424f1b2e48SStefano Zampini pcbddc->benign_n = 0; 22431f4df5f7SStefano Zampini n_interior_dofs = 0; 22441f4df5f7SStefano Zampini interior_dofs = NULL; 22451f4df5f7SStefano Zampini if (pcbddc->current_level) { /* need to compute interior nodes */ 22461f4df5f7SStefano Zampini PetscInt n,i,j; 22471f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 22481f4df5f7SStefano Zampini PetscInt *iwork; 22491f4df5f7SStefano Zampini 22501f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 22511f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 22521f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 22531f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 225490648384SStefano Zampini for (i=1;i<n_neigh;i++) 22551f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 22561f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 22571f4df5f7SStefano Zampini for (i=0;i<n;i++) 22581f4df5f7SStefano Zampini if (!iwork[i]) 22591f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 22601f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 22611f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 22621f4df5f7SStefano Zampini } 22634f1b2e48SStefano Zampini if (has_null_pressures) { 22644f1b2e48SStefano Zampini IS *subs; 22651f4df5f7SStefano Zampini PetscInt nsubs,i,j,nl; 22661f4df5f7SStefano Zampini const PetscInt *idxs; 22671f4df5f7SStefano Zampini PetscScalar *array; 22681f4df5f7SStefano Zampini Vec *work; 22691f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 22704f1b2e48SStefano Zampini 22714f1b2e48SStefano Zampini subs = pcbddc->local_subs; 22724f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 22731f4df5f7SStefano 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) */ 22741f4df5f7SStefano Zampini if (pcbddc->current_level) { 22751f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 22761f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 22771f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 22781f4df5f7SStefano Zampini /* work[0] = 1_p */ 22791f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 22801f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 22811f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 22821f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 22831f4df5f7SStefano Zampini /* work[0] = 1_v */ 22841f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 22851f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 22861f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 22871f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 22881f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 22891f4df5f7SStefano Zampini } 22904f1b2e48SStefano Zampini if (nsubs > 1) { 22914f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 22924f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 22934f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 22944f1b2e48SStefano Zampini IS t_zerodiag_subs; 22954f1b2e48SStefano Zampini PetscInt nl; 22964f1b2e48SStefano Zampini 22974f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 22984f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 22994f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 23004f1b2e48SStefano Zampini if (nl) { 23014f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 23024f1b2e48SStefano Zampini 23031f4df5f7SStefano Zampini if (pcbddc->current_level) { 23041f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 23051f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 23061f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 23071f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 23081f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 23091f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 23101f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 23111f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 23121f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 23131f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 23141f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 23151f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 23161f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 23171f4df5f7SStefano Zampini valid = PETSC_FALSE; 23181f4df5f7SStefano Zampini break; 23191f4df5f7SStefano Zampini } 23201f4df5f7SStefano Zampini } 23211f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 23221f4df5f7SStefano Zampini } 23231f4df5f7SStefano Zampini if (valid && pcbddc->NeumannBoundariesLocal) { 23241f4df5f7SStefano Zampini IS t_bc; 23251f4df5f7SStefano Zampini PetscInt nzb; 23261f4df5f7SStefano Zampini 23271f4df5f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pcbddc->NeumannBoundariesLocal,&t_bc);CHKERRQ(ierr); 23281f4df5f7SStefano Zampini ierr = ISGetLocalSize(t_bc,&nzb);CHKERRQ(ierr); 23291f4df5f7SStefano Zampini ierr = ISDestroy(&t_bc);CHKERRQ(ierr); 23301f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 23311f4df5f7SStefano Zampini } 23321f4df5f7SStefano Zampini if (valid && pressures) { 23334f1b2e48SStefano Zampini IS t_pressure_subs; 23344f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 23354f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 23364f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 23374f1b2e48SStefano Zampini } 23384f1b2e48SStefano Zampini if (valid) { 23394f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 23404f1b2e48SStefano Zampini pcbddc->benign_n++; 23414f1b2e48SStefano Zampini } else { 23424f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 23434f1b2e48SStefano Zampini } 23444f1b2e48SStefano Zampini } 23454f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 23464f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 23474f1b2e48SStefano Zampini } 23484f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 23494f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 23501f4df5f7SStefano Zampini 23511f4df5f7SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 23521f4df5f7SStefano Zampini PetscInt nzb; 23531f4df5f7SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nzb);CHKERRQ(ierr); 23541f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 23551f4df5f7SStefano Zampini } 23561f4df5f7SStefano Zampini if (valid && pressures) { 23574f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 23584f1b2e48SStefano Zampini } 23591f4df5f7SStefano Zampini if (valid && pcbddc->current_level) { 23601f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 23611f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 23621f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 23631f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 23641f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 23651f4df5f7SStefano Zampini valid = PETSC_FALSE; 23661f4df5f7SStefano Zampini break; 23671f4df5f7SStefano Zampini } 23681f4df5f7SStefano Zampini } 23691f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 23701f4df5f7SStefano Zampini } 23714f1b2e48SStefano Zampini if (valid) { 23724f1b2e48SStefano Zampini pcbddc->benign_n = 1; 2373ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 23744f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 23754f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 23764f1b2e48SStefano Zampini } 23774f1b2e48SStefano Zampini } 23781f4df5f7SStefano Zampini if (pcbddc->current_level) { 23791f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 23804f1b2e48SStefano Zampini } 23811f4df5f7SStefano Zampini } 23821f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 23834f1b2e48SStefano Zampini 23844f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2385b9b0e38cSStefano Zampini PetscInt n; 2386b9b0e38cSStefano Zampini 23874f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 23884f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2389b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2390b9b0e38cSStefano Zampini if (n) { 23914f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 23924f1b2e48SStefano Zampini have_null = PETSC_FALSE; 23934f1b2e48SStefano Zampini } 2394b9b0e38cSStefano Zampini } 23954f1b2e48SStefano Zampini 23964f1b2e48SStefano Zampini /* final check for null pressures */ 23974f1b2e48SStefano Zampini if (zerodiag && pressures) { 23984f1b2e48SStefano Zampini PetscInt nz,np; 23994f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 24004f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 24014f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 24024f1b2e48SStefano Zampini } 24034f1b2e48SStefano Zampini 24044f1b2e48SStefano Zampini if (recompute_zerodiag) { 24054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 24064f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 24074f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 24084f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 24094f1b2e48SStefano Zampini } else { 24104f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 24114f1b2e48SStefano Zampini 24124f1b2e48SStefano Zampini nzn = 0; 24134f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 24144f1b2e48SStefano Zampini PetscInt ns; 24154f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 24164f1b2e48SStefano Zampini nzn += ns; 24174f1b2e48SStefano Zampini } 24184f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 24194f1b2e48SStefano Zampini nzn = 0; 24204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 24214f1b2e48SStefano Zampini PetscInt ns,*idxs; 24224f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 24234f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 24244f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 24254f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 24264f1b2e48SStefano Zampini nzn += ns; 24274f1b2e48SStefano Zampini } 24284f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 24294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 24304f1b2e48SStefano Zampini } 24314f1b2e48SStefano Zampini have_null = PETSC_FALSE; 24324f1b2e48SStefano Zampini } 24334f1b2e48SStefano Zampini 2434669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2435a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2436a198735bSStefano Zampini Mat A,loc_divudotp; 2437a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2438a198735bSStefano Zampini IS row,col,isused = NULL; 2439a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2440a198735bSStefano Zampini 24411f4df5f7SStefano Zampini if (pressures) { 24421f4df5f7SStefano Zampini isused = pressures; 24431f4df5f7SStefano Zampini } else { 24441f4df5f7SStefano Zampini isused = zerodiag; 24451f4df5f7SStefano Zampini } 2446a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2447669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 24481ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 24491ae86dd6SStefano 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"); 2450a198735bSStefano Zampini n_isused = 0; 2451a198735bSStefano Zampini if (isused) { 2452a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2453a198735bSStefano Zampini } 2454a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2455a198735bSStefano Zampini st = st-n_isused; 24561ae86dd6SStefano Zampini if (n) { 2457a198735bSStefano Zampini const PetscInt *gidxs; 2458a198735bSStefano Zampini 2459a198735bSStefano Zampini ierr = MatGetSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2460a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2461a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2462a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2463a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2464a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 24651ae86dd6SStefano Zampini } else { 2466a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2467a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2468a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2469a198735bSStefano Zampini } 2470a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2471a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2472a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2473a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2474a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2475a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2476a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2477a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2478a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2479a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2480a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2481a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2482a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2483a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 24841ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 24851ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 24861ae86dd6SStefano Zampini } 2487b3afcdbeSStefano Zampini 2488b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 24894f1b2e48SStefano Zampini if (has_null_pressures) { 24904f1b2e48SStefano Zampini IS zerodiagc; 24914f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 24924f1b2e48SStefano Zampini PetscInt i,s,*nnz; 24934f1b2e48SStefano Zampini 24944f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2495339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 2496339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 2497339f8db1SStefano Zampini /* local change of basis for pressures */ 2498339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 249997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2500339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2501339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2502339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 25034f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 25044f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 25054f1b2e48SStefano Zampini PetscInt nzs,j; 25064f1b2e48SStefano Zampini 25074f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 25084f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 25094f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 25104f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 25114f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 25124f1b2e48SStefano Zampini } 2513339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2514339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2515339f8db1SStefano Zampini /* set identity on velocities */ 2516339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 2517339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 2518339f8db1SStefano Zampini } 25194f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 25204f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 25219f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 25224f1b2e48SStefano 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); 2523339f8db1SStefano Zampini /* set change on pressures */ 25244f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 25254f1b2e48SStefano Zampini PetscScalar *array; 25264f1b2e48SStefano Zampini PetscInt nzs; 25274f1b2e48SStefano Zampini 25284f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 25294f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 25304f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2531339f8db1SStefano Zampini PetscScalar vals[2]; 2532339f8db1SStefano Zampini PetscInt cols[2]; 2533339f8db1SStefano Zampini 2534339f8db1SStefano Zampini cols[0] = idxs[i]; 25354f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2536339f8db1SStefano Zampini vals[0] = 1.; 2537b0f5fe93SStefano Zampini vals[1] = 1.; 25384f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2539339f8db1SStefano Zampini } 25404f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 25414f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 25424f1b2e48SStefano Zampini array[nzs-1] = 1.; 25434f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 25444f1b2e48SStefano Zampini /* store local idxs for p0 */ 25454f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 25464f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2547339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 25484f1b2e48SStefano Zampini } 2549339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2550339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2551a3df083aSStefano Zampini /* project if needed */ 2552a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 25531dd7afcfSStefano Zampini Mat M; 25541dd7afcfSStefano Zampini 25551dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2556339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25571dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25581dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2559a3df083aSStefano Zampini } 25604f1b2e48SStefano Zampini /* store global idxs for p0 */ 25614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2562339f8db1SStefano Zampini } 2563ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 25644f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 2565b0f5fe93SStefano Zampini 2566b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 2567b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 256827b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 256927b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2570339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2571339f8db1SStefano Zampini PetscFunctionReturn(0); 2572339f8db1SStefano Zampini } 2573339f8db1SStefano Zampini 2574339f8db1SStefano Zampini #undef __FUNCT__ 2575015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 2576015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 2577efc2fbd9SStefano Zampini { 2578efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2579de9d7bd0SStefano Zampini PetscScalar *array; 2580efc2fbd9SStefano Zampini PetscErrorCode ierr; 2581efc2fbd9SStefano Zampini 2582efc2fbd9SStefano Zampini PetscFunctionBegin; 2583efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 2584efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 25854f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2586efc2fbd9SStefano Zampini } 2587de9d7bd0SStefano Zampini if (get) { 2588efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 25894f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 25904f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 2591efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 2592de9d7bd0SStefano Zampini } else { 2593de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 2594de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2595de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2596de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 2597efc2fbd9SStefano Zampini } 2598efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2599efc2fbd9SStefano Zampini } 2600efc2fbd9SStefano Zampini 2601efc2fbd9SStefano Zampini #undef __FUNCT__ 2602c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 2603c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 2604c263805aSStefano Zampini { 2605c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2606c263805aSStefano Zampini PetscErrorCode ierr; 2607c263805aSStefano Zampini 2608c263805aSStefano Zampini PetscFunctionBegin; 2609c263805aSStefano Zampini /* TODO: add error checking 2610c263805aSStefano Zampini - avoid nested pop (or push) calls. 2611c263805aSStefano Zampini - cannot push before pop. 26121c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 2613c263805aSStefano Zampini */ 26144f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2615efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2616efc2fbd9SStefano Zampini } 2617c263805aSStefano Zampini if (pop) { 2618a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 26194f1b2e48SStefano Zampini IS is_p0; 26204f1b2e48SStefano Zampini MatReuse reuse; 2621c263805aSStefano Zampini 2622c263805aSStefano Zampini /* extract B_0 */ 26234f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 26244f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 26254f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 26264f1b2e48SStefano Zampini } 26274f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 26284f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 2629c263805aSStefano Zampini /* remove rows and cols from local problem */ 2630c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 263197d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 26324f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 26334f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 2634a3df083aSStefano Zampini } else { 2635a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2636a3df083aSStefano Zampini PetscScalar *vals; 2637a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 2638a3df083aSStefano Zampini 2639a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 2640a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 2641a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 26420b5adadeSStefano Zampini PetscInt *nnz; 2643a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 2644a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 2645a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2646331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 2647331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2648331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 2649331e053bSStefano Zampini nnz[i] = n - nnz[i]; 2650331e053bSStefano Zampini } 2651331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 2652331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2653331e053bSStefano Zampini } 2654a3df083aSStefano Zampini 2655a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2656a3df083aSStefano Zampini PetscScalar *array; 2657a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 2658a3df083aSStefano Zampini 2659a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 2660a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2661a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2662a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 2663a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2664a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 2665a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 2666a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 2667a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 2668a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 2669a3df083aSStefano Zampini cum = 0; 2670a3df083aSStefano Zampini for (j=0;j<n;j++) { 267122db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 2672a3df083aSStefano Zampini vals[cum] = array[j]; 2673a3df083aSStefano Zampini idxs_ins[cum] = j; 2674a3df083aSStefano Zampini cum++; 2675a3df083aSStefano Zampini } 2676a3df083aSStefano Zampini } 2677a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 2678a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 2679a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2680a3df083aSStefano Zampini } 2681a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2682a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2683a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 2684a3df083aSStefano Zampini } 2685c263805aSStefano Zampini } else { /* push */ 2686a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 26874f1b2e48SStefano Zampini PetscInt i; 26884f1b2e48SStefano Zampini 26894f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 26904f1b2e48SStefano Zampini PetscScalar *B0_vals; 26914f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 26924f1b2e48SStefano Zampini 26934f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 26944f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 26957b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 26964f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 26974f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 26984f1b2e48SStefano Zampini } 2699c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2700c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2701a3df083aSStefano Zampini } else { 2702a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 2703a3df083aSStefano Zampini } 2704c263805aSStefano Zampini } 2705c263805aSStefano Zampini PetscFunctionReturn(0); 2706c263805aSStefano Zampini } 2707c263805aSStefano Zampini 2708c263805aSStefano Zampini #undef __FUNCT__ 2709b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 271008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 2711b1b3d7a2SStefano Zampini { 2712b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 271308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 271408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 271508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 271608122e43SStefano Zampini PetscScalar *work,lwork; 271708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 271808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 271908122e43SStefano Zampini PetscReal *eigs,thresh; 27201b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 2721f6f667cfSStefano Zampini PetscBool allocated_S_St; 272208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 272308122e43SStefano Zampini PetscReal *rwork; 272408122e43SStefano Zampini #endif 2725b1b3d7a2SStefano Zampini PetscErrorCode ierr; 2726b1b3d7a2SStefano Zampini 2727b1b3d7a2SStefano Zampini PetscFunctionBegin; 2728b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 2729af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 2730af25d912SStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 273106a4e24aSStefano Zampini 2732fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2733fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2734fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2735fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 27361575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2737fd14bc51SStefano Zampini } 2738fd14bc51SStefano Zampini 2739e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 2740e496cd5dSStefano 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); 2741e496cd5dSStefano Zampini } 2742e496cd5dSStefano Zampini 274308122e43SStefano Zampini /* max size of subsets */ 274408122e43SStefano Zampini mss = 0; 274508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 274608122e43SStefano Zampini PetscInt subset_size; 2747862806e4SStefano Zampini 274808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 274908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 275008122e43SStefano Zampini } 275108122e43SStefano Zampini 275208122e43SStefano Zampini /* min/max and threshold */ 275308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 2754f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 275508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 2756f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 2757f6f667cfSStefano Zampini if (nmin) { 2758f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 2759f6f667cfSStefano Zampini } 276008122e43SStefano Zampini 276108122e43SStefano Zampini /* allocate lapack workspace */ 276208122e43SStefano Zampini cum = cum2 = 0; 276308122e43SStefano Zampini maxneigs = 0; 276408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 276508122e43SStefano Zampini PetscInt n,subset_size; 2766f6f667cfSStefano Zampini 276708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 276808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 27699162d606SStefano Zampini cum += subset_size; 27709162d606SStefano Zampini cum2 += subset_size*n; 277108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 277208122e43SStefano Zampini } 277308122e43SStefano Zampini if (mss) { 27749ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 277508122e43SStefano Zampini PetscBLASInt B_itype = 1; 277608122e43SStefano Zampini PetscBLASInt B_N = mss; 27774c6709b3SStefano Zampini PetscReal zero = 0.0; 27784c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 277908122e43SStefano Zampini 278008122e43SStefano Zampini B_lwork = -1; 278108122e43SStefano Zampini S = NULL; 278208122e43SStefano Zampini St = NULL; 2783a58a30b4SStefano Zampini eigs = NULL; 2784a58a30b4SStefano Zampini eigv = NULL; 2785a58a30b4SStefano Zampini B_iwork = NULL; 2786a58a30b4SStefano Zampini B_ifail = NULL; 2787d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2788d1710679SStefano Zampini rwork = NULL; 2789d1710679SStefano Zampini #endif 27908bec7fa6SStefano Zampini thresh = 1.0; 279108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 279208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 279308122e43SStefano 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)); 279408122e43SStefano Zampini #else 279508122e43SStefano 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)); 279608122e43SStefano Zampini #endif 279708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 279808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 279908122e43SStefano Zampini } else { 280008122e43SStefano Zampini /* TODO */ 280108122e43SStefano Zampini } 280208122e43SStefano Zampini } else { 280308122e43SStefano Zampini lwork = 0; 280408122e43SStefano Zampini } 280508122e43SStefano Zampini 280608122e43SStefano Zampini nv = 0; 2807d62866d3SStefano 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) */ 2808d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 280908122e43SStefano Zampini } 28104c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 2811f6f667cfSStefano Zampini if (allocated_S_St) { 2812f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 2813f6f667cfSStefano Zampini } 2814f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 281508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 281608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 281708122e43SStefano Zampini #endif 28189162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 28199162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 28209162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 282108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 28229162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 282308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 282408122e43SStefano Zampini 282508122e43SStefano Zampini maxneigs = 0; 282672b8c272SStefano Zampini cum = cumarray = 0; 28279162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 28289162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 2829d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 283008122e43SStefano Zampini const PetscInt *idxs; 283108122e43SStefano Zampini 2832d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 283308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 283408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 283508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 283608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 28379162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 28389162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 283908122e43SStefano Zampini } 2840d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 284108122e43SStefano Zampini } 284208122e43SStefano Zampini 284308122e43SStefano Zampini if (mss) { /* multilevel */ 284408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 284508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 284608122e43SStefano Zampini } 284708122e43SStefano Zampini 2848ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 284908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 285008122e43SStefano Zampini const PetscInt *idxs; 28519d54b7f4SStefano Zampini PetscReal upper,lower; 2852862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 285308122e43SStefano Zampini PetscBLASInt B_N; 2854aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 285508122e43SStefano Zampini 28569d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 28579d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 28589d54b7f4SStefano Zampini lower = thresh; 28599d54b7f4SStefano Zampini } else { 28609d54b7f4SStefano Zampini upper = 1./thresh; 28619d54b7f4SStefano Zampini lower = 0.; 28629d54b7f4SStefano Zampini } 2863862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 2864ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 2865f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 2866f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 28679ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 2868aff50787SStefano Zampini PetscInt j,k; 2869aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 2870aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 2871aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 287208122e43SStefano Zampini } 287308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 2874aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 2875aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 2876aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 2877aff50787SStefano Zampini } 287808122e43SStefano Zampini } 287908122e43SStefano Zampini } else { 288008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 288108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 288208122e43SStefano Zampini } 28838bec7fa6SStefano Zampini } else { 2884f6f667cfSStefano Zampini S = Sarray + cumarray; 2885f6f667cfSStefano Zampini St = Starray + cumarray; 28868bec7fa6SStefano Zampini } 2887aff50787SStefano Zampini /* see if we can save some work */ 2888b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 2889aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 2890aff50787SStefano Zampini } 2891aff50787SStefano Zampini 2892b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 2893aff50787SStefano Zampini B_neigs = 0; 2894aff50787SStefano Zampini } else { 28959ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 289608122e43SStefano Zampini PetscBLASInt B_itype = 1; 2897f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 28984c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 28999552c7c7SStefano Zampini PetscInt nmin_s; 2900b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 290108122e43SStefano Zampini 2902fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2903eee23b56SStefano 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]]); 2904fd14bc51SStefano Zampini } 2905d16cbb6bSStefano Zampini 2906b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 2907b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 2908b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 2909b7ab4a40SStefano Zampini } 2910b7ab4a40SStefano Zampini 291108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2912b7ab4a40SStefano Zampini if (compute_range) { 2913d16cbb6bSStefano Zampini 2914d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 291508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 29169d54b7f4SStefano 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)); 291708122e43SStefano Zampini #else 29189d54b7f4SStefano 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)); 291908122e43SStefano Zampini #endif 2920b7ab4a40SStefano Zampini } else if (!same_data) { 2921d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 2922d16cbb6bSStefano Zampini B_IL = 1; 2923d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 29249d54b7f4SStefano 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)); 2925d16cbb6bSStefano Zampini #else 29269d54b7f4SStefano 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)); 2927d16cbb6bSStefano Zampini #endif 2928b03ebc13SStefano Zampini } else { /* same_data is true, so just get the adaptive functional requested by the user */ 2929b7ab4a40SStefano Zampini PetscInt k; 2930b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 2931b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 2932b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 2933b7ab4a40SStefano Zampini nmin = nmax; 2934b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 2935b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 2936b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 2937b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 2938b7ab4a40SStefano Zampini } 2939d16cbb6bSStefano Zampini } 294008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 294108122e43SStefano Zampini if (B_ierr) { 29426c4ed002SBarry 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); 29436c4ed002SBarry 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); 29446c4ed002SBarry 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); 294508122e43SStefano Zampini } 294608122e43SStefano Zampini 294708122e43SStefano Zampini if (B_neigs > nmax) { 2948fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2949fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 2950fd14bc51SStefano Zampini } 29519d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 295208122e43SStefano Zampini B_neigs = nmax; 295308122e43SStefano Zampini } 295408122e43SStefano Zampini 29559552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 29569552c7c7SStefano Zampini if (B_neigs < nmin_s) { 295708122e43SStefano Zampini PetscBLASInt B_neigs2; 295808122e43SStefano Zampini 29599d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2960f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 29619d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 29629d54b7f4SStefano Zampini } else { 29639d54b7f4SStefano Zampini B_IL = B_neigs + 1; 29649d54b7f4SStefano Zampini B_IU = nmin_s; 29659d54b7f4SStefano Zampini } 2966fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2967fd14bc51SStefano 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); 2968fd14bc51SStefano Zampini } 29699ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 29701ae86dd6SStefano Zampini PetscInt j,k; 297108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 29721ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 29731ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 29741ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 297508122e43SStefano Zampini } 297608122e43SStefano Zampini } 297708122e43SStefano Zampini } else { 297808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 297908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 298008122e43SStefano Zampini } 298108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 298208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 29839d54b7f4SStefano 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)); 298408122e43SStefano Zampini #else 29859d54b7f4SStefano 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)); 298608122e43SStefano Zampini #endif 298708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 298808122e43SStefano Zampini B_neigs += B_neigs2; 298908122e43SStefano Zampini } 299008122e43SStefano Zampini if (B_ierr) { 29916c4ed002SBarry 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); 29926c4ed002SBarry 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); 29936c4ed002SBarry 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); 299408122e43SStefano Zampini } 2995fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2996ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 299708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 299808122e43SStefano Zampini if (eigs[j] == 0.0) { 2999ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 300008122e43SStefano Zampini } else { 30019d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3002ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 30039d54b7f4SStefano Zampini } else { 30049d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 30059d54b7f4SStefano Zampini } 3006fd14bc51SStefano Zampini } 300708122e43SStefano Zampini } 300808122e43SStefano Zampini } 300908122e43SStefano Zampini } else { 301008122e43SStefano Zampini /* TODO */ 301108122e43SStefano Zampini } 3012aff50787SStefano Zampini } 30136c3e6151SStefano Zampini /* change the basis back to the original one */ 30146c3e6151SStefano Zampini if (sub_schurs->change) { 301572b8c272SStefano Zampini Mat change,phi,phit; 30166c3e6151SStefano Zampini 30176c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 30186c3e6151SStefano Zampini PetscInt ii; 30196c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 30206c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 30216c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 3022684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3023684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 3024684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 3025684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3026684229deSStefano Zampini #else 30276c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 3028684229deSStefano Zampini #endif 30296c3e6151SStefano Zampini } 30306c3e6151SStefano Zampini } 30316c3e6151SStefano Zampini } 303272b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 30336c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 303472b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 30356c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 30366c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 30376c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 30386c3e6151SStefano Zampini } 30398bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 30408bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 30419162d606SStefano Zampini if (B_neigs) { 30429162d606SStefano 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); 3043fd14bc51SStefano Zampini 3044fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 30459552c7c7SStefano Zampini PetscInt ii; 30469552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 3047ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 30489552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 3049ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3050ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3051ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3052ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3053ac47001eSStefano Zampini #else 3054ac47001eSStefano 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); 3055ac47001eSStefano Zampini #endif 30569552c7c7SStefano Zampini } 30579552c7c7SStefano Zampini } 3058fd14bc51SStefano Zampini } 30599162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 30609162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 30619162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 30629162d606SStefano Zampini cum++; 306308122e43SStefano Zampini } 306408122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 306508122e43SStefano Zampini /* shift for next computation */ 306608122e43SStefano Zampini cumarray += subset_size*subset_size; 306708122e43SStefano Zampini } 3068fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3069fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3070fd14bc51SStefano Zampini } 307108122e43SStefano Zampini 307208122e43SStefano Zampini if (mss) { 307308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 307408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 3075f6f667cfSStefano Zampini /* destroy matrices (junk) */ 3076f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 3077f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 307808122e43SStefano Zampini } 3079f6f667cfSStefano Zampini if (allocated_S_St) { 3080f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 3081f6f667cfSStefano Zampini } 3082f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 308308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 308408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 308508122e43SStefano Zampini #endif 308608122e43SStefano Zampini if (pcbddc->dbg_flag) { 30871b968477SStefano Zampini PetscInt maxneigs_r; 3088b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 30899b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 309008122e43SStefano Zampini } 309108122e43SStefano Zampini PetscFunctionReturn(0); 309208122e43SStefano Zampini } 3093b1b3d7a2SStefano Zampini 3094674ae819SStefano Zampini #undef __FUNCT__ 3095c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 3096c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 3097c8587f34SStefano Zampini { 30988629588bSStefano Zampini PetscScalar *coarse_submat_vals; 3099c8587f34SStefano Zampini PetscErrorCode ierr; 3100c8587f34SStefano Zampini 3101c8587f34SStefano Zampini PetscFunctionBegin; 3102f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 31035e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 3104c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 3105c8587f34SStefano Zampini 3106684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 31070fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 3108684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3109c8587f34SStefano Zampini 31108629588bSStefano Zampini /* 31118629588bSStefano Zampini Setup local correction and local part of coarse basis. 31128629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 31138629588bSStefano Zampini */ 311447f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 31158629588bSStefano Zampini 31168629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 31178629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 31188629588bSStefano Zampini 31198629588bSStefano Zampini /* free */ 31208629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 3121c8587f34SStefano Zampini PetscFunctionReturn(0); 3122c8587f34SStefano Zampini } 3123c8587f34SStefano Zampini 3124c8587f34SStefano Zampini #undef __FUNCT__ 3125674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 3126674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 3127674ae819SStefano Zampini { 3128674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3129674ae819SStefano Zampini PetscErrorCode ierr; 3130674ae819SStefano Zampini 3131674ae819SStefano Zampini PetscFunctionBegin; 3132674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 313330368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3134674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 3135785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3136674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 3137f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3138f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3139785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 314063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 314163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 3142674ae819SStefano Zampini PetscFunctionReturn(0); 3143674ae819SStefano Zampini } 3144674ae819SStefano Zampini 3145674ae819SStefano Zampini #undef __FUNCT__ 3146674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 3147674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 3148674ae819SStefano Zampini { 3149674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 31504f1b2e48SStefano Zampini PetscInt i; 3151674ae819SStefano Zampini PetscErrorCode ierr; 3152674ae819SStefano Zampini 3153674ae819SStefano Zampini PetscFunctionBegin; 31541e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 31551e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 3156a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 3157b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3158674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 315916909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 31601dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 3161674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3162669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 3163fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 3164a13144ffSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3165674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 31664f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 31674f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 31684f1b2e48SStefano Zampini } 31694f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 3170b334f244SStefano Zampini if (pcbddc->sub_schurs) { 3171b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 3172b334f244SStefano Zampini } 3173c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 31748af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 3175674ae819SStefano Zampini PetscFunctionReturn(0); 3176674ae819SStefano Zampini } 3177674ae819SStefano Zampini 3178674ae819SStefano Zampini #undef __FUNCT__ 3179674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 3180674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 3181674ae819SStefano Zampini { 3182674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3183674ae819SStefano Zampini PetscErrorCode ierr; 3184674ae819SStefano Zampini 3185674ae819SStefano Zampini PetscFunctionBegin; 3186674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 318758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 3188ca92afb2SStefano Zampini PetscScalar *array; 318906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 319006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 319158da7f69SStefano Zampini } 3192674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3193674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 319415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 319515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3196674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 3197674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 3198674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 319906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 3200674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3201674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 32028ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 3203674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 3204674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 3205674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 3206f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 3207f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 3208f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 3209f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3210727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 32110e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3212f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 321370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 321481d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 32150369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 32161dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 32174f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 32188b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 3219ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 3220ca92afb2SStefano Zampini PetscInt i; 3221ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3222ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 3223ca92afb2SStefano Zampini } 3224ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 3225ca92afb2SStefano Zampini } 32264f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 3227674ae819SStefano Zampini PetscFunctionReturn(0); 3228674ae819SStefano Zampini } 3229674ae819SStefano Zampini 3230674ae819SStefano Zampini #undef __FUNCT__ 3231f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 3232f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 32336bfb1811SStefano Zampini { 32346bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 32356bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 32366bfb1811SStefano Zampini VecType impVecType; 32374f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 32386bfb1811SStefano Zampini PetscErrorCode ierr; 32396bfb1811SStefano Zampini 32406bfb1811SStefano Zampini PetscFunctionBegin; 32414f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 3242b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 32436bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 3244e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 3245e7b262bdSStefano Zampini /* R nodes */ 3246e7b262bdSStefano Zampini old_size = -1; 3247e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 3248e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 3249e7b262bdSStefano Zampini } 3250e7b262bdSStefano Zampini if (n_R != old_size) { 3251e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3252e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 32536bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 32546bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 32556bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 32566bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 3257e7b262bdSStefano Zampini } 3258e7b262bdSStefano Zampini /* local primal dofs */ 3259e7b262bdSStefano Zampini old_size = -1; 3260e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 3261e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 3262e7b262bdSStefano Zampini } 3263e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 3264e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 326583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 3266e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 32676bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 3268e7b262bdSStefano Zampini } 3269e7b262bdSStefano Zampini /* local explicit constraints */ 3270e7b262bdSStefano Zampini old_size = -1; 3271e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 3272e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 3273e7b262bdSStefano Zampini } 3274e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 3275e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 327683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 327783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 327883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 327983b7ccabSStefano Zampini } 32806bfb1811SStefano Zampini PetscFunctionReturn(0); 32816bfb1811SStefano Zampini } 32826bfb1811SStefano Zampini 32836bfb1811SStefano Zampini #undef __FUNCT__ 328447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 328547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 328688ebb749SStefano Zampini { 328725084f0cSStefano Zampini PetscErrorCode ierr; 328825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 328988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 329088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3291d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 329225084f0cSStefano Zampini /* submatrices of local problem */ 329380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 329406656605SStefano Zampini /* submatrices of local coarse problem */ 329506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 329625084f0cSStefano Zampini /* working matrices */ 329706656605SStefano Zampini Mat C_CR; 329825084f0cSStefano Zampini /* additional working stuff */ 329906656605SStefano Zampini PC pc_R; 33004f1b2e48SStefano Zampini Mat F; 33015cbda25cSStefano Zampini Vec dummy_vec; 3302a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 330325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 330406656605SStefano Zampini PetscScalar *work; 330506656605SStefano Zampini PetscInt *idx_V_B; 3306ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 330706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 3308ffd830a3SStefano Zampini 330925084f0cSStefano Zampini /* some shortcuts to scalars */ 331006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 331188ebb749SStefano Zampini 331288ebb749SStefano Zampini PetscFunctionBegin; 33139a962809SStefano 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"); 3314ffd830a3SStefano Zampini 3315ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 3316b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 33174f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 3318b371cd4fSStefano Zampini n_B = pcis->n_B; 3319b371cd4fSStefano Zampini n_D = pcis->n - n_B; 332088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 332188ebb749SStefano Zampini 332288ebb749SStefano Zampini /* vertices in boundary numbering */ 3323785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 33240e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 33256c4ed002SBarry 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); 332688ebb749SStefano Zampini 332706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 3328019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 332906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 333006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 333106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 333206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 333306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 333406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 333506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 333606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 333706656605SStefano Zampini 333806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 333906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 334006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 334106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 334206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 3343ffd830a3SStefano Zampini lda_rhs = n_R; 3344a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 334506656605SStefano Zampini if (isLU || isILU || isCHOL) { 334606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 3347b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 3348df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3349d62866d3SStefano Zampini MatFactorType type; 3350d62866d3SStefano Zampini 3351df4d28bfSStefano Zampini F = reuse_solver->F; 33526816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 3353d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 3354ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 335522db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 335606656605SStefano Zampini } else { 335706656605SStefano Zampini F = NULL; 335806656605SStefano Zampini } 335906656605SStefano Zampini 3360ffd830a3SStefano Zampini /* allocate workspace */ 3361ffd830a3SStefano Zampini n = 0; 3362ffd830a3SStefano Zampini if (n_constraints) { 3363ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 3364ffd830a3SStefano Zampini } 3365ffd830a3SStefano Zampini if (n_vertices) { 3366ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 3367ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 3368ffd830a3SStefano Zampini } 3369ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 3370ffd830a3SStefano Zampini 33715cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 33725cbda25cSStefano Zampini dummy_vec = NULL; 33735cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 33745cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 33755cbda25cSStefano Zampini } 33765cbda25cSStefano Zampini 337788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 337888ebb749SStefano Zampini if (n_constraints) { 337972b8c272SStefano Zampini Mat M1,M2,M3,C_B; 338006656605SStefano Zampini IS is_aux; 338180677318SStefano Zampini PetscScalar *array,*array2; 338206656605SStefano Zampini 3383f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 338480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 338588ebb749SStefano Zampini 338625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 338725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 33888ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 338972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 339088ebb749SStefano Zampini 339180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 339280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 3393ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 339488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 339506656605SStefano Zampini const PetscScalar *row_cmat_values; 339606656605SStefano Zampini const PetscInt *row_cmat_indices; 339706656605SStefano Zampini PetscInt size_of_constraint,j; 339888ebb749SStefano Zampini 339906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 340006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 3401ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 340206656605SStefano Zampini } 340306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 340406656605SStefano Zampini } 3405ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 340606656605SStefano Zampini if (F) { 340706656605SStefano Zampini Mat B; 340806656605SStefano Zampini 3409ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 3410a3df083aSStefano Zampini if (need_benign_correction) { 3411df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3412a3df083aSStefano Zampini 341372b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 341472b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 3415a3df083aSStefano Zampini } 341680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 3417a3df083aSStefano Zampini if (need_benign_correction) { 3418a3df083aSStefano Zampini PetscScalar *marr; 3419df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3420a3df083aSStefano Zampini 3421a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 34225cbda25cSStefano Zampini if (lda_rhs != n_R) { 34235cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 34245cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 34255cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 34265cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 34275cbda25cSStefano Zampini } 34285cbda25cSStefano Zampini } else { 3429a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 3430a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 34315cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3432a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3433a3df083aSStefano Zampini } 34345cbda25cSStefano Zampini } 3435a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 3436a3df083aSStefano Zampini } 343706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 343806656605SStefano Zampini } else { 343980677318SStefano Zampini PetscScalar *marr; 344080677318SStefano Zampini 344180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 344206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 3443ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3444ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 344506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 344606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 344706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 344806656605SStefano Zampini } 344980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 345006656605SStefano Zampini } 345180677318SStefano Zampini if (!pcbddc->switch_static) { 345280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 345380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 345480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 345580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 3456ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 345780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 345880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 345980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 346080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 346180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 346280677318SStefano Zampini } 346380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 346480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 346572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 346680677318SStefano Zampini } else { 3467ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3468ffd830a3SStefano Zampini IS dummy; 3469ffd830a3SStefano Zampini 3470ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 347172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 3472ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 3473ffd830a3SStefano Zampini } else { 347480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 347580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 3476ffd830a3SStefano Zampini } 347725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 347880677318SStefano Zampini } 347980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 348080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 348180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 348206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 348306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 348480677318SStefano Zampini if (isCHOL) { 348580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 348680677318SStefano Zampini } else { 348725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 348880677318SStefano Zampini } 348980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 349006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 349125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 349225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 349325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 349480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 349572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 349672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 349706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 349806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 3499f4ddd8eeSStefano Zampini } 3500fc227af8SStefano Zampini 3501fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 350288ebb749SStefano Zampini if (n_vertices) { 350306656605SStefano Zampini IS is_aux; 35043a50541eSStefano Zampini 3505b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 35066816873aSStefano Zampini IS tis; 35076816873aSStefano Zampini 35086816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 35096816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 35106816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 35116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 35126816873aSStefano Zampini } else { 35133a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 35146816873aSStefano Zampini } 35159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 35169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 351704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 351825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 351988ebb749SStefano Zampini } 352088ebb749SStefano Zampini 352188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 3522f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 352306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 352406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 352506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 352606656605SStefano Zampini } 3527f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 352806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 352906656605SStefano Zampini PetscScalar *marray; 353006656605SStefano Zampini 353106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 353206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 3533f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3534f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3535f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3536f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3537f4ddd8eeSStefano Zampini } 3538f4ddd8eeSStefano Zampini } 353906656605SStefano Zampini 3540f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 354106656605SStefano Zampini PetscScalar *marray; 354288ebb749SStefano Zampini 354306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 35448eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 354506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 354688ebb749SStefano Zampini } 35473301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 354806656605SStefano Zampini n *= 2; 354988ebb749SStefano Zampini } 355006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 355106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 355206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 35538eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 355406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 355506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 355688ebb749SStefano Zampini } 35573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 355806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 35598eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 356006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 356106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 356288ebb749SStefano Zampini } 356388ebb749SStefano Zampini } else { 3564c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 3565c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 35661b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3567c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 3568c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 3569c0553b1fSStefano Zampini } 357088ebb749SStefano Zampini } 357106656605SStefano Zampini } 3572019a44ceSStefano Zampini 357306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 35744f1b2e48SStefano Zampini p0_lidx_I = NULL; 35754f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 3576d12edf2fSStefano Zampini const PetscInt *idxs; 3577d12edf2fSStefano Zampini 3578d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 35794f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 35804f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 35814f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 35824f1b2e48SStefano Zampini } 3583d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 3584d12edf2fSStefano Zampini } 3585d16cbb6bSStefano Zampini 358606656605SStefano Zampini /* vertices */ 358706656605SStefano Zampini if (n_vertices) { 358816f15bc4SStefano Zampini 3589af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 359004708bb6SStefano Zampini 359116f15bc4SStefano Zampini if (n_R) { 359214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 359306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 359416f15bc4SStefano Zampini PetscScalar *x,*y; 359504708bb6SStefano Zampini PetscBool isseqaij; 359606656605SStefano Zampini 359721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 359814393ed6SStefano Zampini if (need_benign_correction) { 359914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 360014393ed6SStefano Zampini IS is_p0; 360114393ed6SStefano Zampini PetscInt *idxs_p0,n; 360214393ed6SStefano Zampini 360314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 360414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 360514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 3606af25d912SStefano 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); 360714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 360814393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 360914393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 361014393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 361114393ed6SStefano Zampini } 361214393ed6SStefano Zampini 3613ffd830a3SStefano Zampini if (lda_rhs == n_R) { 3614af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3615ffd830a3SStefano Zampini } else { 3616ca92afb2SStefano Zampini PetscScalar *av,*array; 3617ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 3618ca92afb2SStefano Zampini PetscInt n; 3619ca92afb2SStefano Zampini PetscBool flg_row; 3620ffd830a3SStefano Zampini 3621ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 3622ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 36239d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3624ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3625ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 3626ca92afb2SStefano Zampini for (i=0;i<n;i++) { 3627ca92afb2SStefano Zampini PetscInt j; 3628ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 3629ffd830a3SStefano Zampini } 3630ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3631ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3632ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 3633ffd830a3SStefano Zampini } 3634ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3635a3df083aSStefano Zampini if (need_benign_correction) { 3636df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3637a3df083aSStefano Zampini PetscScalar *marr; 3638a3df083aSStefano Zampini 3639a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 364014393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 364114393ed6SStefano Zampini 364214393ed6SStefano Zampini | 0 0 0 | (V) 364314393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 364414393ed6SStefano Zampini | 0 0 -1 | (p0) 364514393ed6SStefano Zampini 364614393ed6SStefano Zampini */ 3647df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 364814393ed6SStefano Zampini const PetscScalar *vals; 364914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 365014393ed6SStefano Zampini PetscInt n,j,nz; 365114393ed6SStefano Zampini 3652df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3653df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 365414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 365514393ed6SStefano Zampini for (j=0;j<n;j++) { 365614393ed6SStefano Zampini PetscScalar val = vals[j]; 365714393ed6SStefano Zampini PetscInt k,col = idxs[j]; 365814393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 365914393ed6SStefano Zampini } 366014393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3661df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 366214393ed6SStefano Zampini } 366372b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 366472b8c272SStefano Zampini } 366572b8c272SStefano Zampini if (F) { 366614393ed6SStefano Zampini /* need to correct the rhs */ 366772b8c272SStefano Zampini if (need_benign_correction) { 366872b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 366972b8c272SStefano Zampini PetscScalar *marr; 367072b8c272SStefano Zampini 367172b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 36725cbda25cSStefano Zampini if (lda_rhs != n_R) { 36735cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 36745cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 36755cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 36765cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 36775cbda25cSStefano Zampini } 36785cbda25cSStefano Zampini } else { 3679a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3680a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 36815cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3682a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3683a3df083aSStefano Zampini } 36845cbda25cSStefano Zampini } 3685a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 3686a3df083aSStefano Zampini } 368706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 368814393ed6SStefano Zampini /* need to correct the solution */ 3689a3df083aSStefano Zampini if (need_benign_correction) { 3690df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3691a3df083aSStefano Zampini PetscScalar *marr; 3692a3df083aSStefano Zampini 3693a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 36945cbda25cSStefano Zampini if (lda_rhs != n_R) { 36955cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 36965cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 36975cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 36985cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 36995cbda25cSStefano Zampini } 37005cbda25cSStefano Zampini } else { 3701a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3702a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 37035cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3704a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3705a3df083aSStefano Zampini } 37065cbda25cSStefano Zampini } 3707a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 3708a3df083aSStefano Zampini } 370906656605SStefano Zampini } else { 371006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 371106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 3712ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 3713ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 371406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 371506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 371606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 371706656605SStefano Zampini } 371806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 371906656605SStefano Zampini } 372080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3721ffd830a3SStefano Zampini /* S_VV and S_CV */ 372206656605SStefano Zampini if (n_constraints) { 372306656605SStefano Zampini Mat B; 372480677318SStefano Zampini 3725ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 372680677318SStefano Zampini for (i=0;i<n_vertices;i++) { 3727ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3728ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 372980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 373080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 373180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 373280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 373380677318SStefano Zampini } 3734ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 373580677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 373680677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3737ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 373880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 373906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 3740ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 3741ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 374206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 374306656605SStefano Zampini } 374404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 374504708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 3746511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 374704708bb6SStefano Zampini } 3748ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3749ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 3750ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3751ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 3752ffd830a3SStefano Zampini } 375306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 375414393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 375514393ed6SStefano Zampini if (need_benign_correction) { 3756df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 375714393ed6SStefano Zampini PetscScalar *marr,*sums; 375814393ed6SStefano Zampini 375914393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 3760f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 3761df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 376214393ed6SStefano Zampini const PetscScalar *vals; 376314393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 376414393ed6SStefano Zampini PetscInt n,j,nz; 376514393ed6SStefano Zampini 3766df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3767df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 376814393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 376914393ed6SStefano Zampini PetscInt k; 377014393ed6SStefano Zampini sums[j] = 0.; 377114393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 377214393ed6SStefano Zampini } 377314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 377414393ed6SStefano Zampini for (j=0;j<n;j++) { 377514393ed6SStefano Zampini PetscScalar val = vals[j]; 377614393ed6SStefano Zampini PetscInt k; 377714393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 377814393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 377914393ed6SStefano Zampini } 378014393ed6SStefano Zampini } 378114393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3782df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 378314393ed6SStefano Zampini } 378414393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 3785f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 378614393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 378714393ed6SStefano Zampini } 378880677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 378906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 379006656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 379106656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 379206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 379306656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 379406656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 379506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3796d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 3797019a44ceSStefano Zampini } else { 3798d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3799d16cbb6bSStefano Zampini } 380021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 3801d16cbb6bSStefano Zampini 380206656605SStefano Zampini /* coarse basis functions */ 380306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 380416f15bc4SStefano Zampini PetscScalar *y; 380516f15bc4SStefano Zampini 3806ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 380706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 380806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 380906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 381006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 381106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 381206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 381306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 381406656605SStefano Zampini 381506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 38164f1b2e48SStefano Zampini PetscInt j; 38174f1b2e48SStefano Zampini 381806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 381906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 382006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 382106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 382206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 38234f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 382406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 382506656605SStefano Zampini } 382606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 382706656605SStefano Zampini } 382804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 382904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 383006656605SStefano Zampini } 38315cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 383206656605SStefano Zampini 383306656605SStefano Zampini if (n_constraints) { 383406656605SStefano Zampini Mat B; 383506656605SStefano Zampini 3836ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 383706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 383880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 383906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 384006656605SStefano Zampini if (n_vertices) { 384180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 384280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 384380677318SStefano Zampini } else { 384480677318SStefano Zampini Mat S_VCt; 384580677318SStefano Zampini 3846ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3847ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 384872b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 3849ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 3850ffd830a3SStefano Zampini } 385180677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 385280677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 385380677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 385480677318SStefano Zampini } 385506656605SStefano Zampini } 385606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 385706656605SStefano Zampini /* coarse basis functions */ 385806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 385906656605SStefano Zampini PetscScalar *y; 386006656605SStefano Zampini 3861ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 386206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 386306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 386406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 386506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 386606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 386706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 386806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 38694f1b2e48SStefano Zampini PetscInt j; 38704f1b2e48SStefano Zampini 387106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 387206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 387306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 387406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 387506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 38764f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 387706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 387806656605SStefano Zampini } 387906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 388006656605SStefano Zampini } 388106656605SStefano Zampini } 388280677318SStefano Zampini if (n_constraints) { 388380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 388480677318SStefano Zampini } 38854f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 388672b8c272SStefano Zampini 388772b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 388872b8c272SStefano Zampini if (pcbddc->benign_n) { 388972b8c272SStefano Zampini Mat B0_B,B0_BPHI; 389072b8c272SStefano Zampini IS is_dummy; 389172b8c272SStefano Zampini PetscScalar *data; 389272b8c272SStefano Zampini PetscInt j; 389372b8c272SStefano Zampini 389472b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 389572b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 389672b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 389772b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 389886c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 389972b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 390072b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 390172b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 390272b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 390372b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 390472b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 390572b8c272SStefano Zampini } 390672b8c272SStefano Zampini } 390772b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 390872b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 390972b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 391072b8c272SStefano Zampini } 3911019a44ceSStefano Zampini 391206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 39133301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 3914ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 3915ffd830a3SStefano Zampini PetscScalar *marray; 391606656605SStefano Zampini 391706656605SStefano Zampini if (n_constraints) { 3918ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 391906656605SStefano Zampini 3920af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 392106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 3922ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 392316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 392406656605SStefano Zampini if (n_vertices) { 3925ffd830a3SStefano Zampini Mat S_VCT; 392606656605SStefano Zampini 392706656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 3928ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 392916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 393006656605SStefano Zampini } 3931ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 39325b782168SStefano Zampini } else { 39335b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 393406656605SStefano Zampini } 393516f15bc4SStefano Zampini if (n_vertices && n_R) { 3936ffd830a3SStefano Zampini PetscScalar *av,*marray; 3937ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 3938ffd830a3SStefano Zampini PetscInt n; 3939ffd830a3SStefano Zampini PetscBool flg_row; 394006656605SStefano Zampini 3941ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 3942af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 3943ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3944ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 3945ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 3946ffd830a3SStefano Zampini for (i=0;i<n;i++) { 3947ffd830a3SStefano Zampini PetscInt j; 3948ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 3949ffd830a3SStefano Zampini } 3950ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 3951ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3952ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 395306656605SStefano Zampini } 395406656605SStefano Zampini 3955ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 3956ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 3957ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 3958ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 3959ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 396006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 396106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 396206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 396306656605SStefano Zampini } 3964ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 39655b782168SStefano Zampini if (B_C) { 3966ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 3967ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 3968ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 3969ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 3970ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 3971ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3972ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 397306656605SStefano Zampini } 3974ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 39755b782168SStefano Zampini } 397606656605SStefano Zampini /* coarse basis functions */ 397706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 397806656605SStefano Zampini PetscScalar *y; 397906656605SStefano Zampini 3980ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 398106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 398206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 398306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 398406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 398506656605SStefano Zampini if (i<n_vertices) { 398606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 398706656605SStefano Zampini } 398806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 398906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 399006656605SStefano Zampini 399106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 399206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 399306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 399406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 399506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 399606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 399706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 399806656605SStefano Zampini } 399906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 400006656605SStefano Zampini } 4001ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 4002ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 400306656605SStefano Zampini } 4004d62866d3SStefano Zampini /* free memory */ 400588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 400606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 400706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 400806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 400906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 4010d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4011d62866d3SStefano Zampini if (n_vertices) { 4012d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 4013d62866d3SStefano Zampini } 4014d62866d3SStefano Zampini if (n_constraints) { 4015d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 4016d62866d3SStefano Zampini } 401788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 401888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 401988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 4020d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 402188ebb749SStefano Zampini Mat coarse_sub_mat; 402225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 402388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 402488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 402588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 40268bec7fa6SStefano Zampini Mat C_B,CPHI; 40278bec7fa6SStefano Zampini IS is_dummy; 40288bec7fa6SStefano Zampini Vec mones; 402988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 403088ebb749SStefano Zampini PetscReal real_value; 403188ebb749SStefano Zampini 4032a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 4033a3df083aSStefano Zampini Mat A; 4034a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 4035a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 4036a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 4037a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 4038a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4039a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 4040a3df083aSStefano Zampini } else { 404188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 404288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 404388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 404488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4045a3df083aSStefano Zampini } 404688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 404788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 4048ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 404988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 405088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 405188ebb749SStefano Zampini } 405288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 405388ebb749SStefano Zampini 405425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 40553301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 405625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4057ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 405888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 405988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 406088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 406188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 406288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 406388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 406488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 406588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 406688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 406788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 406888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 406988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 407088ebb749SStefano Zampini } else { 407188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 407288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 407388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 407488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 407588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 407688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 407788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 407888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 407988ebb749SStefano Zampini } 408088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 408188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 408288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 4083511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 40844f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4085fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 4086d12edf2fSStefano Zampini PetscScalar *data,*data2; 40874f1b2e48SStefano Zampini PetscInt j; 4088d12edf2fSStefano Zampini 40894f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 4090fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 4091d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 409286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 4093d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 4094d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 40954f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 40964f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 4097d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 40984f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 40994f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 41004f1b2e48SStefano Zampini } 4101d12edf2fSStefano Zampini } 4102d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 4103d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 4104d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 4105d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 4106d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 4107d12edf2fSStefano Zampini } 4108d12edf2fSStefano Zampini #if 0 4109d12edf2fSStefano Zampini { 4110d12edf2fSStefano Zampini PetscViewer viewer; 4111d12edf2fSStefano Zampini char filename[256]; 4112ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4113d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4114d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4115ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 4116ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 4117ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 4118d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 411972b8c272SStefano Zampini if (save_change) { 412072b8c272SStefano Zampini Mat phi_B; 412172b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 412272b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 412372b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 412472b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 412572b8c272SStefano Zampini } else { 4126ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 4127ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 412872b8c272SStefano Zampini } 4129ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 4130ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 4131ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 4132ffd830a3SStefano Zampini } 4133ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 4134ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 4135ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 4136ffd830a3SStefano Zampini } 413772b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 4138ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 4139ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 4140ffd830a3SStefano Zampini } 4141d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4142d12edf2fSStefano Zampini } 4143d12edf2fSStefano Zampini #endif 414481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 41458bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 41461575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 414706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 41488bec7fa6SStefano Zampini 41498bec7fa6SStefano Zampini /* check constraints */ 4150a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 4151a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 41524f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 41538bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4154a00504b5SStefano Zampini } else { 4155a00504b5SStefano Zampini PetscScalar *data; 4156a00504b5SStefano Zampini Mat tmat; 4157a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4158a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 4159a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4160a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4161a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4162a00504b5SStefano Zampini } 41638bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 41648bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 41658bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 41668bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4167bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 4168ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 4169bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4170bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 4171bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 4172bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4173bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 417488ebb749SStefano Zampini } 41758bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 41768bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 41778bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 41788bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 417925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 418088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 418188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 418288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 418388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 418488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 418588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 418688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 418788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 418888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 418988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 4190ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 419188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 419288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 419388ebb749SStefano Zampini } 419488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 419588ebb749SStefano Zampini } 41968629588bSStefano Zampini /* get back data */ 41978629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 419888ebb749SStefano Zampini PetscFunctionReturn(0); 419988ebb749SStefano Zampini } 420088ebb749SStefano Zampini 420188ebb749SStefano Zampini #undef __FUNCT__ 4202d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 4203d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 4204aa0d41d4SStefano Zampini { 4205d65f70fdSStefano Zampini Mat *work_mat; 4206d65f70fdSStefano Zampini IS isrow_s,iscol_s; 4207d65f70fdSStefano Zampini PetscBool rsorted,csorted; 4208c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 4209aa0d41d4SStefano Zampini PetscErrorCode ierr; 4210aa0d41d4SStefano Zampini 4211aa0d41d4SStefano Zampini PetscFunctionBegin; 4212d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 4213d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 4214d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 4215d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 4216aa0d41d4SStefano Zampini 4217d65f70fdSStefano Zampini if (!rsorted) { 4218906d46d4SStefano Zampini const PetscInt *idxs; 4219906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 4220aa0d41d4SStefano Zampini 4221d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 4222d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 4223d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4224d65f70fdSStefano Zampini idxs_perm_r[i] = i; 4225aa0d41d4SStefano Zampini } 4226d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 4227d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 4228d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4229d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 4230aa0d41d4SStefano Zampini } 4231d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 4232d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 4233d65f70fdSStefano Zampini } else { 4234d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 4235d65f70fdSStefano Zampini isrow_s = isrow; 4236aa0d41d4SStefano Zampini } 4237906d46d4SStefano Zampini 4238d65f70fdSStefano Zampini if (!csorted) { 4239d65f70fdSStefano Zampini if (isrow == iscol) { 4240d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 4241d65f70fdSStefano Zampini iscol_s = isrow_s; 4242d65f70fdSStefano Zampini } else { 4243d65f70fdSStefano Zampini const PetscInt *idxs; 4244d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 4245906d46d4SStefano Zampini 4246d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 4247d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 4248d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4249d65f70fdSStefano Zampini idxs_perm_c[i] = i; 4250d65f70fdSStefano Zampini } 4251d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 4252d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 4253d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4254d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 4255d65f70fdSStefano Zampini } 4256d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 4257d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 4258d65f70fdSStefano Zampini } 4259d65f70fdSStefano Zampini } else { 4260d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 4261d65f70fdSStefano Zampini iscol_s = iscol; 4262d65f70fdSStefano Zampini } 4263d65f70fdSStefano Zampini 4264d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4265d65f70fdSStefano Zampini 4266d65f70fdSStefano Zampini if (!rsorted || !csorted) { 4267906d46d4SStefano Zampini Mat new_mat; 4268d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 4269906d46d4SStefano Zampini 4270d65f70fdSStefano Zampini if (!rsorted) { 4271d65f70fdSStefano Zampini PetscInt *idxs_r,i; 4272d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 4273d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4274d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 4275906d46d4SStefano Zampini } 4276d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 4277d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 4278d65f70fdSStefano Zampini } else { 4279d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 4280906d46d4SStefano Zampini } 4281d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 4282d65f70fdSStefano Zampini 4283d65f70fdSStefano Zampini if (!csorted) { 4284d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 4285d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 4286d65f70fdSStefano Zampini is_perm_c = is_perm_r; 4287d65f70fdSStefano Zampini } else { 4288d65f70fdSStefano Zampini PetscInt *idxs_c,i; 4289f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 4290d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 4291d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4292d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 4293d65f70fdSStefano Zampini } 4294d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 4295d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 4296d65f70fdSStefano Zampini } 4297d65f70fdSStefano Zampini } else { 4298d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 4299d65f70fdSStefano Zampini } 4300d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 4301d65f70fdSStefano Zampini 4302d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 4303d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 4304d65f70fdSStefano Zampini work_mat[0] = new_mat; 4305d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 4306d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 4307d65f70fdSStefano Zampini } 4308d65f70fdSStefano Zampini 4309d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 4310d65f70fdSStefano Zampini *B = work_mat[0]; 4311d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 4312d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 4313d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 4314d65f70fdSStefano Zampini PetscFunctionReturn(0); 4315d65f70fdSStefano Zampini } 4316d65f70fdSStefano Zampini 4317d65f70fdSStefano Zampini #undef __FUNCT__ 43185e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 43195e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 4320aa0d41d4SStefano Zampini { 4321aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 43225e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4323d65f70fdSStefano Zampini Mat new_mat; 43245e8657edSStefano Zampini IS is_local,is_global; 4325d65f70fdSStefano Zampini PetscInt local_size; 4326d65f70fdSStefano Zampini PetscBool isseqaij; 4327aa0d41d4SStefano Zampini PetscErrorCode ierr; 4328aa0d41d4SStefano Zampini 4329aa0d41d4SStefano Zampini PetscFunctionBegin; 4330aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 43315e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 43325e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 4333b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 4334aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 4335d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 4336aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 4337906d46d4SStefano Zampini 4338906d46d4SStefano Zampini /* check */ 4339906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4340906d46d4SStefano Zampini Vec x,x_change; 4341906d46d4SStefano Zampini PetscReal error; 4342906d46d4SStefano Zampini 43435e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 4344906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 43455e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 4346e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4347e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4348d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 434988428137SStefano Zampini if (!pcbddc->change_interior) { 435088428137SStefano Zampini const PetscScalar *x,*y,*v; 435188428137SStefano Zampini PetscReal lerror = 0.; 435288428137SStefano Zampini PetscInt i; 435388428137SStefano Zampini 435488428137SStefano Zampini ierr = VecGetArrayRead(matis->x,&x);CHKERRQ(ierr); 435588428137SStefano Zampini ierr = VecGetArrayRead(matis->y,&y);CHKERRQ(ierr); 435688428137SStefano Zampini ierr = VecGetArrayRead(matis->counter,&v);CHKERRQ(ierr); 435788428137SStefano Zampini for (i=0;i<local_size;i++) 435888428137SStefano Zampini if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror) 435988428137SStefano Zampini lerror = PetscAbsScalar(x[i]-y[i]); 436088428137SStefano Zampini ierr = VecRestoreArrayRead(matis->x,&x);CHKERRQ(ierr); 436188428137SStefano Zampini ierr = VecRestoreArrayRead(matis->y,&y);CHKERRQ(ierr); 436288428137SStefano Zampini ierr = VecRestoreArrayRead(matis->counter,&v);CHKERRQ(ierr); 436388428137SStefano Zampini ierr = MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4364637e8532SStefano Zampini if (error > PETSC_SMALL) { 4365637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4366637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on I: %1.6e\n",error); 4367637e8532SStefano Zampini } else { 4368637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on I: %1.6e\n",error); 4369637e8532SStefano Zampini } 4370637e8532SStefano Zampini } 437188428137SStefano Zampini } 4372e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4373e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4374906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4375906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4376637e8532SStefano Zampini if (error > PETSC_SMALL) { 4377637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4378637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 4379637e8532SStefano Zampini } else { 4380637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on N: %1.6e\n",error); 4381637e8532SStefano Zampini } 4382637e8532SStefano Zampini } 4383906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4384906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4385906d46d4SStefano Zampini } 4386906d46d4SStefano Zampini 438722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 43889b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 438922d5777bSStefano Zampini if (isseqaij) { 4390a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4391a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 4392aa0d41d4SStefano Zampini } else { 4393a00504b5SStefano Zampini Mat work_mat; 43941cf9b237SStefano Zampini 4395a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4396aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4397a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 43981d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 4399aa0d41d4SStefano Zampini } 44003301b35fSStefano Zampini if (matis->A->symmetric_set) { 44013301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 4402e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 44033301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 4404e496cd5dSStefano Zampini #endif 44053301b35fSStefano Zampini } 4406d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 4407aa0d41d4SStefano Zampini PetscFunctionReturn(0); 4408aa0d41d4SStefano Zampini } 4409aa0d41d4SStefano Zampini 4410aa0d41d4SStefano Zampini #undef __FUNCT__ 4411a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 44128ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 4413a64d13efSStefano Zampini { 4414a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4415a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4416d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 441753892102SStefano Zampini PetscInt *idx_R_local=NULL; 44183a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 44193a50541eSStefano Zampini PetscInt vbs,bs; 44206816873aSStefano Zampini PetscBT bitmask=NULL; 4421a64d13efSStefano Zampini PetscErrorCode ierr; 4422a64d13efSStefano Zampini 4423a64d13efSStefano Zampini PetscFunctionBegin; 4424b23d619eSStefano Zampini /* 4425b23d619eSStefano Zampini No need to setup local scatters if 4426b23d619eSStefano Zampini - primal space is unchanged 4427b23d619eSStefano Zampini AND 4428b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 4429b23d619eSStefano Zampini AND 4430b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 4431b23d619eSStefano Zampini */ 4432b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 4433f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 4434f4ddd8eeSStefano Zampini } 4435f4ddd8eeSStefano Zampini /* destroy old objects */ 4436f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 4437f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 4438f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 4439a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 4440b371cd4fSStefano Zampini n_B = pcis->n_B; 4441b371cd4fSStefano Zampini n_D = pcis->n - n_B; 4442b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 44433a50541eSStefano Zampini 4444a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 44456816873aSStefano Zampini 444653892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 4447b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4448854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 4449a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 4450a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 44510e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 4452a64d13efSStefano Zampini } 4453a64d13efSStefano Zampini 4454a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 44554641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 44566816873aSStefano Zampini idx_R_local[n_R++] = i; 4457a64d13efSStefano Zampini } 4458a64d13efSStefano Zampini } 4459df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 4460df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 44616816873aSStefano Zampini 4462df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4463df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 44646816873aSStefano Zampini } 44653a50541eSStefano Zampini 44663a50541eSStefano Zampini /* Block code */ 44673a50541eSStefano Zampini vbs = 1; 44683a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 44693a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 44703a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 44713a50541eSStefano Zampini PetscInt *vary; 4472b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4473785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 44743a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 4475d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 4476d3df7717SStefano 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 */ 44770e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 4478d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 44793a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 44803a50541eSStefano Zampini is_blocked = PETSC_FALSE; 44813a50541eSStefano Zampini break; 44823a50541eSStefano Zampini } 44833a50541eSStefano Zampini } 4484d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 4485d3df7717SStefano Zampini } else { 4486d3df7717SStefano Zampini /* Verify directly the R set */ 4487d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 4488d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 4489d3df7717SStefano Zampini for (j=1; j<bs; j++) { 4490d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 4491d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 4492d3df7717SStefano Zampini break; 4493d3df7717SStefano Zampini } 4494d3df7717SStefano Zampini } 4495d3df7717SStefano Zampini } 4496d3df7717SStefano Zampini } 44973a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 44983a50541eSStefano Zampini vbs = bs; 44993a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 45003a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 45013a50541eSStefano Zampini } 45023a50541eSStefano Zampini } 45033a50541eSStefano Zampini } 45043a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 4505b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4506df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 450753892102SStefano Zampini 4508df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4509df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 451053892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 4511df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 451253892102SStefano Zampini } else { 45133a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 451453892102SStefano Zampini } 4515a64d13efSStefano Zampini 4516a64d13efSStefano Zampini /* print some info if requested */ 4517a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 4518a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4519a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45201575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4521a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 4522a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 45234f1b2e48SStefano 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); 4524a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4525a64d13efSStefano Zampini } 4526a64d13efSStefano Zampini 4527a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 4528b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 45296816873aSStefano Zampini IS is_aux1,is_aux2; 45306816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 45316816873aSStefano Zampini 45323a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4533854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 4534854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 4535a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 45364641a718SStefano Zampini for (i=0; i<n_D; i++) { 45374641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 45384641a718SStefano Zampini } 4539a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4540a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 45414641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 45424641a718SStefano Zampini aux_array1[j++] = i; 4543a64d13efSStefano Zampini } 4544a64d13efSStefano Zampini } 4545a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4546a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4547a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 45484641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 45494641a718SStefano Zampini aux_array2[j++] = i; 4550a64d13efSStefano Zampini } 4551a64d13efSStefano Zampini } 4552a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4553a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 4554a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 4555a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4556a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 4557a64d13efSStefano Zampini 45588eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4559785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 4560a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 45614641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 45624641a718SStefano Zampini aux_array1[j++] = i; 4563a64d13efSStefano Zampini } 4564a64d13efSStefano Zampini } 4565a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4566a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 4567a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4568a64d13efSStefano Zampini } 45694641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 45703a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4571d62866d3SStefano Zampini } else { 4572df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 45736816873aSStefano Zampini IS tis; 45746816873aSStefano Zampini PetscInt schur_size; 45756816873aSStefano Zampini 4576df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 45776816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 4578df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 45796816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 45806816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 45816816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 45826816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 45836816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 4584d62866d3SStefano Zampini } 4585d62866d3SStefano Zampini } 4586a64d13efSStefano Zampini PetscFunctionReturn(0); 4587a64d13efSStefano Zampini } 4588a64d13efSStefano Zampini 4589304d26faSStefano Zampini 4590304d26faSStefano Zampini #undef __FUNCT__ 4591304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 4592684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 4593304d26faSStefano Zampini { 4594304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4595304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4596304d26faSStefano Zampini PC pc_temp; 4597304d26faSStefano Zampini Mat A_RR; 4598f4ddd8eeSStefano Zampini MatReuse reuse; 4599304d26faSStefano Zampini PetscScalar m_one = -1.0; 4600304d26faSStefano Zampini PetscReal value; 460104708bb6SStefano Zampini PetscInt n_D,n_R; 4602c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 4603304d26faSStefano Zampini PetscErrorCode ierr; 4604e604994aSStefano Zampini /* prefixes stuff */ 4605312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 4606e604994aSStefano Zampini size_t len; 4607304d26faSStefano Zampini 4608304d26faSStefano Zampini PetscFunctionBegin; 4609304d26faSStefano Zampini 4610e604994aSStefano Zampini /* compute prefixes */ 4611e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 4612e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 4613e604994aSStefano Zampini if (!pcbddc->current_level) { 4614e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4615e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4616e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4617e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4618e604994aSStefano Zampini } else { 4619e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4620312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4621e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4622e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 4623312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4624312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 462534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 462634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4627e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4628e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4629e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 4630e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 4631e604994aSStefano Zampini } 4632e604994aSStefano Zampini 4633304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 4634684f6988SStefano Zampini if (dirichlet) { 4635d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4636450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 46379a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 4638450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 4639a3df083aSStefano Zampini Mat A_IIn; 4640a3df083aSStefano Zampini 4641a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 4642a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 4643a3df083aSStefano Zampini pcis->A_II = A_IIn; 4644a3df083aSStefano Zampini } 4645450f8f5eSStefano Zampini } 46463301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 46473301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 4648964fefecSStefano Zampini } 4649ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 4650964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 4651304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 4652304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 4653304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 4654304d26faSStefano Zampini /* default */ 4655304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 4656e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 46579577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 4658304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 46599577ea80SStefano Zampini if (issbaij) { 46609577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 46619577ea80SStefano Zampini } else { 4662304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 46639577ea80SStefano Zampini } 4664304d26faSStefano Zampini /* Allow user's customization */ 4665304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 4666304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4667304d26faSStefano Zampini } 4668d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 4669b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4670df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4671d62866d3SStefano Zampini 4672df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 4673d5574798SStefano Zampini } 4674304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4675304d26faSStefano Zampini if (!n_D) { 4676304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 4677304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4678304d26faSStefano Zampini } 4679304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 4680304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 4681304d26faSStefano Zampini /* set ksp_D into pcis data */ 4682304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 4683304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 4684304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 4685684f6988SStefano Zampini } 4686304d26faSStefano Zampini 4687304d26faSStefano Zampini /* NEUMANN PROBLEM */ 4688684f6988SStefano Zampini A_RR = 0; 4689684f6988SStefano Zampini if (neumann) { 4690d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 469104708bb6SStefano Zampini PetscInt ibs,mbs; 469204708bb6SStefano Zampini PetscBool issbaij; 469304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4694f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 46958ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 4696f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 4697f4ddd8eeSStefano Zampini PetscInt nn_R; 469881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 4699f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4700f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 4701f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 4702f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 4703f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4704f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4705f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 4706727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 4707f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4708f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4709f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 4710f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 4711f4ddd8eeSStefano Zampini } 4712f4ddd8eeSStefano Zampini } 4713f4ddd8eeSStefano Zampini /* last check */ 4714d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 4715f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4716f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4717f4ddd8eeSStefano Zampini } 4718f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 4719f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4720f4ddd8eeSStefano Zampini } 4721a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 4722af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 4723af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 472404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 472504708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 472604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 472704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 472804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4729af732b37SStefano Zampini } else { 4730511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 47316816873aSStefano Zampini } 473204708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 473304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 473404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 473504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 473604708bb6SStefano Zampini } else { 4737511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 473804708bb6SStefano Zampini } 473904708bb6SStefano Zampini } 4740a00504b5SStefano Zampini /* extract A_RR */ 4741b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4742a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4743a00504b5SStefano Zampini 4744a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 474516e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4746a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 474716e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 474816e386b8SStefano Zampini } else { 4749a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 4750a00504b5SStefano Zampini } 4751a00504b5SStefano Zampini } else { 4752a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4753a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 4754a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4755a00504b5SStefano Zampini } 4756a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 4757f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 475816e386b8SStefano Zampini } 47593301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 47603301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 47616816873aSStefano Zampini } 4762f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 4763304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 4764304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 4765304d26faSStefano Zampini /* default */ 4766304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 4767e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 4768304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 47699577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 47709577ea80SStefano Zampini if (issbaij) { 47719577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 47729577ea80SStefano Zampini } else { 4773304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 47749577ea80SStefano Zampini } 4775304d26faSStefano Zampini /* Allow user's customization */ 4776304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 4777304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4778304d26faSStefano Zampini } 4779304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4780304d26faSStefano Zampini if (!n_R) { 4781304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 4782304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4783304d26faSStefano Zampini } 47845cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 4785df4d28bfSStefano Zampini /* Reuse solver if it is present */ 4786b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4787df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4788d62866d3SStefano Zampini 4789df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 4790d62866d3SStefano Zampini } 4791304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 4792304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 4793684f6988SStefano Zampini } 4794304d26faSStefano Zampini 4795684f6988SStefano Zampini if (pcbddc->dbg_flag) { 4796684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 47971575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4798684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4799684f6988SStefano Zampini } 4800c7017625SStefano Zampini 4801c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 4802c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 4803c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 4804c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 4805c7017625SStefano Zampini } 4806c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 4807c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 4808c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 4809c7017625SStefano Zampini } 4810c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 4811c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 4812c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 4813c7017625SStefano Zampini } 4814c7017625SStefano Zampini 4815c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 4816c7017625SStefano Zampini if (pcbddc->dbg_flag) { 4817684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 48180fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 48190fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 48200fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 48210fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 48220fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 4823e604994aSStefano 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); 4824c7017625SStefano Zampini if (check_corr[0]) { 4825c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 4826c7017625SStefano Zampini } 4827304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4828304d26faSStefano Zampini } 4829684f6988SStefano Zampini if (neumann) { /* Neumann */ 48300fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 48310fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 48320fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 48330fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 48340fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 4835e604994aSStefano 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); 4836c7017625SStefano Zampini if (check_corr[1]) { 4837c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 4838c7017625SStefano Zampini } 4839304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4840304d26faSStefano Zampini } 4841684f6988SStefano Zampini } 48425cbda25cSStefano Zampini /* free Neumann problem's matrix */ 48435cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4844304d26faSStefano Zampini PetscFunctionReturn(0); 4845304d26faSStefano Zampini } 4846304d26faSStefano Zampini 4847304d26faSStefano Zampini #undef __FUNCT__ 4848ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 484980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 4850674ae819SStefano Zampini { 4851674ae819SStefano Zampini PetscErrorCode ierr; 4852674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4853be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4854b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 4855674ae819SStefano Zampini 4856674ae819SStefano Zampini PetscFunctionBegin; 4857b334f244SStefano Zampini if (!reuse_solver) { 485880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 485920c7b377SStefano Zampini } 486080677318SStefano Zampini if (!pcbddc->switch_static) { 486180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 486280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 486380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 486420c7b377SStefano Zampini } 4865b334f244SStefano Zampini if (!reuse_solver) { 486680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 486780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 486820c7b377SStefano Zampini } else { 4869df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4870be83ff47SStefano Zampini 4871df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4872df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 487320c7b377SStefano Zampini } 4874be83ff47SStefano Zampini } else { 487580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 487680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 487780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 487880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 487980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 488080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 488180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 488280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 488380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4884674ae819SStefano Zampini } 4885674ae819SStefano Zampini } 4886b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 488780677318SStefano Zampini if (applytranspose) { 488880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 488980677318SStefano Zampini } else { 489080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 489180677318SStefano Zampini } 4892be83ff47SStefano Zampini } else { 4893df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4894be83ff47SStefano Zampini 4895be83ff47SStefano Zampini if (applytranspose) { 4896df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4897be83ff47SStefano Zampini } else { 4898df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4899be83ff47SStefano Zampini } 4900be83ff47SStefano Zampini } 490180677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 490280677318SStefano Zampini if (!pcbddc->switch_static) { 4903b334f244SStefano Zampini if (!reuse_solver) { 490480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 490580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4906be83ff47SStefano Zampini } else { 4907df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4908be83ff47SStefano Zampini 4909df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4910df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4911be83ff47SStefano Zampini } 491280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 491380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 491480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 491580677318SStefano Zampini } 491680677318SStefano Zampini } else { 491780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 491880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 491980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 492080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 492180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 492280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 492380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 492480677318SStefano Zampini } 492580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 492680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 492780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 492880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4929674ae819SStefano Zampini } 4930674ae819SStefano Zampini PetscFunctionReturn(0); 4931674ae819SStefano Zampini } 4932674ae819SStefano Zampini 4933dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 4934674ae819SStefano Zampini #undef __FUNCT__ 4935674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 4936dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 4937674ae819SStefano Zampini { 4938674ae819SStefano Zampini PetscErrorCode ierr; 4939674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4940674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 4941674ae819SStefano Zampini const PetscScalar zero = 0.0; 4942674ae819SStefano Zampini 4943674ae819SStefano Zampini PetscFunctionBegin; 4944dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 49454fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 4946dc359a40SStefano Zampini if (applytranspose) { 4947674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 49488eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 4949dc359a40SStefano Zampini } else { 4950674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 4951674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 495215aaf578SStefano Zampini } 49534fee134fSStefano Zampini } else { 49544fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 49554fee134fSStefano Zampini } 4956efc2fbd9SStefano Zampini 4957efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 49584f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4959efc2fbd9SStefano Zampini PetscScalar *array; 49604f1b2e48SStefano Zampini PetscInt j; 4961efc2fbd9SStefano Zampini 4962efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 49634f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 4964efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 4965efc2fbd9SStefano Zampini } 4966efc2fbd9SStefano Zampini 496712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 496812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 496912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 497012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 497112edc857SStefano Zampini 49729f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 497312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 497451694757SStefano Zampini Mat coarse_mat; 4975964fefecSStefano Zampini Vec rhs,sol; 497651694757SStefano Zampini MatNullSpace nullsp; 497727b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 4978964fefecSStefano Zampini 497927b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 498027b6a85dSStefano Zampini PC coarse_pc; 498127b6a85dSStefano Zampini 498227b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 498327b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 498427b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 498527b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 498627b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 498727b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 49883bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 498927b6a85dSStefano Zampini } 499027b6a85dSStefano Zampini } 4991964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 4992964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 499351694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 499451694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 499551694757SStefano Zampini if (nullsp) { 499651694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 499751694757SStefano Zampini } 499812edc857SStefano Zampini if (applytranspose) { 49999a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 5000964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 50012701bc32SStefano Zampini } else { 50021f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 50032701bc32SStefano Zampini PC coarse_pc; 50042701bc32SStefano Zampini 50052701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 50062701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 50073e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 50082701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 500912edc857SStefano Zampini } else { 5010964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 501112edc857SStefano Zampini } 50122701bc32SStefano Zampini } 50131d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 501427b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 501527b6a85dSStefano Zampini PC coarse_pc; 501627b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 501727b6a85dSStefano Zampini 501827b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 501927b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 502027b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 50213bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 502227b6a85dSStefano Zampini } 502351694757SStefano Zampini if (nullsp) { 502451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 502551694757SStefano Zampini } 502612edc857SStefano Zampini } 5027674ae819SStefano Zampini 5028674ae819SStefano Zampini /* Local solution on R nodes */ 50294fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 503080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 50319f00e9b4SStefano Zampini } 50329f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 50339f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 503412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5035674ae819SStefano Zampini 50364fee134fSStefano Zampini /* Sum contributions from the two levels */ 50374fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5038dc359a40SStefano Zampini if (applytranspose) { 5039dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5040dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5041dc359a40SStefano Zampini } else { 5042674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 50438eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5044dc359a40SStefano Zampini } 5045efc2fbd9SStefano Zampini /* store p0 */ 50464f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5047efc2fbd9SStefano Zampini PetscScalar *array; 50484f1b2e48SStefano Zampini PetscInt j; 5049efc2fbd9SStefano Zampini 5050efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 50514f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 5052efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5053efc2fbd9SStefano Zampini } 50544fee134fSStefano Zampini } else { /* expand the coarse solution */ 50554fee134fSStefano Zampini if (applytranspose) { 50564fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 50574fee134fSStefano Zampini } else { 50584fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 50594fee134fSStefano Zampini } 50604fee134fSStefano Zampini } 5061674ae819SStefano Zampini PetscFunctionReturn(0); 5062674ae819SStefano Zampini } 5063674ae819SStefano Zampini 5064674ae819SStefano Zampini #undef __FUNCT__ 5065674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 506612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 5067674ae819SStefano Zampini { 5068674ae819SStefano Zampini PetscErrorCode ierr; 5069674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 507058da7f69SStefano Zampini PetscScalar *array; 507112edc857SStefano Zampini Vec from,to; 5072674ae819SStefano Zampini 5073674ae819SStefano Zampini PetscFunctionBegin; 507412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 507512edc857SStefano Zampini from = pcbddc->coarse_vec; 507612edc857SStefano Zampini to = pcbddc->vec1_P; 507712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 507812edc857SStefano Zampini Vec tvec; 507958da7f69SStefano Zampini 508058da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 508158da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 508212edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 508358da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 508458da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 508558da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 508612edc857SStefano Zampini } 508712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 508812edc857SStefano Zampini from = pcbddc->vec1_P; 508912edc857SStefano Zampini to = pcbddc->coarse_vec; 509012edc857SStefano Zampini } 509112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 5092674ae819SStefano Zampini PetscFunctionReturn(0); 5093674ae819SStefano Zampini } 5094674ae819SStefano Zampini 5095674ae819SStefano Zampini #undef __FUNCT__ 5096674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 509712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 5098674ae819SStefano Zampini { 5099674ae819SStefano Zampini PetscErrorCode ierr; 5100674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 510158da7f69SStefano Zampini PetscScalar *array; 510212edc857SStefano Zampini Vec from,to; 5103674ae819SStefano Zampini 5104674ae819SStefano Zampini PetscFunctionBegin; 510512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 510612edc857SStefano Zampini from = pcbddc->coarse_vec; 510712edc857SStefano Zampini to = pcbddc->vec1_P; 510812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 510912edc857SStefano Zampini from = pcbddc->vec1_P; 511012edc857SStefano Zampini to = pcbddc->coarse_vec; 511112edc857SStefano Zampini } 511212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 511312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 511412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 511512edc857SStefano Zampini Vec tvec; 511658da7f69SStefano Zampini 511712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 511858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 511958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 512058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 512158da7f69SStefano Zampini } 512258da7f69SStefano Zampini } else { 512358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 512458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 512512edc857SStefano Zampini } 512612edc857SStefano Zampini } 5127674ae819SStefano Zampini PetscFunctionReturn(0); 5128674ae819SStefano Zampini } 5129674ae819SStefano Zampini 5130984c4197SStefano Zampini /* uncomment for testing purposes */ 5131984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 5132674ae819SStefano Zampini #undef __FUNCT__ 5133674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 5134674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 5135674ae819SStefano Zampini { 5136674ae819SStefano Zampini PetscErrorCode ierr; 5137674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 5138674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5139674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5140984c4197SStefano Zampini /* one and zero */ 5141984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 5142984c4197SStefano Zampini /* space to store constraints and their local indices */ 51439162d606SStefano Zampini PetscScalar *constraints_data; 51449162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 51459162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 51469162d606SStefano Zampini PetscInt *constraints_n; 5147984c4197SStefano Zampini /* iterators */ 5148b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 5149984c4197SStefano Zampini /* BLAS integers */ 5150e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 5151e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 5152c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 5153727cdba6SStefano Zampini /* reuse */ 51540e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 51550e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 5156984c4197SStefano Zampini /* change of basis */ 5157b3d85658SStefano Zampini PetscBool qr_needed; 51589162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 5159984c4197SStefano Zampini /* auxiliary stuff */ 516064efe560SStefano Zampini PetscInt *nnz,*is_indices; 51618a0068c3SStefano Zampini PetscInt ncc; 5162984c4197SStefano Zampini /* some quantities */ 516345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 5164a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 5165984c4197SStefano Zampini 5166674ae819SStefano Zampini PetscFunctionBegin; 51678e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 51688e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 51698e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 517016909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 5171088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 5172088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 51730e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 51740e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 51750e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 51760e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 51770e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 5178088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5179cf5a6209SStefano Zampini 5180cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 51819162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 5182cf5a6209SStefano Zampini MatNullSpace nearnullsp; 5183cf5a6209SStefano Zampini const Vec *nearnullvecs; 5184cf5a6209SStefano Zampini Vec *localnearnullsp; 5185cf5a6209SStefano Zampini PetscScalar *array; 5186cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 5187cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 5188674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 5189b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 5190674ae819SStefano Zampini PetscScalar *work; 5191674ae819SStefano Zampini PetscReal *singular_vals; 5192674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5193674ae819SStefano Zampini PetscReal *rwork; 5194674ae819SStefano Zampini #endif 5195674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5196674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 5197674ae819SStefano Zampini #else 5198964fefecSStefano Zampini PetscBLASInt dummy_int=1; 5199964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 5200674ae819SStefano Zampini #endif 5201674ae819SStefano Zampini 5202674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 5203d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 5204e4d548c7SStefano Zampini /* print some info */ 52055c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 5206e4d548c7SStefano Zampini PetscInt nv; 5207e4d548c7SStefano Zampini 5208c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5209e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 5210e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5211e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5212e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 5213e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 5214e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 5215e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5216e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5217e4d548c7SStefano Zampini } 5218e4d548c7SStefano Zampini 5219d06fc5fdSStefano Zampini /* free unneeded index sets */ 5220d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 5221d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 5222674ae819SStefano Zampini } 5223d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 5224d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5225d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5226d06fc5fdSStefano Zampini } 5227d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5228d06fc5fdSStefano Zampini n_ISForEdges = 0; 5229d06fc5fdSStefano Zampini } 5230d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 5231d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5232d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5233d06fc5fdSStefano Zampini } 5234d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5235d06fc5fdSStefano Zampini n_ISForFaces = 0; 5236d06fc5fdSStefano Zampini } 523770022509SStefano Zampini 5238674ae819SStefano Zampini /* check if near null space is attached to global mat */ 5239674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 5240674ae819SStefano Zampini if (nearnullsp) { 5241674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 5242f4ddd8eeSStefano Zampini /* remove any stored info */ 5243f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 5244f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5245f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 5246f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 5247f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 5248473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5249f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 5250f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 5251f4ddd8eeSStefano Zampini } 5252984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 5253984c4197SStefano Zampini nnsp_size = 0; 5254674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 5255674ae819SStefano Zampini } 5256984c4197SStefano Zampini /* get max number of constraints on a single cc */ 5257984c4197SStefano Zampini max_constraints = nnsp_size; 5258984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 5259984c4197SStefano Zampini 5260674ae819SStefano Zampini /* 5261674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 52629162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 52639162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 52649162d606SStefano Zampini There can be multiple constraints per connected component 5265674ae819SStefano Zampini */ 5266674ae819SStefano Zampini n_vertices = 0; 5267674ae819SStefano Zampini if (ISForVertices) { 5268674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 5269674ae819SStefano Zampini } 52709162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 52719162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 52729162d606SStefano Zampini 52739162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 52749162d606SStefano Zampini total_counts *= max_constraints; 5275674ae819SStefano Zampini total_counts += n_vertices; 52764641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 52779162d606SStefano Zampini 5278674ae819SStefano Zampini total_counts = 0; 5279674ae819SStefano Zampini max_size_of_constraint = 0; 5280674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 52819162d606SStefano Zampini IS used_is; 5282674ae819SStefano Zampini if (i<n_ISForEdges) { 52839162d606SStefano Zampini used_is = ISForEdges[i]; 5284674ae819SStefano Zampini } else { 52859162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 5286674ae819SStefano Zampini } 52879162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 5288674ae819SStefano Zampini total_counts += j; 5289674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 5290674ae819SStefano Zampini } 52919162d606SStefano 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); 52929162d606SStefano Zampini 5293984c4197SStefano Zampini /* get local part of global near null space vectors */ 5294785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 5295984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5296984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 5297e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5298e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5299984c4197SStefano Zampini } 5300674ae819SStefano Zampini 5301242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 5302242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 5303a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 5304242a89d7SStefano Zampini 5305984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 5306a773dcb8SStefano Zampini if (!skip_lapack) { 5307674ae819SStefano Zampini PetscScalar temp_work; 5308911cabfeSStefano Zampini 5309674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5310984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 5311785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 5312785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 5313785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 5314674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5315785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 5316674ae819SStefano Zampini #endif 5317674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5318c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 5319c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 5320674ae819SStefano Zampini lwork = -1; 5321674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5322674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5323c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 5324674ae819SStefano Zampini #else 5325c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 5326674ae819SStefano Zampini #endif 5327674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5328984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 5329674ae819SStefano Zampini #else /* on missing GESVD */ 5330674ae819SStefano Zampini /* SVD */ 5331674ae819SStefano Zampini PetscInt max_n,min_n; 5332674ae819SStefano Zampini max_n = max_size_of_constraint; 5333984c4197SStefano Zampini min_n = max_constraints; 5334984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 5335674ae819SStefano Zampini min_n = max_size_of_constraint; 5336984c4197SStefano Zampini max_n = max_constraints; 5337674ae819SStefano Zampini } 5338785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 5339674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5340785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 5341674ae819SStefano Zampini #endif 5342674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5343674ae819SStefano Zampini lwork = -1; 5344e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 5345e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 5346b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 5347674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5348674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 53499162d606SStefano 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)); 5350674ae819SStefano Zampini #else 53519162d606SStefano 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)); 5352674ae819SStefano Zampini #endif 5353674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5354984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 5355984c4197SStefano Zampini #endif /* on missing GESVD */ 5356674ae819SStefano Zampini /* Allocate optimal workspace */ 5357674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 5358854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 5359674ae819SStefano Zampini } 5360674ae819SStefano Zampini /* Now we can loop on constraining sets */ 5361674ae819SStefano Zampini total_counts = 0; 53629162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 53639162d606SStefano Zampini constraints_data_ptr[0] = 0; 5364674ae819SStefano Zampini /* vertices */ 53659162d606SStefano Zampini if (n_vertices) { 5366674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 53679162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 5368674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 53699162d606SStefano Zampini constraints_n[total_counts] = 1; 53709162d606SStefano Zampini constraints_data[total_counts] = 1.0; 53719162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 53729162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 5373674ae819SStefano Zampini total_counts++; 5374674ae819SStefano Zampini } 5375674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5376674ae819SStefano Zampini n_vertices = total_counts; 5377674ae819SStefano Zampini } 5378984c4197SStefano Zampini 5379674ae819SStefano Zampini /* edges and faces */ 53809162d606SStefano Zampini total_counts_cc = total_counts; 5381911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 53829162d606SStefano Zampini IS used_is; 53839162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 53849162d606SStefano Zampini 5385911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 53869162d606SStefano Zampini used_is = ISForEdges[ncc]; 5387984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 5388674ae819SStefano Zampini } else { 53899162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 5390984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 5391674ae819SStefano Zampini } 5392674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 53939162d606SStefano Zampini 53949162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 53959162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5396984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 5397984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 5398674ae819SStefano Zampini if (nnsp_has_cnst) { 53995b08dc53SStefano Zampini PetscScalar quad_value; 54009162d606SStefano Zampini 54019162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 54029162d606SStefano Zampini idxs_copied = PETSC_TRUE; 54039162d606SStefano Zampini 5404a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 5405674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 5406a773dcb8SStefano Zampini } else { 5407a773dcb8SStefano Zampini quad_value = 1.0; 5408a773dcb8SStefano Zampini } 5409674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 54109162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 5411674ae819SStefano Zampini } 54129162d606SStefano Zampini temp_constraints++; 5413674ae819SStefano Zampini total_counts++; 5414674ae819SStefano Zampini } 5415674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 5416984c4197SStefano Zampini PetscReal real_value; 54179162d606SStefano Zampini PetscScalar *ptr_to_data; 54189162d606SStefano Zampini 5419984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 54209162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 5421674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 54229162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 5423674ae819SStefano Zampini } 5424984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 5425984c4197SStefano Zampini /* check if array is null on the connected component */ 5426e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 54279162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 54285b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 5429674ae819SStefano Zampini temp_constraints++; 5430674ae819SStefano Zampini total_counts++; 54319162d606SStefano Zampini if (!idxs_copied) { 54329162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 54339162d606SStefano Zampini idxs_copied = PETSC_TRUE; 5434674ae819SStefano Zampini } 5435674ae819SStefano Zampini } 54369162d606SStefano Zampini } 54379162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 543845a1bb75SStefano Zampini valid_constraints = temp_constraints; 5439eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 5440a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 54419162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 54429162d606SStefano Zampini 54439162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5444a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 54459162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 5446a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 54479162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 5448a773dcb8SStefano Zampini } else { /* perform SVD */ 5449984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 54509162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5451674ae819SStefano Zampini 5452674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5453984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 5454984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 5455984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 5456984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 5457984c4197SStefano Zampini from that computed using LAPACKgesvd 5458984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 5459984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 5460984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 5461674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 5462e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5463984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5464674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 5465674ae819SStefano Zampini for (k=0;k<j+1;k++) { 54669162d606SStefano 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)); 5467674ae819SStefano Zampini } 5468674ae819SStefano Zampini } 5469e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 5470e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5471e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 5472674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5473c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 5474674ae819SStefano Zampini #else 5475c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 5476674ae819SStefano Zampini #endif 5477674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5478984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 5479984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 5480674ae819SStefano Zampini j = 0; 5481984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 5482674ae819SStefano Zampini total_counts = total_counts-j; 548345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 5484e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 5485c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5486c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5487c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 5488c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5489c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 5490c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5491674ae819SStefano Zampini if (j<temp_constraints) { 5492984c4197SStefano Zampini PetscInt ii; 5493984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 5494674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 54959162d606SStefano 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)); 5496674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5497984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 5498674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 54999162d606SStefano 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]; 5500674ae819SStefano Zampini } 5501674ae819SStefano Zampini } 5502674ae819SStefano Zampini } 5503674ae819SStefano Zampini #else /* on missing GESVD */ 5504e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5505e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5506b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5507674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5508674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 55099162d606SStefano 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)); 5510674ae819SStefano Zampini #else 55119162d606SStefano 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)); 5512674ae819SStefano Zampini #endif 5513984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 5514674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5515984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 5516e310c8b4SStefano Zampini k = temp_constraints; 5517e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 5518674ae819SStefano Zampini j = 0; 5519e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 552045a1bb75SStefano Zampini valid_constraints = k-j; 5521911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 5522984c4197SStefano Zampini #endif /* on missing GESVD */ 5523674ae819SStefano Zampini } 5524a773dcb8SStefano Zampini } 55259162d606SStefano Zampini /* update pointers information */ 55269162d606SStefano Zampini if (valid_constraints) { 55279162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 55289162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 55299162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 55309162d606SStefano Zampini /* set change_of_basis flag */ 553145a1bb75SStefano Zampini if (boolforchange) { 5532b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 55339162d606SStefano Zampini } 5534b3d85658SStefano Zampini total_counts_cc++; 553545a1bb75SStefano Zampini } 553645a1bb75SStefano Zampini } 5537984c4197SStefano Zampini /* free workspace */ 55388f1c130eSStefano Zampini if (!skip_lapack) { 5539984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 5540984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5541984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 5542984c4197SStefano Zampini #endif 5543984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 5544984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5545984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 5546984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 5547984c4197SStefano Zampini #endif 5548984c4197SStefano Zampini } 5549984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5550984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 5551984c4197SStefano Zampini } 5552984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 5553cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 5554cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5555cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5556cf5a6209SStefano Zampini } 5557cf5a6209SStefano Zampini if (n_ISForFaces) { 5558cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5559cf5a6209SStefano Zampini } 5560cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5561cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5562cf5a6209SStefano Zampini } 5563cf5a6209SStefano Zampini if (n_ISForEdges) { 5564cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5565cf5a6209SStefano Zampini } 5566cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 556708122e43SStefano Zampini } else { 556808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5569984c4197SStefano Zampini 557008122e43SStefano Zampini total_counts = 0; 557108122e43SStefano Zampini n_vertices = 0; 5572d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 5573d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 557408122e43SStefano Zampini } 557508122e43SStefano Zampini max_constraints = 0; 55769162d606SStefano Zampini total_counts_cc = 0; 557708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 557808122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 55799162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 558008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 558108122e43SStefano Zampini } 55829162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 55839162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 55849162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 55859162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 558674d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 55879162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 55889162d606SStefano Zampini total_counts_cc = 0; 55899162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 55909162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 55919162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 559208122e43SStefano Zampini } 559308122e43SStefano Zampini } 55949162d606SStefano Zampini #if 0 55959162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 55969162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 55979162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 55989162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 55999162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 56009162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 56019162d606SStefano Zampini } 56029162d606SStefano Zampini printf("\n"); 56039162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 56049162d606SStefano Zampini } 56051b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 56068bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 56071b968477SStefano Zampini } 56081b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 56098bec7fa6SStefano 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]); 56101b968477SStefano Zampini } 561108122e43SStefano Zampini #endif 561208122e43SStefano Zampini 56138bec7fa6SStefano Zampini max_size_of_constraint = 0; 56149162d606SStefano 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]); 56159162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 561608122e43SStefano Zampini /* Change of basis */ 5617b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 561808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 561908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 562008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 5621b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 562208122e43SStefano Zampini } 562308122e43SStefano Zampini } 562408122e43SStefano Zampini } 562508122e43SStefano Zampini } 5626984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 56274f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 562808122e43SStefano Zampini 56299162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 56309162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 5631eee23b56SStefano 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); 5632674ae819SStefano Zampini 5633674ae819SStefano Zampini /* Create constraint matrix */ 5634674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 563516f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 5636984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 5637984c4197SStefano Zampini 5638984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 5639a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 5640a717540cSStefano Zampini qr_needed = PETSC_FALSE; 564174d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 5642984c4197SStefano Zampini total_primal_vertices=0; 5643b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 56449162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 56459162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 564672b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 56479162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 5648b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 564964efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 56509162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 56519162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 5652a717540cSStefano Zampini } 5653b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 565491af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 5655a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 5656a717540cSStefano Zampini qr_needed = PETSC_TRUE; 5657a717540cSStefano Zampini } 5658fa434743SStefano Zampini } else { 5659b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 5660fa434743SStefano Zampini } 5661a717540cSStefano Zampini } 5662b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 5663b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 5664674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 566570022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 56664f1b2e48SStefano 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); 56670e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 56680e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 5669984c4197SStefano Zampini 5670984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 567174d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 5672785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 5673984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 567474d5cdf7SStefano Zampini 5675984c4197SStefano Zampini j = total_primal_vertices; 567674d5cdf7SStefano Zampini total_counts = total_primal_vertices; 5677b3d85658SStefano Zampini cum = total_primal_vertices; 56789162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 56794641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 5680b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 5681b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 5682b3d85658SStefano Zampini cum++; 56839162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 568474d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 568574d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 568674d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 568774d5cdf7SStefano Zampini } 56889162d606SStefano Zampini j += constraints_n[i]; 5689674ae819SStefano Zampini } 5690674ae819SStefano Zampini } 5691674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 5692674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 5693088faed8SStefano Zampini 5694674ae819SStefano Zampini /* set values in constraint matrix */ 5695984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 56960e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5697674ae819SStefano Zampini } 5698984c4197SStefano Zampini total_counts = total_primal_vertices; 56999162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 57004641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 57019162d606SStefano Zampini PetscInt *cols; 57029162d606SStefano Zampini 57039162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 57049162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 57059162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 57069162d606SStefano Zampini PetscInt row = total_counts+k; 57079162d606SStefano Zampini PetscScalar *vals; 57089162d606SStefano Zampini 57099162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 57109162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 57119162d606SStefano Zampini } 57129162d606SStefano Zampini total_counts += constraints_n[i]; 5713674ae819SStefano Zampini } 5714674ae819SStefano Zampini } 5715674ae819SStefano Zampini /* assembling */ 5716674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5717674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5718088faed8SStefano Zampini 5719984c4197SStefano Zampini /* 57206a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5721984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 5722f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5723984c4197SStefano Zampini */ 5724674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 5725674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 5726026de310SStefano Zampini /* dual and primal dofs on a single cc */ 5727984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 5728984c4197SStefano Zampini /* working stuff for GEQRF */ 572981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 5730984c4197SStefano Zampini PetscBLASInt lqr_work; 5731984c4197SStefano Zampini /* working stuff for UNGQR */ 5732984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 5733984c4197SStefano Zampini PetscBLASInt lgqr_work; 5734984c4197SStefano Zampini /* working stuff for TRTRS */ 5735984c4197SStefano Zampini PetscScalar *trs_rhs; 57363f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 5737984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 5738984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 5739984c4197SStefano Zampini PetscScalar *start_vals; 5740984c4197SStefano Zampini /* working stuff for values insertion */ 57414641a718SStefano Zampini PetscBT is_primal; 574264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 5743906d46d4SStefano Zampini /* matrix sizes */ 5744906d46d4SStefano Zampini PetscInt global_size,local_size; 5745906d46d4SStefano Zampini /* temporary change of basis */ 5746906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 5747cf5a6209SStefano Zampini /* extra space for debugging */ 5748cf5a6209SStefano Zampini PetscScalar *dbg_work; 5749984c4197SStefano Zampini 5750906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 5751906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 575216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 5753bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 5754906d46d4SStefano Zampini /* nonzeros for local mat */ 5755bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 57561dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5757bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 57581dd7afcfSStefano Zampini } else { 57591dd7afcfSStefano Zampini const PetscInt *ii; 57601dd7afcfSStefano Zampini PetscInt n; 57611dd7afcfSStefano Zampini PetscBool flg_row; 57621dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 57631dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 57641dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 57651dd7afcfSStefano Zampini } 57669162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 5767a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 57689162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 5769a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 57709162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 5771a717540cSStefano Zampini } else { 57729162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 57739162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 5774a717540cSStefano Zampini } 5775a717540cSStefano Zampini } 5776a717540cSStefano Zampini } 5777906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 5778bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 57791dd7afcfSStefano Zampini /* Set interior change in the matrix */ 57801dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5781bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 5782906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 5783a717540cSStefano Zampini } 57841dd7afcfSStefano Zampini } else { 57851dd7afcfSStefano Zampini const PetscInt *ii,*jj; 57861dd7afcfSStefano Zampini PetscScalar *aa; 57871dd7afcfSStefano Zampini PetscInt n; 57881dd7afcfSStefano Zampini PetscBool flg_row; 57891dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 57901dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 57911dd7afcfSStefano Zampini for (i=0;i<n;i++) { 57921dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 57931dd7afcfSStefano Zampini } 57941dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 57951dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 57961dd7afcfSStefano Zampini } 5797a717540cSStefano Zampini 5798a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5799a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5800a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5801a717540cSStefano Zampini } 5802a717540cSStefano Zampini 5803a717540cSStefano Zampini 5804a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 5805a717540cSStefano Zampini /* 5806a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 5807a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 5808a717540cSStefano Zampini 5809a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 5810a717540cSStefano Zampini 5811a6b551f4SStefano 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) 5812a6b551f4SStefano Zampini 5813a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 5814a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 5815a717540cSStefano Zampini | ... | 5816a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 5817a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 5818a717540cSStefano Zampini 5819a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 5820a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 5821a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 5822a6b551f4SStefano Zampini 5823a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 5824a717540cSStefano Zampini */ 5825a717540cSStefano Zampini if (qr_needed) { 5826984c4197SStefano Zampini /* space to store Q */ 5827854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 58284e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 58294e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 5830984c4197SStefano Zampini /* first we issue queries for optimal work */ 58313f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 58323f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 58333f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5834984c4197SStefano Zampini lqr_work = -1; 58353f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 5836984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 5837984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 5838785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 5839984c4197SStefano Zampini lgqr_work = -1; 58403f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 58413f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 58423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 58433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 58443f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 58453f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 5846984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 5847984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 5848785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 5849984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 5850785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 5851a717540cSStefano Zampini /* allocating workspace for check */ 5852a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5853cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 5854a717540cSStefano Zampini } 5855a717540cSStefano Zampini } 5856984c4197SStefano Zampini /* array to store whether a node is primal or not */ 58574641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 5858473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 58590e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 5860eee23b56SStefano 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); 586139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 586239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 586339e2fb2aSStefano Zampini } 586439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 5865984c4197SStefano Zampini 5866a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 58679162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 58689162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 58694641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 5870984c4197SStefano Zampini /* get constraint info */ 58719162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 5872984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 5873984c4197SStefano Zampini 5874984c4197SStefano Zampini if (pcbddc->dbg_flag) { 58759162d606SStefano 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); 5876674ae819SStefano Zampini } 5877984c4197SStefano Zampini 5878fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 5879a717540cSStefano Zampini 5880a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 5881a717540cSStefano Zampini if (pcbddc->dbg_flag) { 58829162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5883a717540cSStefano Zampini } 5884984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 58859162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5886984c4197SStefano Zampini 5887984c4197SStefano Zampini /* compute QR decomposition of constraints */ 58883f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 58893f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 58903f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5891674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 58923f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 5893984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 5894674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5895984c4197SStefano Zampini 5896984c4197SStefano Zampini /* explictly compute R^-T */ 5897984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 5898984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 58993f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 59003f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 59013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 59023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 5903984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59043f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 5905984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 5906984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5907984c4197SStefano Zampini 5908a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 59093f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 59103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 59113f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 59123f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5913984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59143f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 5915984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 5916984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5917984c4197SStefano Zampini 5918984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 5919984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 5920984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 59213f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 59223f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 59233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 59243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 59253f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 59263f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5927984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59289162d606SStefano 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)); 5929984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 59309162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5931984c4197SStefano Zampini 5932984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 59339162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 5934984c4197SStefano Zampini /* insert cols for primal dofs */ 5935984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 5936984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 59379162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 5938906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 5939984c4197SStefano Zampini } 5940984c4197SStefano Zampini /* insert cols for dual dofs */ 5941984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 59429162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 5943984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 59449162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 5945906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 5946984c4197SStefano Zampini j++; 5947674ae819SStefano Zampini } 5948674ae819SStefano Zampini } 5949984c4197SStefano Zampini 5950984c4197SStefano Zampini /* check change of basis */ 5951984c4197SStefano Zampini if (pcbddc->dbg_flag) { 5952984c4197SStefano Zampini PetscInt ii,jj; 5953984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 5954c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 5955c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5956c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 5957c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5958c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 5959c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 5960984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5961cf5a6209SStefano 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)); 5962984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5963984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 5964984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 5965cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 5966cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) valid_qr = PETSC_FALSE; 5967674ae819SStefano Zampini } 5968674ae819SStefano Zampini } 5969984c4197SStefano Zampini if (!valid_qr) { 597022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 5971984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 5972984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 5973cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 5974cf5a6209SStefano 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])); 5975674ae819SStefano Zampini } 5976cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 5977cf5a6209SStefano 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])); 5978984c4197SStefano Zampini } 5979984c4197SStefano Zampini } 5980984c4197SStefano Zampini } 5981674ae819SStefano Zampini } else { 598222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 5983674ae819SStefano Zampini } 5984674ae819SStefano Zampini } 5985a717540cSStefano Zampini } else { /* simple transformation block */ 5986a717540cSStefano Zampini PetscInt row,col; 5987a6b551f4SStefano Zampini PetscScalar val,norm; 5988a6b551f4SStefano Zampini 5989a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 59909162d606SStefano 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)); 5991a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 59929162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 59939162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 5994bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 59959162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 5996906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 59979162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 5998a717540cSStefano Zampini } else { 5999a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 60009162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6001a717540cSStefano Zampini if (row != col) { 60029162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 6003a717540cSStefano Zampini } else { 60049162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 6005a717540cSStefano Zampini } 6006906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 6007a717540cSStefano Zampini } 6008a717540cSStefano Zampini } 6009a717540cSStefano Zampini } 601098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 601122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 6012a717540cSStefano Zampini } 6013674ae819SStefano Zampini } 6014984c4197SStefano Zampini } else { 6015984c4197SStefano Zampini if (pcbddc->dbg_flag) { 60169162d606SStefano 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); 6017674ae819SStefano Zampini } 6018674ae819SStefano Zampini } 6019674ae819SStefano Zampini } 6020a717540cSStefano Zampini 6021a717540cSStefano Zampini /* free workspace */ 6022a717540cSStefano Zampini if (qr_needed) { 6023984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6024cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 6025984c4197SStefano Zampini } 6026984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 6027984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 6028984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 6029984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 6030984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 6031674ae819SStefano Zampini } 6032a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 6033906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6034906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6035906d46d4SStefano Zampini 6036906d46d4SStefano Zampini /* assembling of global change of variable */ 603788c03ad3SStefano Zampini if (!pcbddc->fake_change) { 6038bbb9e6c6SStefano Zampini Mat tmat; 603916f15bc4SStefano Zampini PetscInt bs; 604016f15bc4SStefano Zampini 6041906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 6042906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 6043bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 6044bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 6045bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6046bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 604716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 604816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 6049906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 6050bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 6051bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6052bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6053bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6054bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6055e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6056e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6057bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 6058bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 605988c03ad3SStefano Zampini 6060906d46d4SStefano Zampini /* check */ 6061906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 6062906d46d4SStefano Zampini PetscReal error; 6063906d46d4SStefano Zampini Vec x,x_change; 6064906d46d4SStefano Zampini 6065906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 6066906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 6067906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 6068906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 6069e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6070e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6071bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 6072e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6073e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6074906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 6075906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 6076906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 6077637e8532SStefano Zampini if (error > PETSC_SMALL) { 6078637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 6079637e8532SStefano Zampini } 6080906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 6081906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 6082906d46d4SStefano Zampini } 6083b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 6084b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 6085b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6086bf3a8328SStefano Zampini 60879a962809SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 6088b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 6089ac632422SStefano Zampini Mat S_new,tmat; 6090bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 6091bbb9e6c6SStefano Zampini 6092bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 60936816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 6094bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6095bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 6096bf3a8328SStefano Zampini IS is_V; 6097b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 6098b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 6099b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 6100b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 6101b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 6102bf3a8328SStefano Zampini } 6103bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 6104ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6105b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 6106ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6107bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6108bf3a8328SStefano Zampini const PetscScalar *array; 6109bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 6110bf3a8328SStefano Zampini PetscInt i,n_V; 6111bf3a8328SStefano Zampini 6112b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6113b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 6114b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6115b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6116b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 6117b087196eSStefano Zampini for (i=0;i<n_V;i++) { 6118b087196eSStefano Zampini PetscScalar val; 6119b087196eSStefano Zampini PetscInt idx; 6120b087196eSStefano Zampini 6121b087196eSStefano Zampini idx = idxs_V[i]; 6122b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 6123b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 6124b087196eSStefano Zampini } 6125b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6126b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6127bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 6128bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6129bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6130bf3a8328SStefano Zampini } 6131ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 6132ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6133ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 6134ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6135b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 6136ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6137bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6138b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6139bf3a8328SStefano Zampini } 6140ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 6141ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6142ac632422SStefano Zampini } 6143b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 614488c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6145b96c3477SStefano Zampini } 6146c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 6147b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 6148c9db6a07SStefano Zampini PetscInt i; 6149c9db6a07SStefano Zampini 6150c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6151c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 6152c9db6a07SStefano Zampini } 6153c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 6154c9db6a07SStefano Zampini } 6155b96c3477SStefano Zampini } 615616909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 615716909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 615816909a7fSStefano Zampini } else { 6159906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 616016909a7fSStefano Zampini } 61611dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 616227b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 616372b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 616472b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 616572b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 616672b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 616772b8c272SStefano Zampini } 61681dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 616927b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 6170b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 6171b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 6172906d46d4SStefano Zampini } else { 61731dd7afcfSStefano Zampini Mat benign_global = NULL; 617427b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 61751dd7afcfSStefano Zampini Mat tmat; 61761dd7afcfSStefano Zampini 61771dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 61781dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 61791dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 61801dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 61811dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 61821dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 61831dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 61841dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 61851dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 61861dd7afcfSStefano Zampini if (pcbddc->benign_change) { 61871dd7afcfSStefano Zampini Mat M; 61881dd7afcfSStefano Zampini 61891dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 61901dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 61911dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 61921dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 6193906d46d4SStefano Zampini } else { 61941dd7afcfSStefano Zampini Mat eye; 61951dd7afcfSStefano Zampini PetscScalar *array; 61961dd7afcfSStefano Zampini 61971dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 61981dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 61991dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 62001dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 6201906d46d4SStefano Zampini } 62021dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 62031dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 62041dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 62051dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 62061dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 62071dd7afcfSStefano Zampini } 62081dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 62091dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 62101dd7afcfSStefano Zampini } 62111dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 62121dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 62131dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 621427b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 62151dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 62161dd7afcfSStefano Zampini } 62171dd7afcfSStefano Zampini } 621816909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 621916909a7fSStefano Zampini IS is_global; 622016909a7fSStefano Zampini const PetscInt *gidxs; 622116909a7fSStefano Zampini 622216909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 622316909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 622416909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 622516909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 622616909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 622716909a7fSStefano Zampini } 62281dd7afcfSStefano Zampini } 62291dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 62301dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 6231b9b85e73SStefano Zampini } 6232a717540cSStefano Zampini 623372b8c272SStefano Zampini if (!pcbddc->fake_change) { 62344f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 62354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 62364f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 62374f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 6238019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 6239019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 6240019a44ceSStefano Zampini pcbddc->local_primal_size++; 6241019a44ceSStefano Zampini } 6242019a44ceSStefano Zampini 6243019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 6244727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 6245727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 62469f47a83aSStefano 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); 6247c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 62480e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 62499f47a83aSStefano 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); 6250727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 6251727cdba6SStefano Zampini } 62520e6343abSStefano Zampini } 6253727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 6254b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 625572b8c272SStefano Zampini } 625672b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 6257727cdba6SStefano Zampini 6258a717540cSStefano Zampini /* flush dbg viewer */ 6259b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 6260b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6261b8ffe317SStefano Zampini } 6262a717540cSStefano Zampini 6263e310c8b4SStefano Zampini /* free workspace */ 6264a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 62654641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 626608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 62679162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 62689162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 626908122e43SStefano Zampini } else { 62709162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 62719162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 62729162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 627308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 627408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 62759162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 62769162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 627708122e43SStefano Zampini } 6278674ae819SStefano Zampini PetscFunctionReturn(0); 6279674ae819SStefano Zampini } 6280674ae819SStefano Zampini 6281674ae819SStefano Zampini #undef __FUNCT__ 6282674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 6283674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 6284674ae819SStefano Zampini { 628571582508SStefano Zampini ISLocalToGlobalMapping map; 6286674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 6287674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 6288*66da6bd7Sstefano_zampini PetscInt i,N; 6289*66da6bd7Sstefano_zampini PetscBool rcsr = PETSC_FALSE; 6290*66da6bd7Sstefano_zampini PetscErrorCode ierr; 6291674ae819SStefano Zampini 6292674ae819SStefano Zampini PetscFunctionBegin; 62938af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 6294b03ebc13SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 62958e61c736SStefano Zampini /* Reset previously computed graph */ 62968e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 6297674ae819SStefano Zampini /* Init local Graph struct */ 62987fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 629971582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 6300be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 6301674ae819SStefano Zampini 6302575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 63039a962809SStefano 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); 63049577ea80SStefano Zampini 6305674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 6306*66da6bd7Sstefano_zampini if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) { 63074d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 63084d379d7bSStefano Zampini PetscInt nvtxs; 6309e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 6310674ae819SStefano Zampini 63112fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 63122fffb893SStefano Zampini if (flg_row) { 63134d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 6314b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 63152fffb893SStefano Zampini } 63162fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 6317*66da6bd7Sstefano_zampini rcsr = PETSC_TRUE; 6318674ae819SStefano Zampini } 63199b28b941SStefano Zampini if (pcbddc->dbg_flag) { 63209b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6321674ae819SStefano Zampini } 6322674ae819SStefano Zampini 6323674ae819SStefano Zampini /* Setup of Graph */ 63244b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 632514f95afaSStefano 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); 6326674ae819SStefano Zampini 63274f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 63284f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 63294f1b2e48SStefano Zampini PetscInt *local_subs; 63304f1b2e48SStefano Zampini 63314f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 63324f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 63334f1b2e48SStefano Zampini const PetscInt *idxs; 63344f1b2e48SStefano Zampini PetscInt nl,j; 63354f1b2e48SStefano Zampini 63364f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 63374f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 633871582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 63394f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 63404f1b2e48SStefano Zampini } 63414f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 63424f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 63434f1b2e48SStefano Zampini } 63448af8fcf9SStefano Zampini } 63454f1b2e48SStefano Zampini 6346cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 6347674ae819SStefano Zampini /* Graph's connected components analysis */ 6348674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 634971582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 63508af8fcf9SStefano Zampini } 6351*66da6bd7Sstefano_zampini if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0; 6352674ae819SStefano Zampini PetscFunctionReturn(0); 6353674ae819SStefano Zampini } 6354674ae819SStefano Zampini 63559a7d3425SStefano Zampini #undef __FUNCT__ 63569a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 63579a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 63589a7d3425SStefano Zampini { 63599a7d3425SStefano Zampini PetscInt i,j; 63609a7d3425SStefano Zampini PetscScalar *alphas; 63619a7d3425SStefano Zampini PetscErrorCode ierr; 63629a7d3425SStefano Zampini 63639a7d3425SStefano Zampini PetscFunctionBegin; 6364785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 63659a7d3425SStefano Zampini for (i=0;i<n;i++) { 63669a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 6367669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 6368669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 6369669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 63709a7d3425SStefano Zampini } 63719a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 63729a7d3425SStefano Zampini PetscFunctionReturn(0); 63739a7d3425SStefano Zampini } 63749a7d3425SStefano Zampini 6375e7931f94SStefano Zampini #undef __FUNCT__ 6376bb360cb4SStefano Zampini #define __FUNCT__ "PCBDDCMatISGetSubassemblingPattern" 6377bb360cb4SStefano Zampini PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 6378e7931f94SStefano Zampini { 637957de7509SStefano Zampini Mat A; 6380e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 6381e7931f94SStefano Zampini PetscMPIInt size,rank,color; 638252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 638352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 6384bb360cb4SStefano Zampini PetscInt im_active,active_procs,N,n,i,j,threshold = 2; 638557de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 638627b6a85dSStefano Zampini PetscInt xadj_count,*count; 638727b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 638827b6a85dSStefano Zampini PetscSubcomm psubcomm; 638927b6a85dSStefano Zampini MPI_Comm subcomm; 639052e5ac9dSStefano Zampini PetscErrorCode ierr; 6391a57a6d2fSStefano Zampini 6392e7931f94SStefano Zampini PetscFunctionBegin; 639357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 639457de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 639557de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 639657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 639757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 639857de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 639957de7509SStefano Zampini 640057de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 640157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 640257de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 640357de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 640457de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 6405bb360cb4SStefano Zampini im_active = !!n; 640657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 640757de7509SStefano Zampini void_procs = size - active_procs; 640857de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 640957de7509SStefano Zampini if (void_procs) { 641057de7509SStefano Zampini PetscInt ncand; 641157de7509SStefano Zampini 641257de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 641357de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 641457de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 641557de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 641657de7509SStefano Zampini if (!procs_candidates[i]) { 641757de7509SStefano Zampini procs_candidates[ncand++] = i; 641857de7509SStefano Zampini } 641957de7509SStefano Zampini } 642057de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 642157de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 642257de7509SStefano Zampini } 642357de7509SStefano Zampini 6424bb360cb4SStefano Zampini /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix 642514f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 6426bb360cb4SStefano Zampini ierr = MatGetSize(mat,&N,NULL);CHKERRQ(ierr); 6427bb360cb4SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) { 642814f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 642914f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 643014f0bfb9SStefano Zampini else dest = rank; 643157de7509SStefano Zampini if (im_active) { 643257de7509SStefano Zampini issize = 1; 643357de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 643414f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 643557de7509SStefano Zampini } else { 643614f0bfb9SStefano Zampini isidx = dest; 643757de7509SStefano Zampini } 643857de7509SStefano Zampini } else { 643957de7509SStefano Zampini issize = 0; 644057de7509SStefano Zampini isidx = -1; 644157de7509SStefano Zampini } 6442bb360cb4SStefano Zampini if (*n_subdomains != 1) *n_subdomains = active_procs; 644357de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 6444daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 644557de7509SStefano Zampini PetscFunctionReturn(0); 644657de7509SStefano Zampini } 6447c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 6448c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 644927b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 6450e7931f94SStefano Zampini 6451e7931f94SStefano Zampini /* Get info on mapping */ 64523bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6453e7931f94SStefano Zampini 6454e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 6455785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 6456e7931f94SStefano Zampini xadj[0] = 0; 6457e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 6458785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 6459785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 6460bb360cb4SStefano Zampini ierr = PetscCalloc1(n,&count);CHKERRQ(ierr); 646127b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 646227b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 646327b6a85dSStefano Zampini count[shared[i][j]] += 1; 6464e7931f94SStefano Zampini 646527b6a85dSStefano Zampini xadj_count = 0; 64662b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 646727b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 646827b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 6469d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 6470d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 6471d023bfaeSStefano Zampini xadj_count++; 647227b6a85dSStefano Zampini break; 647327b6a85dSStefano Zampini } 6474e7931f94SStefano Zampini } 6475e7931f94SStefano Zampini } 6476d023bfaeSStefano Zampini xadj[1] = xadj_count; 647727b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 64783bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6479e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 6480e7931f94SStefano Zampini 64813837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 6482e7931f94SStefano Zampini 648327b6a85dSStefano Zampini /* Restrict work on active processes only */ 648427b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 648527b6a85dSStefano Zampini if (void_procs) { 648627b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 648727b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 648827b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 648927b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 649027b6a85dSStefano Zampini } else { 649127b6a85dSStefano Zampini psubcomm = NULL; 649227b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 649327b6a85dSStefano Zampini } 649427b6a85dSStefano Zampini 649527b6a85dSStefano Zampini v_wgt = NULL; 649627b6a85dSStefano Zampini if (!color) { 6497e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 6498e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 6499e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6500c8587f34SStefano Zampini } else { 650152e5ac9dSStefano Zampini Mat subdomain_adj; 650252e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 650352e5ac9dSStefano Zampini MatPartitioning partitioner; 650427b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 650552e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 650657de7509SStefano Zampini PetscMPIInt size; 6507b0c7d250SStefano Zampini PetscBool aggregate; 6508b0c7d250SStefano Zampini 650927b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 651027b6a85dSStefano Zampini if (void_procs) { 651127b6a85dSStefano Zampini PetscInt prank = rank; 6512785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 651327b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 6514e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 6515e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 6516c8587f34SStefano Zampini } 6517e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 651827b6a85dSStefano Zampini } else { 651927b6a85dSStefano Zampini oldranks = NULL; 652027b6a85dSStefano Zampini } 6521b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 652227b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 6523b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 6524b0c7d250SStefano Zampini PetscMPIInt nrank; 6525b0c7d250SStefano Zampini PetscScalar *vals; 6526b0c7d250SStefano Zampini 652727b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 6528b0c7d250SStefano Zampini lrows = 0; 6529b0c7d250SStefano Zampini if (nrank<redprocs) { 6530b0c7d250SStefano Zampini lrows = size/redprocs; 6531b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 6532b0c7d250SStefano Zampini } 653327b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 6534b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 6535b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6536b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6537b0c7d250SStefano Zampini row = nrank; 6538b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 6539b0c7d250SStefano Zampini cols = adjncy; 6540b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 6541b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 6542b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 6543b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6544b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 654552e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 654652e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 654752e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6548b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 654927b6a85dSStefano Zampini if (use_vwgt) { 655027b6a85dSStefano Zampini Vec v; 655127b6a85dSStefano Zampini const PetscScalar *array; 655227b6a85dSStefano Zampini PetscInt nl; 655327b6a85dSStefano Zampini 655427b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 6555bb360cb4SStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)n,INSERT_VALUES);CHKERRQ(ierr); 655627b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 655727b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 655827b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 655927b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 656027b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 656122db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 656227b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 656327b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 656427b6a85dSStefano Zampini } 6565b0c7d250SStefano Zampini } else { 656627b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 656727b6a85dSStefano Zampini if (use_vwgt) { 656827b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 6569bb360cb4SStefano Zampini v_wgt[0] = n; 657027b6a85dSStefano Zampini } 6571b0c7d250SStefano Zampini } 657222b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 6573e7931f94SStefano Zampini 6574e7931f94SStefano Zampini /* Partition */ 657527b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 6576e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 657727b6a85dSStefano Zampini if (v_wgt) { 6578e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 6579c8587f34SStefano Zampini } 658057de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 658157de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 6582e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 6583e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 658422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 6585e7931f94SStefano Zampini 658652e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 65876583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 658852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 658952e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 659057de7509SStefano Zampini if (!aggregate) { 659157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 659227b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 659327b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 659427b6a85dSStefano Zampini #endif 659557de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 659627b6a85dSStefano Zampini } else if (oldranks) { 6597b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 659827b6a85dSStefano Zampini } else { 659927b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 660057de7509SStefano Zampini } 660128143c3dSStefano Zampini } else { 6602b0c7d250SStefano Zampini PetscInt idxs[1]; 6603b0c7d250SStefano Zampini PetscMPIInt tag; 6604b0c7d250SStefano Zampini MPI_Request *reqs; 6605b0c7d250SStefano Zampini 6606b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 6607b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 6608b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 660927b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 661028143c3dSStefano Zampini } 661127b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 6612b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6613b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 661457de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 661527b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 661627b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 661727b6a85dSStefano Zampini #endif 661857de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 661927b6a85dSStefano Zampini } else if (oldranks) { 6620b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 662127b6a85dSStefano Zampini } else { 662227b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 6623e7931f94SStefano Zampini } 662457de7509SStefano Zampini } 662552e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6626e7931f94SStefano Zampini /* clean up */ 6627e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 662852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 6629e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 6630e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 6631e7931f94SStefano Zampini } 663227b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 663357de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 6634e7931f94SStefano Zampini 6635e7931f94SStefano Zampini /* assemble parallel IS for sends */ 6636e7931f94SStefano Zampini i = 1; 663727b6a85dSStefano Zampini if (!color) i=0; 663857de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 6639e7931f94SStefano Zampini PetscFunctionReturn(0); 6640e7931f94SStefano Zampini } 6641e7931f94SStefano Zampini 6642e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 6643e7931f94SStefano Zampini 6644e7931f94SStefano Zampini #undef __FUNCT__ 66451e0482f5SStefano Zampini #define __FUNCT__ "PCBDDCMatISSubassemble" 66461e0482f5SStefano 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[]) 6647e7931f94SStefano Zampini { 664870cf5478SStefano Zampini Mat local_mat; 6649e7931f94SStefano Zampini IS is_sends_internal; 66509d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 66511ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 66529d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 6653e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 6654e7931f94SStefano Zampini PetscInt* l2gmap_indices; 6655e7931f94SStefano Zampini const PetscInt* is_indices; 6656e7931f94SStefano Zampini MatType new_local_type; 6657e7931f94SStefano Zampini /* buffers */ 6658e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 665928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 66609d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 6661e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 66621ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 6663e7931f94SStefano Zampini /* MPI */ 666428143c3dSStefano Zampini MPI_Comm comm,comm_n; 666528143c3dSStefano Zampini PetscSubcomm subcomm; 6666e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 666728143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 666828143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 66691ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 66701ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 66711ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 6672e7931f94SStefano Zampini PetscErrorCode ierr; 6673e7931f94SStefano Zampini 6674e7931f94SStefano Zampini PetscFunctionBegin; 667557de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 6676e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 667728143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 667857de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 667957de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 668057de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 668157de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 668257de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 66831ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 66841ae86dd6SStefano Zampini if (nvecs) { 66851ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 66861ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 66871ae86dd6SStefano Zampini } 668857de7509SStefano Zampini /* further checks */ 6689e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6690e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 6691e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 6692e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 6693e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 669457de7509SStefano Zampini if (reuse && *mat_n) { 669570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 669657de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 669770cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 669828143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 669970cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 670070cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 670170cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 670270cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 670370cf5478SStefano Zampini } 6704e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 6705e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 670657de7509SStefano Zampini 6707e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 6708e7931f94SStefano Zampini if (!is_sends) { 670928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 6710bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 6711c8587f34SStefano Zampini } else { 6712e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 6713e7931f94SStefano Zampini is_sends_internal = is_sends; 6714c8587f34SStefano Zampini } 6715e7931f94SStefano Zampini 6716e7931f94SStefano Zampini /* get comm */ 6717a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 6718e7931f94SStefano Zampini 6719e7931f94SStefano Zampini /* compute number of sends */ 6720e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 6721e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 6722e7931f94SStefano Zampini 6723e7931f94SStefano Zampini /* compute number of receives */ 6724e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 6725785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 6726e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 6727e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6728e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 6729e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 6730e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 6731e7931f94SStefano Zampini 673228143c3dSStefano Zampini /* restrict comm if requested */ 673328143c3dSStefano Zampini subcomm = 0; 673428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 673528143c3dSStefano Zampini if (restrict_comm) { 6736779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 6737779c1cceSStefano Zampini 673828143c3dSStefano Zampini color = 0; 673953a05cb3SStefano Zampini if (restrict_full) { 674053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 674153a05cb3SStefano Zampini } else { 674253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 674353a05cb3SStefano Zampini } 6744b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 674528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 674628143c3dSStefano Zampini /* check if reuse has been requested */ 674757de7509SStefano Zampini if (reuse) { 674828143c3dSStefano Zampini if (*mat_n) { 674928143c3dSStefano Zampini PetscMPIInt subcommsize2; 675028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 675128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 675228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 675328143c3dSStefano Zampini } else { 675428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 675528143c3dSStefano Zampini } 675628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 6757779c1cceSStefano Zampini PetscMPIInt rank; 6758779c1cceSStefano Zampini 6759779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 676028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 676128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 676228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 6763306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 676428143c3dSStefano Zampini } 676528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 676628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 676728143c3dSStefano Zampini } else { 676828143c3dSStefano Zampini comm_n = comm; 676928143c3dSStefano Zampini } 677028143c3dSStefano Zampini 6771e7931f94SStefano Zampini /* prepare send/receive buffers */ 6772785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 6773e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 6774785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 6775e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 677628143c3dSStefano Zampini if (nis) { 6777854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 677828143c3dSStefano Zampini } 6779e7931f94SStefano Zampini 678028143c3dSStefano Zampini /* Get data from local matrices */ 67816c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 6782e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 6783e7931f94SStefano Zampini /* 6784e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 6785e7931f94SStefano Zampini send_buffer_idxs should contain: 6786e7931f94SStefano Zampini - MatType_PRIVATE type 6787e7931f94SStefano Zampini - PetscInt size_of_l2gmap 6788e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 6789e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 6790e7931f94SStefano Zampini */ 67916c4ed002SBarry Smith else { 6792e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 67933bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 6794854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 6795e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 6796e7931f94SStefano Zampini send_buffer_idxs[1] = i; 67973bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6798e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 67993bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6800e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 6801e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6802e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 6803e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 6804c8587f34SStefano Zampini } 6805c8587f34SStefano Zampini } 6806e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 680728143c3dSStefano Zampini /* additional is (if any) */ 680828143c3dSStefano Zampini if (nis) { 680928143c3dSStefano Zampini PetscMPIInt psum; 681028143c3dSStefano Zampini PetscInt j; 681128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 681228143c3dSStefano Zampini PetscInt plen; 681328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 681428143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 681528143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 681628143c3dSStefano Zampini } 6817854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 681828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 681928143c3dSStefano Zampini PetscInt plen; 682028143c3dSStefano Zampini const PetscInt *is_array_idxs; 682128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 682228143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 682328143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 682428143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 682528143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 682628143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 682728143c3dSStefano Zampini } 682828143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 682928143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 683028143c3dSStefano Zampini } 683128143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 683228143c3dSStefano Zampini } 683328143c3dSStefano Zampini 6834e7931f94SStefano Zampini buf_size_idxs = 0; 6835e7931f94SStefano Zampini buf_size_vals = 0; 683628143c3dSStefano Zampini buf_size_idxs_is = 0; 68371ae86dd6SStefano Zampini buf_size_vecs = 0; 6838e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6839e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 6840e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 684128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 68421ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 6843e7931f94SStefano Zampini } 6844785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 6845785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 684695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 68471ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 6848e7931f94SStefano Zampini 6849e7931f94SStefano Zampini /* get new tags for clean communications */ 6850e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 6851e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 685228143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 68531ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 6854e7931f94SStefano Zampini 6855e7931f94SStefano Zampini /* allocate for requests */ 6856785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 6857785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 685895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 68591ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 6860785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 6861785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 686295ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 68631ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 6864e7931f94SStefano Zampini 6865e7931f94SStefano Zampini /* communications */ 6866e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6867e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 686828143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 68691ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 6870e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6871e7931f94SStefano Zampini source_dest = onodes[i]; 6872e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 6873e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 6874e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6875e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 687628143c3dSStefano Zampini if (nis) { 687757de7509SStefano Zampini source_dest = onodes_is[i]; 687828143c3dSStefano 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); 687928143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 688028143c3dSStefano Zampini } 68811ae86dd6SStefano Zampini if (nvecs) { 68821ae86dd6SStefano Zampini source_dest = onodes[i]; 68831ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 68841ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 68851ae86dd6SStefano Zampini } 6886e7931f94SStefano Zampini } 6887e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6888e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 6889e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 6890e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 689128143c3dSStefano Zampini if (nis) { 689228143c3dSStefano 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); 689328143c3dSStefano Zampini } 68941ae86dd6SStefano Zampini if (nvecs) { 68951ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 68961ae86dd6SStefano 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); 68971ae86dd6SStefano Zampini } 6898e7931f94SStefano Zampini } 6899e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6900e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 6901e7931f94SStefano Zampini 6902e7931f94SStefano Zampini /* assemble new l2g map */ 6903e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6904e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 69059d30be91SStefano Zampini new_local_rows = 0; 6906e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 69079d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6908e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6909e7931f94SStefano Zampini } 69109d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 6911e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 69129d30be91SStefano Zampini new_local_rows = 0; 6913e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 69149d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 69159d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6916e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6917e7931f94SStefano Zampini } 69189d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 69199d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 6920e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 6921e7931f94SStefano Zampini 6922e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 6923e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 6924e7931f94SStefano 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) */ 6925e7931f94SStefano Zampini if (n_recvs) { 692628143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 6927e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6928e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6929e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 6930e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 6931e7931f94SStefano Zampini break; 6932e7931f94SStefano Zampini } 6933e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6934e7931f94SStefano Zampini } 6935e7931f94SStefano Zampini switch (new_local_type_private) { 693628143c3dSStefano Zampini case MATDENSE_PRIVATE: 693728143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 6938e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 6939e7931f94SStefano Zampini bs = 1; 694028143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 694128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 694228143c3dSStefano Zampini bs = 1; 694328143c3dSStefano Zampini } 6944e7931f94SStefano Zampini break; 6945e7931f94SStefano Zampini case MATAIJ_PRIVATE: 6946e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 6947e7931f94SStefano Zampini bs = 1; 6948e7931f94SStefano Zampini break; 6949e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 6950e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 6951e7931f94SStefano Zampini break; 6952e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 6953e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 6954e7931f94SStefano Zampini break; 6955e7931f94SStefano Zampini default: 69569d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 6957e7931f94SStefano Zampini break; 6958e7931f94SStefano Zampini } 695928143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 696028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 696128143c3dSStefano Zampini bs = 1; 6962e7931f94SStefano Zampini } 6963e7931f94SStefano Zampini 696470cf5478SStefano Zampini /* create MATIS object if needed */ 696557de7509SStefano Zampini if (!reuse) { 6966e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 6967e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 696870cf5478SStefano Zampini } else { 696970cf5478SStefano Zampini /* it also destroys the local matrices */ 697057de7509SStefano Zampini if (*mat_n) { 697170cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 697257de7509SStefano Zampini } else { /* this is a fake object */ 697357de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 697457de7509SStefano Zampini } 697570cf5478SStefano Zampini } 697670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 6977e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 69789d30be91SStefano Zampini 69799d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 69809d30be91SStefano Zampini 69819d30be91SStefano Zampini /* Global to local map of received indices */ 69829d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 69839d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 69849d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 69859d30be91SStefano Zampini 69869d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 69879d30be91SStefano Zampini buf_size_idxs = 0; 69889d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 69899d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 69909d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 69919d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 69929d30be91SStefano Zampini } 69939d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 69949d30be91SStefano Zampini 69959d30be91SStefano Zampini /* set preallocation */ 69969d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 69979d30be91SStefano Zampini if (!newisdense) { 69989d30be91SStefano Zampini PetscInt *new_local_nnz=0; 69999d30be91SStefano Zampini 70009d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 70019d30be91SStefano Zampini if (n_recvs) { 70029d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 70039d30be91SStefano Zampini } 70049d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 70059d30be91SStefano Zampini PetscInt j; 70069d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 70079d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 70089d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 70099d30be91SStefano Zampini } 70109d30be91SStefano Zampini } else { 70119d30be91SStefano Zampini /* TODO */ 70129d30be91SStefano Zampini } 70139d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 70149d30be91SStefano Zampini } 70159d30be91SStefano Zampini if (new_local_nnz) { 70169d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 70179d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 70189d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 70199d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 70209d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 70219d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 70229d30be91SStefano Zampini } else { 70239d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 70249d30be91SStefano Zampini } 70259d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 70269d30be91SStefano Zampini } else { 70279d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 70289d30be91SStefano Zampini } 7029e7931f94SStefano Zampini 7030e7931f94SStefano Zampini /* set values */ 7031e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 70329d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 7033e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7034e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 7035e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 70369d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 7037e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7038e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7039e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 704028143c3dSStefano Zampini } else { 704128143c3dSStefano Zampini /* TODO */ 7042e7931f94SStefano Zampini } 7043e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7044e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 7045e7931f94SStefano Zampini } 7046e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7047e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 704870cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 704970cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 70509d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 7051e7931f94SStefano Zampini 7052dfd14d43SStefano Zampini #if 0 705328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 7054e7931f94SStefano Zampini Vec lvec,rvec; 7055e7931f94SStefano Zampini PetscReal infty_error; 7056e7931f94SStefano Zampini 70572a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 7058e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 7059e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 7060e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 706170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 7062e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 7063e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 7064e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 7065e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 7066e7931f94SStefano Zampini } 706728143c3dSStefano Zampini #endif 7068e7931f94SStefano Zampini 706928143c3dSStefano Zampini /* assemble new additional is (if any) */ 707028143c3dSStefano Zampini if (nis) { 707128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 707228143c3dSStefano Zampini 707328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7074854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 707528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 707628143c3dSStefano Zampini psum = 0; 707728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 707828143c3dSStefano Zampini for (j=0;j<nis;j++) { 707928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 708028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 708128143c3dSStefano Zampini psum += plen; 708228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 708328143c3dSStefano Zampini } 708428143c3dSStefano Zampini } 7085854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 7086854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 708728143c3dSStefano Zampini for (i=1;i<nis;i++) { 708828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 708928143c3dSStefano Zampini } 709028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 709128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 709228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 709328143c3dSStefano Zampini for (j=0;j<nis;j++) { 709428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 709528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 709628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 709728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 709828143c3dSStefano Zampini } 709928143c3dSStefano Zampini } 710028143c3dSStefano Zampini for (i=0;i<nis;i++) { 710128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 710228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 710328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 710428143c3dSStefano Zampini } 710528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 710628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 710728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 710828143c3dSStefano Zampini } 7109e7931f94SStefano Zampini /* free workspace */ 711028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 7111e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7112e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 7113e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7114e7931f94SStefano Zampini if (isdense) { 7115e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 7116e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 7117e7931f94SStefano Zampini } else { 7118e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 7119e7931f94SStefano Zampini } 712028143c3dSStefano Zampini if (nis) { 712128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 712228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 712328143c3dSStefano Zampini } 71241ae86dd6SStefano Zampini 71251ae86dd6SStefano Zampini if (nvecs) { 71261ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 71271ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 71281ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 71291ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 71301ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 71311ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 71321ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 71331ae86dd6SStefano Zampini /* set values */ 71341ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 71351ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 71361ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 71371ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 71381ae86dd6SStefano Zampini PetscInt j; 71391ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 71401ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 71411ae86dd6SStefano Zampini } 71421ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 71431ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 71441ae86dd6SStefano Zampini } 71451ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 71461ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 71471ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 71481ae86dd6SStefano Zampini } 71491ae86dd6SStefano Zampini 71501ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 71511ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 7152e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 7153e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 71541ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 715528143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 7156e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 7157e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 71581ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 715928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 7160e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 7161e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 7162e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 7163e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 7164e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 716528143c3dSStefano Zampini if (nis) { 716628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 716728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 716828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 716928143c3dSStefano Zampini } 717028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 717128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 717228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 717328143c3dSStefano Zampini for (i=0;i<nis;i++) { 717428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 717528143c3dSStefano Zampini } 71761ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 71771ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 71781ae86dd6SStefano Zampini } 717953a05cb3SStefano Zampini *mat_n = NULL; 718028143c3dSStefano Zampini } 7181e7931f94SStefano Zampini PetscFunctionReturn(0); 7182e7931f94SStefano Zampini } 7183a57a6d2fSStefano Zampini 718412edc857SStefano Zampini /* temporary hack into ksp private data structure */ 7185af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 718612edc857SStefano Zampini 7187c8587f34SStefano Zampini #undef __FUNCT__ 7188c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 7189c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 7190c8587f34SStefano Zampini { 7191c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 7192c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 719320a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 71941ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 71951e0482f5SStefano Zampini Mat coarseG,t_coarse_mat_is; 71969881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 719720a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 71986e683305SStefano Zampini IS coarse_is,*isarray; 71996e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 720030368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 7201f9eb5b7dSStefano Zampini PC pc_temp; 7202c8587f34SStefano Zampini PCType coarse_pc_type; 7203c8587f34SStefano Zampini KSPType coarse_ksp_type; 7204f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 72054f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 72061e0482f5SStefano Zampini PetscInt ncoarse,nedcfield; 720768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 720822bc73bbSStefano Zampini PetscScalar *array; 720957de7509SStefano Zampini MatReuse coarse_mat_reuse; 721057de7509SStefano Zampini PetscBool restr, full_restr, have_void; 72119881197aSStefano Zampini PetscErrorCode ierr; 7212fdc09c96SStefano Zampini 7213c8587f34SStefano Zampini PetscFunctionBegin; 7214c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 721568457ee5SStefano 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 */ 7216fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 72175a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 7218fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 7219f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 7220f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 7221f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 7222fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 722351bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 722451bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 7225dc4bcba2SStefano Zampini PC pc; 7226dc4bcba2SStefano Zampini PetscBool isbddc; 7227dc4bcba2SStefano Zampini 7228dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 7229dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 7230dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 7231dc4bcba2SStefano Zampini if (isbddc) { 723263c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 723363c961adSStefano Zampini } else { 7234727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 723563c961adSStefano Zampini } 7236fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7237fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 7238fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7239f4ddd8eeSStefano Zampini } 7240fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 7241fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7242f4ddd8eeSStefano Zampini } 724370cf5478SStefano Zampini /* reset any subassembling information */ 724457de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 724570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 724657de7509SStefano Zampini } 72476e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 7248fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7249f4ddd8eeSStefano Zampini } 725057de7509SStefano Zampini /* assemble coarse matrix */ 725157de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 725257de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 725357de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 725457de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 725518a45a71SStefano Zampini } else { 725657de7509SStefano Zampini coarse_mat = NULL; 725757de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 72586e683305SStefano Zampini } 7259e7931f94SStefano Zampini 7260abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 7261abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 7262abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 7263abbbba34SStefano Zampini 7264abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 726522bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 726622bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 726722bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 726822bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 7269e176bc59SStefano 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); 72706e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 72716e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 72726e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7273abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 7274abbbba34SStefano Zampini 727557de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 727657de7509SStefano Zampini im_active = !!(pcis->n); 727757de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 727857de7509SStefano Zampini 727914f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 728057de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 728157de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 728257de7509SStefano Zampini coarse_mat_is = NULL; 728357de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 728457de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 72851ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 728657de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 728757de7509SStefano Zampini if (multilevel_requested) { 728857de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 728957de7509SStefano Zampini restr = PETSC_FALSE; 729057de7509SStefano Zampini full_restr = PETSC_FALSE; 729157de7509SStefano Zampini } else { 729257de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 729357de7509SStefano Zampini restr = PETSC_TRUE; 729457de7509SStefano Zampini full_restr = PETSC_TRUE; 729557de7509SStefano Zampini } 72964b2aedd3SStefano Zampini if (!pcbddc->coarse_size) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 729757de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 729857de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 7299a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 7300bb360cb4SStefano Zampini if (multilevel_requested) { 7301bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7302bb360cb4SStefano Zampini } else { 7303bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7304bb360cb4SStefano Zampini } 7305a198735bSStefano Zampini } else { 7306a198735bSStefano Zampini PetscMPIInt size,rank; 7307a198735bSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 7308a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 7309a198735bSStefano Zampini have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE; 7310a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 7311a198735bSStefano Zampini } 731257de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 731357de7509SStefano Zampini PetscInt psum; 731457de7509SStefano Zampini PetscMPIInt size; 731557de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 731657de7509SStefano Zampini else psum = 0; 731757de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 731857de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 731957de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 732057de7509SStefano Zampini } 732157de7509SStefano Zampini /* determine if we can go multilevel */ 732257de7509SStefano Zampini if (multilevel_requested) { 732357de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 732457de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 732557de7509SStefano Zampini } 732657de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 732757de7509SStefano Zampini 7328e4d548c7SStefano Zampini /* dump subassembling pattern */ 7329e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 7330e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 7331e4d548c7SStefano Zampini } 7332e4d548c7SStefano Zampini 73336e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 73341e0482f5SStefano Zampini nedcfield = -1; 73351e0482f5SStefano Zampini if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal)) { /* protects from unneded computations */ 73366e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 73376e683305SStefano Zampini const PetscInt *idxs; 73386e683305SStefano Zampini ISLocalToGlobalMapping tmap; 73396e683305SStefano Zampini 73406e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 73410be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 73426e683305SStefano Zampini /* allocate space for temporary storage */ 7343854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 7344854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 73456e683305SStefano Zampini /* allocate for IS array */ 73466e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 73471e0482f5SStefano Zampini if (pcbddc->nedclocal) { 73481e0482f5SStefano Zampini if (pcbddc->nedfield > -1) { 73491e0482f5SStefano Zampini nedcfield = pcbddc->nedfield; 73501e0482f5SStefano Zampini } else { 73511e0482f5SStefano Zampini nedcfield = 0; 73521e0482f5SStefano Zampini if (nisdofs) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen (%d)",nisdofs); 73531e0482f5SStefano Zampini nisdofs = 1; 73541e0482f5SStefano Zampini } 73551e0482f5SStefano Zampini } 73566e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 735727b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 735830368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 7359854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 73606e683305SStefano Zampini /* dofs splitting */ 73616e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 73626e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 73631e0482f5SStefano Zampini if (nedcfield != i) { 73646e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 73656e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 73666e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 73676e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 73681e0482f5SStefano Zampini } else { 73691e0482f5SStefano Zampini ierr = ISGetLocalSize(pcbddc->nedclocal,&tsize);CHKERRQ(ierr); 73701e0482f5SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 73711e0482f5SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 7372eee23b56SStefano Zampini if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping coarse nedelec field! %d != %d\n",tsize,nout); 73731e0482f5SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 73741e0482f5SStefano Zampini } 73756e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 737630368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 73776e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 73786e683305SStefano Zampini } 73796e683305SStefano Zampini /* neumann boundaries */ 73806e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 73816e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 73826e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 73836e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 73846e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 73856e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 73866e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 738730368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 73886e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 73896e683305SStefano Zampini } 73906e683305SStefano Zampini /* free memory */ 73916e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 73926e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 73936e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 73946e683305SStefano Zampini } else { 73956e683305SStefano Zampini nis = 0; 73966e683305SStefano Zampini nisdofs = 0; 73976e683305SStefano Zampini nisneu = 0; 739830368db7SStefano Zampini nisvert = 0; 73996e683305SStefano Zampini isarray = NULL; 74006e683305SStefano Zampini } 74016e683305SStefano Zampini /* destroy no longer needed map */ 74026e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 74036e683305SStefano Zampini 740457de7509SStefano Zampini /* subassemble */ 740557de7509SStefano Zampini if (multilevel_allowed) { 74061ae86dd6SStefano Zampini Vec vp[1]; 74071ae86dd6SStefano Zampini PetscInt nvecs = 0; 740857de7509SStefano Zampini PetscBool reuse,reuser; 74091ae86dd6SStefano Zampini 741057de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 741157de7509SStefano Zampini else reuse = PETSC_FALSE; 741257de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 74131ae86dd6SStefano Zampini vp[0] = NULL; 74141ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 74151ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 74161ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 74171ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 74181ae86dd6SStefano Zampini nvecs = 1; 74191ae86dd6SStefano Zampini 74201ae86dd6SStefano Zampini if (pcbddc->divudotp) { 7421a198735bSStefano Zampini Mat B,loc_divudotp; 74221ae86dd6SStefano Zampini Vec v,p; 74231ae86dd6SStefano Zampini IS dummy; 74241ae86dd6SStefano Zampini PetscInt np; 74251ae86dd6SStefano Zampini 7426a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 7427a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 74281ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 7429a198735bSStefano Zampini ierr = MatGetSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 74301ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 74311ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 74321ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 74331ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 74341ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 74351ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 74361ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 74371ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 74381ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 74391ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 74401ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 74411ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 744274e2c79eSStefano Zampini } 74431ae86dd6SStefano Zampini } 74441ae86dd6SStefano Zampini if (reuser) { 74451e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 744674e2c79eSStefano Zampini } else { 74471e0482f5SStefano 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); 74481ae86dd6SStefano Zampini } 74491ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 74501ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 74511ae86dd6SStefano Zampini PetscInt nl; 74521ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 74531ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 74541ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 74551ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 74561ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 74571ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 74581ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 74591ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 7460a198735bSStefano Zampini } else { 7461a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 74621ae86dd6SStefano Zampini } 74631ae86dd6SStefano Zampini } else { 74641e0482f5SStefano 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); 74656e683305SStefano Zampini } 746657de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 746757de7509SStefano Zampini PetscMPIInt size; 7468f913dca9SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size);CHKERRQ(ierr); 746957de7509SStefano Zampini if (!multilevel_allowed) { 747057de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 74716e683305SStefano Zampini } else { 747257de7509SStefano Zampini Mat A; 7473779c1cceSStefano Zampini 747457de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 747557de7509SStefano Zampini if (coarse_mat_is) { 747657de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 747757de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 747857de7509SStefano Zampini coarse_mat = coarse_mat_is; 747957de7509SStefano Zampini } 748057de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 748157de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 748257de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 7483779c1cceSStefano Zampini } 7484779c1cceSStefano Zampini } 748557de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 748657de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 74876e683305SStefano Zampini 74886e683305SStefano Zampini /* create local to global scatters for coarse problem */ 748968457ee5SStefano Zampini if (compute_vecs) { 74906e683305SStefano Zampini PetscInt lrows; 74916e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 749257de7509SStefano Zampini if (coarse_mat) { 749357de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 74946e683305SStefano Zampini } else { 74956e683305SStefano Zampini lrows = 0; 74966e683305SStefano Zampini } 74976e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 74986e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 74996e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 75006e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 75016e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 75026e683305SStefano Zampini } 75036e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 7504c8587f34SStefano Zampini 7505f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 7506f9eb5b7dSStefano Zampini if (multilevel_allowed) { 7507f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 7508f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 7509f9eb5b7dSStefano Zampini } else { 7510f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 7511f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 7512c8587f34SStefano Zampini } 7513c8587f34SStefano Zampini 75146e683305SStefano Zampini /* print some info if requested */ 75156e683305SStefano Zampini if (pcbddc->dbg_flag) { 75166e683305SStefano Zampini if (!multilevel_allowed) { 75176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 75186e683305SStefano Zampini if (multilevel_requested) { 75196e683305SStefano 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); 75206e683305SStefano Zampini } else if (pcbddc->max_levels) { 75216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 75226e683305SStefano Zampini } 75236e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 75246e683305SStefano Zampini } 75256e683305SStefano Zampini } 75266e683305SStefano Zampini 75271e0482f5SStefano Zampini /* communicate coarse discrete gradient */ 75281e0482f5SStefano Zampini coarseG = NULL; 75291e0482f5SStefano Zampini if (pcbddc->nedcG && multilevel_allowed) { 75301e0482f5SStefano Zampini MPI_Comm ccomm; 75311e0482f5SStefano Zampini if (coarse_mat) { 75321e0482f5SStefano Zampini ccomm = PetscObjectComm((PetscObject)coarse_mat); 75331e0482f5SStefano Zampini } else { 75341e0482f5SStefano Zampini ccomm = MPI_COMM_NULL; 75351e0482f5SStefano Zampini } 75361e0482f5SStefano Zampini ierr = MatMPIAIJRestrict(pcbddc->nedcG,ccomm,&coarseG);CHKERRQ(ierr); 75371e0482f5SStefano Zampini } 75381e0482f5SStefano Zampini 7539f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 754057de7509SStefano Zampini if (coarse_mat) { 75416a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 75426e683305SStefano Zampini if (pcbddc->dbg_flag) { 754357de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 75446e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 75456e683305SStefano Zampini } 7546f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 7547312be037SStefano Zampini char prefix[256],str_level[16]; 7548e604994aSStefano Zampini size_t len; 75491e0482f5SStefano Zampini 755057de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 7551422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 7552c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 7553f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 755457de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7555c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 75566e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 7557c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 75581e0482f5SStefano Zampini /* TODO is this logic correct? should check for coarse_mat type */ 7559c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7560e604994aSStefano Zampini /* prefix */ 7561e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 7562e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 7563e604994aSStefano Zampini if (!pcbddc->current_level) { 7564e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 7565e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 7566c8587f34SStefano Zampini } else { 7567e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 7568312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 7569312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 757034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 7571312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 7572e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 7573e604994aSStefano Zampini } 7574e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 75753e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 75763e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 75773e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 75783e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 7579f9eb5b7dSStefano Zampini /* allow user customization */ 7580f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 75813e3c6dadSStefano Zampini } 75823e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 758351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 75843e3c6dadSStefano Zampini if (nisdofs) { 75853e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 75863e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 75873e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 75883e3c6dadSStefano Zampini } 75893e3c6dadSStefano Zampini } 75903e3c6dadSStefano Zampini if (nisneu) { 75913e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 75923e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 7593312be037SStefano Zampini } 759430368db7SStefano Zampini if (nisvert) { 759530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 759630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 759730368db7SStefano Zampini } 75981e0482f5SStefano Zampini if (coarseG) { 75991e0482f5SStefano Zampini ierr = PCBDDCSetDiscreteGradient(pc_temp,coarseG,1,nedcfield,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 76001e0482f5SStefano Zampini } 7601f9eb5b7dSStefano Zampini 7602f9eb5b7dSStefano Zampini /* get some info after set from options */ 7603f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 7604f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 76054f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 76066e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 7607f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7608f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 7609f9eb5b7dSStefano Zampini } 761039f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 76114f3a063dSStefano Zampini if (isredundant) { 76124f3a063dSStefano Zampini KSP inner_ksp; 76134f3a063dSStefano Zampini PC inner_pc; 76144f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 76154f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 76164f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 76174f3a063dSStefano Zampini } 7618f9eb5b7dSStefano Zampini 761957de7509SStefano Zampini /* parameters which miss an API */ 762057de7509SStefano Zampini if (isbddc) { 7621720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 7622720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 762357de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 762427b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 762527b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 7626a198735bSStefano Zampini Mat coarsedivudotp_is; 7627a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 7628a198735bSStefano Zampini IS row,col; 7629a198735bSStefano Zampini const PetscInt *gidxs; 7630a198735bSStefano Zampini PetscInt n,st,M,N; 7631a198735bSStefano Zampini 7632a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 7633a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 7634a198735bSStefano Zampini st = st-n; 7635a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 7636a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 7637a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 7638a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7639a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 7640a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7641a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 7642a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 7643a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 7644a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 7645a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 7646a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 7647a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 7648a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 7649a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 7650a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 7651a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 7652a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 7653a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 7654a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 76558ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 7656a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 7657720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 765859e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 7659720d30f9SStefano Zampini } 7660d4d8cf7bSStefano Zampini } 76619881197aSStefano Zampini 76623301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 76635a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 76643301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 76653301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 76663301b35fSStefano Zampini } 76673301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 76683301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 76693301b35fSStefano Zampini } 76703301b35fSStefano Zampini if (pc->pmat->spd_set) { 76713301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 76723301b35fSStefano Zampini } 767327b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 767427b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 767527b6a85dSStefano Zampini } 76766e683305SStefano Zampini /* set operators */ 76775f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 76786e683305SStefano Zampini if (pcbddc->dbg_flag) { 76796e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 76806e683305SStefano Zampini } 76816e683305SStefano Zampini } 76821e0482f5SStefano Zampini ierr = MatDestroy(&coarseG);CHKERRQ(ierr); 76836e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 7684b1ecc7b1SStefano Zampini #if 0 7685b9b85e73SStefano Zampini { 7686b9b85e73SStefano Zampini PetscViewer viewer; 7687b9b85e73SStefano Zampini char filename[256]; 7688b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 7689b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 76906a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 7691b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 7692f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 7693b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 7694b9b85e73SStefano Zampini } 7695b9b85e73SStefano Zampini #endif 7696f9eb5b7dSStefano Zampini 769798a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 769898a51de6SStefano Zampini Vec crhs,csol; 769904708bb6SStefano Zampini 7700f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 7701f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 7702f347579bSStefano Zampini if (!csol) { 77032a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 7704f9eb5b7dSStefano Zampini } 7705f347579bSStefano Zampini if (!crhs) { 77062a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 7707f347579bSStefano Zampini } 7708b0f5fe93SStefano Zampini } 77091ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 7710b0f5fe93SStefano Zampini 7711b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 7712b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 7713b0f5fe93SStefano Zampini 7714b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 77154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 77164f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 77174f1b2e48SStefano Zampini } 7718b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 7719b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 7720b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7721b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7722b0f5fe93SStefano Zampini if (coarse_mat) { 7723b0f5fe93SStefano Zampini Vec nullv; 7724b0f5fe93SStefano Zampini PetscScalar *array,*array2; 7725b0f5fe93SStefano Zampini PetscInt nl; 7726b0f5fe93SStefano Zampini 7727b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 7728b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 7729b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7730b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 7731b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 7732b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 7733b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7734b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 7735b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 7736b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 7737b0f5fe93SStefano Zampini } 7738b0f5fe93SStefano Zampini } 7739b0f5fe93SStefano Zampini 7740b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 7741b0f5fe93SStefano Zampini PetscBool ispreonly; 7742b0f5fe93SStefano Zampini 7743b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7744b0f5fe93SStefano Zampini PetscBool isnull; 7745b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 7746bef83e63SStefano Zampini if (isnull) { 7747b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 7748b0f5fe93SStefano Zampini } 7749bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 7750b0f5fe93SStefano Zampini } 7751b0f5fe93SStefano Zampini /* setup coarse ksp */ 7752b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 7753cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 7754cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 77556e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 7756c8587f34SStefano Zampini KSP check_ksp; 77572b510759SStefano Zampini KSPType check_ksp_type; 7758c8587f34SStefano Zampini PC check_pc; 77596e683305SStefano Zampini Vec check_vec,coarse_vec; 77606a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 77612b510759SStefano Zampini PetscInt its; 77626e683305SStefano Zampini PetscBool compute_eigs; 77636e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 77646e683305SStefano Zampini PetscInt neigs; 77658e185a42SStefano Zampini const char *prefix; 7766c8587f34SStefano Zampini 77672b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 77686e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 7769422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 777023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7771f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 7772e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 7773e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 7774e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 77752b510759SStefano Zampini if (ispreonly) { 77762b510759SStefano Zampini check_ksp_type = KSPPREONLY; 77776e683305SStefano Zampini compute_eigs = PETSC_FALSE; 77782b510759SStefano Zampini } else { 7779cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 77806e683305SStefano Zampini compute_eigs = PETSC_TRUE; 7781c8587f34SStefano Zampini } 7782c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 77836e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 77846e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 77856e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 7786a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 7787a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 7788a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 7789a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 7790c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 7791c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 7792c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 7793c8587f34SStefano Zampini /* create random vec */ 77942701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 7795c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 77966e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 7797c8587f34SStefano Zampini /* solve coarse problem */ 77986e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 7799cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 78006e683305SStefano Zampini if (compute_eigs) { 7801854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 7802854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 78036e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 78041ae86dd6SStefano Zampini if (neigs) { 78056e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 78066e683305SStefano Zampini lambda_min = eigs_r[0]; 78076e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 78082701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 78092701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 7810cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 7811cbcc2c2aSStefano Zampini } 7812c8587f34SStefano Zampini } 7813c8587f34SStefano Zampini } 78141ae86dd6SStefano Zampini } 7815cbcc2c2aSStefano Zampini 7816c8587f34SStefano Zampini /* check coarse problem residual error */ 78176e683305SStefano Zampini if (pcbddc->dbg_flag) { 78186e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 78196e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 78206e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 7821c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 78226e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 78236e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 7824779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 78256e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 78266e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 78276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 78286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 7829b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7830b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 7831b0f5fe93SStefano Zampini } 78326e683305SStefano Zampini if (compute_eigs) { 78336e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 7834b03ebc13SStefano Zampini KSPConvergedReason reason; 7835deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 7836c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 7837b03ebc13SStefano Zampini ierr = KSPGetConvergedReason(check_ksp,&reason);CHKERRQ(ierr); 78386e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 7839b03ebc13SStefano 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); 78406e683305SStefano Zampini for (i=0;i<neigs;i++) { 78416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 7842c8587f34SStefano Zampini } 78436e683305SStefano Zampini } 78446e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 78456e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 78466e683305SStefano Zampini } 7847e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 78482701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 7849c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 78506e683305SStefano Zampini if (compute_eigs) { 78516e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 78526e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 7853c8587f34SStefano Zampini } 78546e683305SStefano Zampini } 78556e683305SStefano Zampini } 7856bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 7857cbcc2c2aSStefano Zampini /* print additional info */ 7858cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 78596e683305SStefano Zampini /* waits until all processes reaches this point */ 78606e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 7861cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 7862cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7863cbcc2c2aSStefano Zampini } 7864cbcc2c2aSStefano Zampini 78652b510759SStefano Zampini /* free memory */ 7866fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 7867c8587f34SStefano Zampini PetscFunctionReturn(0); 7868c8587f34SStefano Zampini } 7869674ae819SStefano Zampini 7870f34684f1SStefano Zampini #undef __FUNCT__ 7871f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 7872f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 7873f34684f1SStefano Zampini { 7874f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7875f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 7876f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 7877dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 7878dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 787973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 7880dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 7881f34684f1SStefano Zampini PetscErrorCode ierr; 7882f34684f1SStefano Zampini 7883f34684f1SStefano Zampini PetscFunctionBegin; 7884f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 78856c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 7886dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 78873bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 7888dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7889dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 78906583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 7891dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 7892dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 7893dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 78946c4ed002SBarry 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); 7895dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 7896dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7897dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 7898dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7899dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7900f34684f1SStefano Zampini 7901f34684f1SStefano Zampini /* check numbering */ 7902f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 7903019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 7904dc456d91SStefano Zampini PetscInt i; 7905b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 7906f34684f1SStefano Zampini 7907f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7908f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 7909f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 79101575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7911019a44ceSStefano Zampini /* counter */ 7912019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7913019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 7914019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7915019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7916019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7917019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7918f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 7919f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 7920727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 7921f34684f1SStefano Zampini } 7922f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7923f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7924f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7925e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7926e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7927e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7928e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7929f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7930019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 7931f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 7932019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 79332c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 793475c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 7935b9b85e73SStefano Zampini set_error = PETSC_TRUE; 79362c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 79372c66d082SStefano 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); 7938f34684f1SStefano Zampini } 7939f34684f1SStefano Zampini } 7940019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 7941b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 7942f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7943f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 7944f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 7945f34684f1SStefano Zampini } 7946f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 7947f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7948e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7949e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7950f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 7951f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 7952b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 7953ca8b9ea9SStefano Zampini PetscInt *gidxs; 7954ca8b9ea9SStefano Zampini 7955ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 79563bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 7957f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 7958f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7959f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 7960f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 79614bc2dc4bSStefano 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); 7962f34684f1SStefano Zampini } 7963f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7964ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 7965f34684f1SStefano Zampini } 7966f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 79671575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7968302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 7969f34684f1SStefano Zampini } 79708bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 7971f34684f1SStefano Zampini /* get back data */ 7972f34684f1SStefano Zampini *coarse_size_n = coarse_size; 7973f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 7974674ae819SStefano Zampini PetscFunctionReturn(0); 7975674ae819SStefano Zampini } 7976674ae819SStefano Zampini 7977e456f2a8SStefano Zampini #undef __FUNCT__ 7978e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 7979a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 7980e456f2a8SStefano Zampini { 7981e456f2a8SStefano Zampini IS localis_t; 7982a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 7983e456f2a8SStefano Zampini PetscScalar *vals; 7984e456f2a8SStefano Zampini PetscErrorCode ierr; 7985e456f2a8SStefano Zampini 7986e456f2a8SStefano Zampini PetscFunctionBegin; 7987a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 7988e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 7989854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 7990e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 7991e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 7992a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 7993a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 79941035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 7995a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 79961035eff8SStefano Zampini } 7997a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 7998e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 7999e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 8000a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 8001a7dc3881SStefano Zampini /* now compute set in local ordering */ 8002a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8003a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8004a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8005a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 8006a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8007ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8008e456f2a8SStefano Zampini lsize++; 8009e456f2a8SStefano Zampini } 8010e456f2a8SStefano Zampini } 8011854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 8012a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8013ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8014e456f2a8SStefano Zampini idxs[lsize++] = i; 8015e456f2a8SStefano Zampini } 8016e456f2a8SStefano Zampini } 8017a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8018a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 8019e456f2a8SStefano Zampini *localis = localis_t; 8020e456f2a8SStefano Zampini PetscFunctionReturn(0); 8021e456f2a8SStefano Zampini } 8022906d46d4SStefano Zampini 8023b96c3477SStefano Zampini #undef __FUNCT__ 8024b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 802508122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 8026b96c3477SStefano Zampini { 8027a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8028b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8029b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 8030a64f4aa4SStefano Zampini Mat S_j; 8031b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 8032b96c3477SStefano Zampini PetscBool free_used_adj; 8033b96c3477SStefano Zampini PetscErrorCode ierr; 8034b96c3477SStefano Zampini 8035b96c3477SStefano Zampini PetscFunctionBegin; 8036b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 8037b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 803808122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 8039b96c3477SStefano Zampini used_xadj = NULL; 8040b96c3477SStefano Zampini used_adjncy = NULL; 8041b96c3477SStefano Zampini } else { 804208122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 804308122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 804408122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 804508122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 8046b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 8047b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 8048b96c3477SStefano Zampini } else { 80492fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 8050b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 8051b96c3477SStefano Zampini PetscInt nvtxs; 8052b96c3477SStefano Zampini 80532fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 80542fffb893SStefano Zampini if (flg_row) { 8055b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 8056b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 8057b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 8058b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 80592fffb893SStefano Zampini } else { 80602fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 80612fffb893SStefano Zampini used_xadj = NULL; 80622fffb893SStefano Zampini used_adjncy = NULL; 80632fffb893SStefano Zampini } 80642fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 8065b96c3477SStefano Zampini } 8066b96c3477SStefano Zampini } 8067d5574798SStefano Zampini 8068d5574798SStefano Zampini /* setup sub_schurs data */ 8069a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8070df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 8071df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 8072a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 807391af6908SStefano 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); 8074a64f4aa4SStefano Zampini } else { 80756816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 80764d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 8077a3df083aSStefano Zampini PetscInt benign_n; 807872b8c272SStefano Zampini Mat change = NULL; 80799d54b7f4SStefano Zampini Vec scaling = NULL; 808072b8c272SStefano Zampini IS change_primal = NULL; 8081a3df083aSStefano Zampini 80825feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 80835feab87aSStefano Zampini PetscInt n_vertices; 80845feab87aSStefano Zampini 80855feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 80862034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 80875feab87aSStefano Zampini } 808804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 808904708bb6SStefano Zampini if (!isseqaij) { 809004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 809104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 809204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 809304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 809404708bb6SStefano Zampini } else { 8095511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 809604708bb6SStefano Zampini } 809704708bb6SStefano Zampini } 8098a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 8099a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 8100ca92afb2SStefano Zampini } else { 8101a3df083aSStefano Zampini benign_n = 0; 8102ca92afb2SStefano Zampini } 8103b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 8104b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 8105b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 810672b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 810722db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 8108b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 810922db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 8110b7ab4a40SStefano Zampini } 8111b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 8112b7ab4a40SStefano 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 */ 8113b7ab4a40SStefano Zampini if (need_change) { 811488c03ad3SStefano Zampini PC_IS *pcisf; 811588c03ad3SStefano Zampini PC_BDDC *pcbddcf; 811688c03ad3SStefano Zampini PC pcf; 811788c03ad3SStefano Zampini 8118e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 811988c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 812088c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 812188c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 812288c03ad3SStefano Zampini /* hacks */ 812388c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 812472b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 812572b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 812672b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 812772b8c272SStefano Zampini pcisf->n = pcis->n; 812872b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 812988c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 813088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 813188c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 813288c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 813388c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 813488c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 813572b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 813688c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 813788c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 813872b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 813972b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 814072b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 814172b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 814272b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 814388c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 814472b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 814588c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 814688c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 814788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 814888c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 814988c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 815088c03ad3SStefano Zampini pcf->ops->destroy = NULL; 815188c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 815288c03ad3SStefano Zampini } 81539d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 815491af6908SStefano 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); 815572b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 815672b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 8157ca92afb2SStefano Zampini } 8158d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8159b96c3477SStefano Zampini 8160b96c3477SStefano Zampini /* free adjacency */ 8161b96c3477SStefano Zampini if (free_used_adj) { 8162b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 8163b96c3477SStefano Zampini } 8164b96c3477SStefano Zampini PetscFunctionReturn(0); 8165b96c3477SStefano Zampini } 8166b96c3477SStefano Zampini 8167b96c3477SStefano Zampini #undef __FUNCT__ 8168b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 816908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 8170b96c3477SStefano Zampini { 8171b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8172b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8173b96c3477SStefano Zampini PCBDDCGraph graph; 8174b96c3477SStefano Zampini PetscErrorCode ierr; 8175b96c3477SStefano Zampini 8176b96c3477SStefano Zampini PetscFunctionBegin; 8177b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 817808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 81793301b35fSStefano Zampini IS verticesIS,verticescomm; 81803301b35fSStefano Zampini PetscInt vsize,*idxs; 8181b96c3477SStefano Zampini 8182b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 81833301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 81843301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 81853301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 81863301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 8187c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 8188b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 8189be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 8190441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 81913301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 8192b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 8193b96c3477SStefano Zampini } else { 8194b96c3477SStefano Zampini graph = pcbddc->mat_graph; 8195b96c3477SStefano Zampini } 8196e4d548c7SStefano Zampini /* print some info */ 81975c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 8198e4d548c7SStefano Zampini IS vertices; 8199e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 8200c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 8201e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8202e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 8203e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8204e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 8205e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 8206e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 8207e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 8208e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8209e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8210c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8211e4d548c7SStefano Zampini } 8212b96c3477SStefano Zampini 8213b96c3477SStefano Zampini /* sub_schurs init */ 8214b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 8215b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 8216b334f244SStefano Zampini } 82178b6046baSStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 8218a64f4aa4SStefano Zampini 8219b96c3477SStefano Zampini /* free graph struct */ 822008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 8221b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 8222b96c3477SStefano Zampini } 8223b96c3477SStefano Zampini PetscFunctionReturn(0); 8224b96c3477SStefano Zampini } 8225fa34dd3eSStefano Zampini 8226fa34dd3eSStefano Zampini #undef __FUNCT__ 8227fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 8228fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 8229fa34dd3eSStefano Zampini { 8230fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8231fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8232fa34dd3eSStefano Zampini PetscErrorCode ierr; 8233fa34dd3eSStefano Zampini 8234fa34dd3eSStefano Zampini PetscFunctionBegin; 8235fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 8236fa34dd3eSStefano Zampini IS zerodiag = NULL; 82374f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 8238fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 82394f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 824075c01103SStefano Zampini PetscReal norm; 8241fa34dd3eSStefano Zampini PetscInt i; 8242fa34dd3eSStefano Zampini 8243fa34dd3eSStefano Zampini /* B0 and B0_B */ 8244fa34dd3eSStefano Zampini if (zerodiag) { 8245fa34dd3eSStefano Zampini IS dummy; 8246fa34dd3eSStefano Zampini 82474f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 82484f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 8249fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 8250fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 8251fa34dd3eSStefano Zampini } 8252fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 8253fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 8254fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 8255fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8256fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8257fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8258fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8259fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 8260fa34dd3eSStefano Zampini /* S_j */ 8261fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8262fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 8263fa34dd3eSStefano Zampini 8264fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 8265fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 8266fa34dd3eSStefano Zampini /* continuous in primal space */ 8267fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 8268fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8269fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8270fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 82714f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 82724f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 8273fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8274fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8275fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8276fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8277fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8278fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8279fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 8280fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 8281fa34dd3eSStefano Zampini 8282fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 8283fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 8284fa34dd3eSStefano Zampini /* local with Schur */ 8285fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 8286fa34dd3eSStefano Zampini if (zerodiag) { 8287fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 82884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 8289fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8290fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 8291fa34dd3eSStefano Zampini } 8292fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 8293fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8294fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8295fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8296fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8297fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 8298fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8299fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8300fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 8301fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8302fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8303fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8304fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8305fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8306fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 8307fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 8308fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8309fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8310fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8311fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8312fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8313fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8314fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 8315fa34dd3eSStefano Zampini if (zerodiag) { 8316fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 8317fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 83184f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 8319fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8320fa34dd3eSStefano Zampini } 8321fa34dd3eSStefano Zampini /* BDDC */ 8322fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 8323fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 8324fa34dd3eSStefano Zampini 8325fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 8326fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 8327fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 8328fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 83294f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 83304f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 8331fa34dd3eSStefano Zampini } 83324f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 8333fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 8334fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 8335fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 8336fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8337fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 8338fa34dd3eSStefano Zampini } 8339fa34dd3eSStefano Zampini PetscFunctionReturn(0); 8340fa34dd3eSStefano Zampini } 83411e0482f5SStefano Zampini 83421e0482f5SStefano Zampini #include <../src/mat/impls/aij/mpi/mpiaij.h> 83431e0482f5SStefano Zampini #undef __FUNCT__ 83441e0482f5SStefano Zampini #define __FUNCT__ "MatMPIAIJRestrict" 83451e0482f5SStefano Zampini PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B) 83461e0482f5SStefano Zampini { 83471e0482f5SStefano Zampini Mat At; 83481e0482f5SStefano Zampini IS rows; 83491e0482f5SStefano Zampini PetscInt rst,ren; 83501e0482f5SStefano Zampini PetscErrorCode ierr; 83511e0482f5SStefano Zampini PetscLayout rmap; 83521e0482f5SStefano Zampini 83531e0482f5SStefano Zampini PetscFunctionBegin; 83541e0482f5SStefano Zampini rst = ren = 0; 83551e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 83561e0482f5SStefano Zampini ierr = PetscLayoutCreate(ccomm,&rmap);CHKERRQ(ierr); 83571e0482f5SStefano Zampini ierr = PetscLayoutSetSize(rmap,A->rmap->N);CHKERRQ(ierr); 83581e0482f5SStefano Zampini ierr = PetscLayoutSetBlockSize(rmap,1);CHKERRQ(ierr); 83591e0482f5SStefano Zampini ierr = PetscLayoutSetUp(rmap);CHKERRQ(ierr); 83601e0482f5SStefano Zampini ierr = PetscLayoutGetRange(rmap,&rst,&ren);CHKERRQ(ierr); 83611e0482f5SStefano Zampini } 8362e07686f2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)A),ren-rst,rst,1,&rows);CHKERRQ(ierr); 83631e0482f5SStefano Zampini ierr = MatGetSubMatrix(A,rows,NULL,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 83641e0482f5SStefano Zampini ierr = ISDestroy(&rows);CHKERRQ(ierr); 83651e0482f5SStefano Zampini 83661e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 83671e0482f5SStefano Zampini Mat_MPIAIJ *a,*b; 83681e0482f5SStefano Zampini IS from,to; 83691e0482f5SStefano Zampini Vec gvec; 83701e0482f5SStefano Zampini PetscInt lsize; 83711e0482f5SStefano Zampini 83721e0482f5SStefano Zampini ierr = MatCreate(ccomm,B);CHKERRQ(ierr); 83731e0482f5SStefano Zampini ierr = MatSetSizes(*B,ren-rst,PETSC_DECIDE,PETSC_DECIDE,At->cmap->N);CHKERRQ(ierr); 83741e0482f5SStefano Zampini ierr = MatSetType(*B,MATAIJ);CHKERRQ(ierr); 83751e0482f5SStefano Zampini ierr = PetscLayoutDestroy(&((*B)->rmap));CHKERRQ(ierr); 83761e0482f5SStefano Zampini ierr = PetscLayoutSetUp((*B)->cmap);CHKERRQ(ierr); 83771e0482f5SStefano Zampini a = (Mat_MPIAIJ*)At->data; 83781e0482f5SStefano Zampini b = (Mat_MPIAIJ*)(*B)->data; 83791e0482f5SStefano Zampini ierr = MPI_Comm_size(ccomm,&b->size);CHKERRQ(ierr); 83801e0482f5SStefano Zampini ierr = MPI_Comm_rank(ccomm,&b->rank);CHKERRQ(ierr); 83811e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->A);CHKERRQ(ierr); 83821e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->B);CHKERRQ(ierr); 83831e0482f5SStefano Zampini b->A = a->A; 83841e0482f5SStefano Zampini b->B = a->B; 83851e0482f5SStefano Zampini 83861e0482f5SStefano Zampini b->donotstash = a->donotstash; 83871e0482f5SStefano Zampini b->roworiented = a->roworiented; 83881e0482f5SStefano Zampini b->rowindices = 0; 83891e0482f5SStefano Zampini b->rowvalues = 0; 83901e0482f5SStefano Zampini b->getrowactive = PETSC_FALSE; 83911e0482f5SStefano Zampini 83921e0482f5SStefano Zampini (*B)->rmap = rmap; 83931e0482f5SStefano Zampini (*B)->factortype = A->factortype; 83941e0482f5SStefano Zampini (*B)->assembled = PETSC_TRUE; 83951e0482f5SStefano Zampini (*B)->insertmode = NOT_SET_VALUES; 83961e0482f5SStefano Zampini (*B)->preallocated = PETSC_TRUE; 83971e0482f5SStefano Zampini 83981e0482f5SStefano Zampini if (a->colmap) { 83991e0482f5SStefano Zampini #if defined(PETSC_USE_CTABLE) 84001e0482f5SStefano Zampini ierr = PetscTableCreateCopy(a->colmap,&b->colmap);CHKERRQ(ierr); 84011e0482f5SStefano Zampini #else 84021e0482f5SStefano Zampini ierr = PetscMalloc1(At->cmap->N,&b->colmap);CHKERRQ(ierr); 84031e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)*B,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 84041e0482f5SStefano Zampini ierr = PetscMemcpy(b->colmap,a->colmap,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 84051e0482f5SStefano Zampini #endif 84061e0482f5SStefano Zampini } else b->colmap = 0; 84071e0482f5SStefano Zampini if (a->garray) { 84081e0482f5SStefano Zampini PetscInt len; 84091e0482f5SStefano Zampini len = a->B->cmap->n; 84101e0482f5SStefano Zampini ierr = PetscMalloc1(len+1,&b->garray);CHKERRQ(ierr); 84111e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)(*B),len*sizeof(PetscInt));CHKERRQ(ierr); 84121e0482f5SStefano Zampini if (len) { ierr = PetscMemcpy(b->garray,a->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 84131e0482f5SStefano Zampini } else b->garray = 0; 84141e0482f5SStefano Zampini 84151e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->lvec);CHKERRQ(ierr); 84161e0482f5SStefano Zampini b->lvec = a->lvec; 84171e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->lvec);CHKERRQ(ierr); 84181e0482f5SStefano Zampini 84191e0482f5SStefano Zampini /* cannot use VecScatterCopy */ 84201e0482f5SStefano Zampini ierr = VecGetLocalSize(b->lvec,&lsize);CHKERRQ(ierr); 84211e0482f5SStefano Zampini ierr = ISCreateGeneral(ccomm,lsize,b->garray,PETSC_USE_POINTER,&from);CHKERRQ(ierr); 84221e0482f5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,lsize,0,1,&to);CHKERRQ(ierr); 84231e0482f5SStefano Zampini ierr = MatCreateVecs(*B,&gvec,NULL);CHKERRQ(ierr); 84241e0482f5SStefano Zampini ierr = VecScatterCreate(gvec,from,b->lvec,to,&b->Mvctx);CHKERRQ(ierr); 84251e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->Mvctx);CHKERRQ(ierr); 84261e0482f5SStefano Zampini ierr = ISDestroy(&from);CHKERRQ(ierr); 84271e0482f5SStefano Zampini ierr = ISDestroy(&to);CHKERRQ(ierr); 84281e0482f5SStefano Zampini ierr = VecDestroy(&gvec);CHKERRQ(ierr); 84291e0482f5SStefano Zampini } 84301e0482f5SStefano Zampini ierr = MatDestroy(&At);CHKERRQ(ierr); 84311e0482f5SStefano Zampini PetscFunctionReturn(0); 84321e0482f5SStefano Zampini } 8433