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) */ 11f498cd09SStefano Zampini PetscErrorCode MatDenseOrthogonalRangeOrComplement(Mat A, PetscBool range, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B) 12a13144ffSStefano Zampini { 13a13144ffSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 14a13144ffSStefano Zampini PetscScalar *uwork,*data,*U, ds = 0.; 15a13144ffSStefano Zampini PetscReal *sing; 16a13144ffSStefano Zampini PetscBLASInt bM,bN,lwork,lierr,di = 1; 17a13144ffSStefano Zampini PetscInt ulw,i,nr,nc,n; 18a13144ffSStefano Zampini PetscErrorCode ierr; 19a13144ffSStefano Zampini 20a13144ffSStefano Zampini PetscFunctionBegin; 21a13144ffSStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 22a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"LAPACK _GESVD not available"); 23614dbb09SStefano Zampini #else 24a13144ffSStefano Zampini ierr = MatGetSize(A,&nr,&nc);CHKERRQ(ierr); 25a13144ffSStefano Zampini if (!nr || !nc) PetscFunctionReturn(0); 26a13144ffSStefano Zampini 27a13144ffSStefano Zampini /* workspace */ 28a13144ffSStefano Zampini if (!work) { 29a13144ffSStefano Zampini ulw = PetscMax(PetscMax(1,5*PetscMin(nr,nc)),3*PetscMin(nr,nc)+PetscMax(nr,nc)); 30f913dca9SStefano Zampini ierr = PetscMalloc1(ulw,&uwork);CHKERRQ(ierr); 31a13144ffSStefano Zampini } else { 32a13144ffSStefano Zampini ulw = lw; 33a13144ffSStefano Zampini uwork = work; 34a13144ffSStefano Zampini } 35a13144ffSStefano Zampini n = PetscMin(nr,nc); 36a13144ffSStefano Zampini if (!rwork) { 37a13144ffSStefano Zampini ierr = PetscMalloc1(n,&sing);CHKERRQ(ierr); 38a13144ffSStefano Zampini } else { 39a13144ffSStefano Zampini sing = rwork; 40a13144ffSStefano Zampini } 41a13144ffSStefano Zampini 42a13144ffSStefano Zampini /* SVD */ 43a13144ffSStefano Zampini ierr = PetscMalloc1(nr*nr,&U);CHKERRQ(ierr); 44a13144ffSStefano Zampini ierr = PetscBLASIntCast(nr,&bM);CHKERRQ(ierr); 45a13144ffSStefano Zampini ierr = PetscBLASIntCast(nc,&bN);CHKERRQ(ierr); 46a13144ffSStefano Zampini ierr = PetscBLASIntCast(ulw,&lwork);CHKERRQ(ierr); 47a13144ffSStefano Zampini ierr = MatDenseGetArray(A,&data);CHKERRQ(ierr); 48a13144ffSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 49a13144ffSStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,&lierr)); 50a13144ffSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 51a13144ffSStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 52a13144ffSStefano Zampini ierr = MatDenseRestoreArray(A,&data);CHKERRQ(ierr); 53a13144ffSStefano Zampini for (i=0;i<n;i++) if (sing[i] < PETSC_SMALL) break; 54a13144ffSStefano Zampini if (!rwork) { 55a13144ffSStefano Zampini ierr = PetscFree(sing);CHKERRQ(ierr); 56a13144ffSStefano Zampini } 57a13144ffSStefano Zampini if (!work) { 58a13144ffSStefano Zampini ierr = PetscFree(uwork);CHKERRQ(ierr); 59a13144ffSStefano Zampini } 60a13144ffSStefano Zampini /* create B */ 61f498cd09SStefano Zampini if (!range) { 62a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,nr-i,NULL,B);CHKERRQ(ierr); 63a13144ffSStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 64a13144ffSStefano Zampini ierr = PetscMemcpy(data,U+nr*i,(nr-i)*nr*sizeof(PetscScalar));CHKERRQ(ierr); 65f498cd09SStefano Zampini } else { 66f498cd09SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nr,i,NULL,B);CHKERRQ(ierr); 67f498cd09SStefano Zampini ierr = MatDenseGetArray(*B,&data);CHKERRQ(ierr); 68f498cd09SStefano Zampini ierr = PetscMemcpy(data,U,i*nr*sizeof(PetscScalar));CHKERRQ(ierr); 69f498cd09SStefano Zampini } 70a13144ffSStefano Zampini ierr = MatDenseRestoreArray(*B,&data);CHKERRQ(ierr); 71a13144ffSStefano Zampini ierr = PetscFree(U);CHKERRQ(ierr); 72614dbb09SStefano Zampini #endif 73614dbb09SStefano Zampini #else /* PETSC_USE_COMPLEX */ 74a13144ffSStefano Zampini PetscFunctionBegin; 75a13144ffSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for complexes"); 76a13144ffSStefano Zampini #endif 77a13144ffSStefano Zampini PetscFunctionReturn(0); 78a13144ffSStefano Zampini } 79a13144ffSStefano Zampini 801e0482f5SStefano Zampini /* TODO REMOVE */ 811e0482f5SStefano Zampini #if defined(PRINT_GDET) 821e0482f5SStefano Zampini static int inc = 0; 831e0482f5SStefano Zampini static int lev = 0; 841e0482f5SStefano Zampini #endif 851e0482f5SStefano Zampini 861e0482f5SStefano Zampini PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, IS corners, Mat* Gins, Mat* GKins, PetscScalar cvals[2], PetscScalar *work, PetscReal *rwork) 87a13144ffSStefano Zampini { 88a13144ffSStefano Zampini PetscErrorCode ierr; 89a13144ffSStefano Zampini Mat GE,GEd; 90a13144ffSStefano Zampini PetscInt rsize,csize,esize; 91a13144ffSStefano Zampini PetscScalar *ptr; 92a13144ffSStefano Zampini 93a13144ffSStefano Zampini PetscFunctionBegin; 94a13144ffSStefano Zampini ierr = ISGetSize(edge,&esize);CHKERRQ(ierr); 95c3c0e390SStefano Zampini if (!esize) PetscFunctionReturn(0); 96a13144ffSStefano Zampini ierr = ISGetSize(extrow,&rsize);CHKERRQ(ierr); 97a13144ffSStefano Zampini ierr = ISGetSize(extcol,&csize);CHKERRQ(ierr); 98a13144ffSStefano Zampini 99a13144ffSStefano Zampini /* gradients */ 100a13144ffSStefano Zampini ptr = work + 5*esize; 1017dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,extrow,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 102a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,rsize,csize,ptr,Gins);CHKERRQ(ierr); 103a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,Gins);CHKERRQ(ierr); 104a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 105a13144ffSStefano Zampini 106a13144ffSStefano Zampini /* constants */ 107a13144ffSStefano Zampini ptr += rsize*csize; 108a13144ffSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,esize,csize,ptr,&GEd);CHKERRQ(ierr); 1097dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,edge,extcol,MAT_INITIAL_MATRIX,&GE);CHKERRQ(ierr); 110a13144ffSStefano Zampini ierr = MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,&GEd);CHKERRQ(ierr); 111a13144ffSStefano Zampini ierr = MatDestroy(&GE);CHKERRQ(ierr); 112f498cd09SStefano Zampini ierr = MatDenseOrthogonalRangeOrComplement(GEd,PETSC_FALSE,5*esize,work,rwork,GKins);CHKERRQ(ierr); 113a13144ffSStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1141e0482f5SStefano Zampini 1151e0482f5SStefano Zampini if (corners) { 1161e0482f5SStefano Zampini Mat GEc; 1171e0482f5SStefano Zampini PetscScalar *vals,v; 1181e0482f5SStefano Zampini 1197dae84e0SHong Zhang ierr = MatCreateSubMatrix(lG,edge,corners,MAT_INITIAL_MATRIX,&GEc);CHKERRQ(ierr); 1201e0482f5SStefano Zampini ierr = MatTransposeMatMult(GEc,*GKins,MAT_INITIAL_MATRIX,1.0,&GEd);CHKERRQ(ierr); 1211e0482f5SStefano Zampini ierr = MatDenseGetArray(GEd,&vals);CHKERRQ(ierr); 122637e8532SStefano Zampini /* v = PetscAbsScalar(vals[0]) */; 123637e8532SStefano Zampini v = 1.; 1241e0482f5SStefano Zampini cvals[0] = vals[0]/v; 1251e0482f5SStefano Zampini cvals[1] = vals[1]/v; 1261e0482f5SStefano Zampini ierr = MatDenseRestoreArray(GEd,&vals);CHKERRQ(ierr); 1271e0482f5SStefano Zampini ierr = MatScale(*GKins,1./v);CHKERRQ(ierr); 1281e0482f5SStefano Zampini #if defined(PRINT_GDET) 1291e0482f5SStefano Zampini { 1301e0482f5SStefano Zampini PetscViewer viewer; 1311e0482f5SStefano Zampini char filename[256]; 1321e0482f5SStefano Zampini sprintf(filename,"Gdet_l%d_r%d_cc%d.m",lev,PetscGlobalRank,inc++); 1331e0482f5SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1341e0482f5SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1351e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEc,"GEc");CHKERRQ(ierr); 1361e0482f5SStefano Zampini ierr = MatView(GEc,viewer);CHKERRQ(ierr); 1371e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)(*GKins),"GK");CHKERRQ(ierr); 1381e0482f5SStefano Zampini ierr = MatView(*GKins,viewer);CHKERRQ(ierr); 1391e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)GEd,"Gproj");CHKERRQ(ierr); 1401e0482f5SStefano Zampini ierr = MatView(GEd,viewer);CHKERRQ(ierr); 1411e0482f5SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1421e0482f5SStefano Zampini } 1431e0482f5SStefano Zampini #endif 1441e0482f5SStefano Zampini ierr = MatDestroy(&GEd);CHKERRQ(ierr); 1451e0482f5SStefano Zampini ierr = MatDestroy(&GEc);CHKERRQ(ierr); 1461e0482f5SStefano Zampini } 1471e0482f5SStefano Zampini 148a13144ffSStefano Zampini PetscFunctionReturn(0); 149a13144ffSStefano Zampini } 150a13144ffSStefano Zampini 151a13144ffSStefano Zampini PetscErrorCode PCBDDCNedelecSupport(PC pc) 152a13144ffSStefano Zampini { 153a13144ffSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 154a13144ffSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1550569b399SStefano Zampini Mat G,T,conn,lG,lGt,lGis,lGall,lGe,lGinit; 156eee23b56SStefano Zampini Vec tvec; 157a13144ffSStefano Zampini PetscSF sfv; 1581e0482f5SStefano Zampini ISLocalToGlobalMapping el2g,vl2g,fl2g,al2g; 159a13144ffSStefano Zampini MPI_Comm comm; 160c2151214SStefano Zampini IS lned,primals,allprimals,nedfieldlocal; 161c2151214SStefano Zampini IS *eedges,*extrows,*extcols,*alleedges; 1627d871cd7SStefano Zampini PetscBT btv,bte,btvc,btb,btbd,btvcand,btvi,btee,bter; 163a13144ffSStefano Zampini PetscScalar *vals,*work; 164a13144ffSStefano Zampini PetscReal *rwork; 165a13144ffSStefano Zampini const PetscInt *idxs,*ii,*jj,*iit,*jjt; 1661e0482f5SStefano Zampini PetscInt ne,nv,Lv,order,n,field; 167a13144ffSStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 168eee23b56SStefano Zampini PetscInt i,j,extmem,cum,maxsize,nee; 169b03ebc13SStefano Zampini PetscInt *extrow,*extrowcum,*marks,*vmarks,*gidxs; 170a13144ffSStefano Zampini PetscInt *sfvleaves,*sfvroots; 171b03ebc13SStefano Zampini PetscInt *corners,*cedges; 172637e8532SStefano Zampini PetscInt *ecount,**eneighs,*vcount,**vneighs; 173b03ebc13SStefano Zampini #if defined(PETSC_USE_DEBUG) 174b03ebc13SStefano Zampini PetscInt *emarks; 175b03ebc13SStefano Zampini #endif 176213b8bfaSStefano Zampini PetscBool print,eerr,done,lrc[2],conforming,global,singular,setprimal; 177a13144ffSStefano Zampini PetscErrorCode ierr; 178a13144ffSStefano Zampini 179a13144ffSStefano Zampini PetscFunctionBegin; 180213b8bfaSStefano Zampini /* If the discrete gradient is defined for a subset of dofs and global is true, 181213b8bfaSStefano Zampini it assumes G is given in global ordering for all the dofs. 182213b8bfaSStefano Zampini Otherwise, the ordering is global for the Nedelec field */ 183213b8bfaSStefano Zampini order = pcbddc->nedorder; 184213b8bfaSStefano Zampini conforming = pcbddc->conforming; 185213b8bfaSStefano Zampini field = pcbddc->nedfield; 186213b8bfaSStefano Zampini global = pcbddc->nedglobal; 187213b8bfaSStefano Zampini setprimal = PETSC_FALSE; 188a13144ffSStefano Zampini print = PETSC_FALSE; 189213b8bfaSStefano Zampini singular = PETSC_FALSE; 190a13144ffSStefano Zampini 191213b8bfaSStefano Zampini /* Command line customization */ 192213b8bfaSStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC Nedelec options","PC");CHKERRQ(ierr); 193213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_field_primal","All edge dofs set as primals: Toselli's algorithm C",NULL,setprimal,&setprimal,NULL);CHKERRQ(ierr); 194213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_singular","Infer nullspace from discrete gradient",NULL,singular,&singular,NULL);CHKERRQ(ierr); 195213b8bfaSStefano Zampini ierr = PetscOptionsInt ("-pc_bddc_nedelec_order","Test variable order code (to be removed)",NULL,order,&order,NULL);CHKERRQ(ierr); 196213b8bfaSStefano Zampini /* print debug info TODO: to be removed */ 197213b8bfaSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_nedelec_print","Print debug info",NULL,print,&print,NULL);CHKERRQ(ierr); 198213b8bfaSStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 199213b8bfaSStefano Zampini 200213b8bfaSStefano Zampini /* Return if there are no edges in the decomposition and the problem is not singular */ 2011e0482f5SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&al2g,NULL);CHKERRQ(ierr); 2021e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(al2g,&n);CHKERRQ(ierr); 203213b8bfaSStefano Zampini ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr); 204213b8bfaSStefano Zampini if (!singular) { 205a13144ffSStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 206a13144ffSStefano Zampini lrc[0] = PETSC_FALSE; 207c2151214SStefano Zampini for (i=0;i<n;i++) { 208a13144ffSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 209a13144ffSStefano Zampini lrc[0] = PETSC_TRUE; 210a13144ffSStefano Zampini break; 211a13144ffSStefano Zampini } 212a13144ffSStefano Zampini } 213a13144ffSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 214a13144ffSStefano Zampini ierr = MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 215a13144ffSStefano Zampini if (!lrc[1]) PetscFunctionReturn(0); 216213b8bfaSStefano Zampini } 217a13144ffSStefano Zampini 218213b8bfaSStefano Zampini /* Get Nedelec field */ 219213b8bfaSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 220c2151214SStefano 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); 221213b8bfaSStefano Zampini if (pcbddc->n_ISForDofsLocal && field >= 0) { 222c2151214SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);CHKERRQ(ierr); 223c2151214SStefano Zampini nedfieldlocal = pcbddc->ISForDofsLocal[field]; 224c2151214SStefano Zampini ierr = ISGetLocalSize(nedfieldlocal,&ne);CHKERRQ(ierr); 225213b8bfaSStefano Zampini } else if (!pcbddc->n_ISForDofsLocal && field != PETSC_DECIDE) { 226213b8bfaSStefano Zampini ne = n; 227213b8bfaSStefano Zampini nedfieldlocal = NULL; 228213b8bfaSStefano Zampini global = PETSC_TRUE; 229213b8bfaSStefano Zampini } else if (field == PETSC_DECIDE) { 230213b8bfaSStefano Zampini PetscInt rst,ren,*idx; 231213b8bfaSStefano Zampini 232213b8bfaSStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 233213b8bfaSStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 234213b8bfaSStefano Zampini ierr = MatGetOwnershipRange(pcbddc->discretegradient,&rst,&ren);CHKERRQ(ierr); 235213b8bfaSStefano Zampini for (i=rst;i<ren;i++) { 236213b8bfaSStefano Zampini PetscInt nc; 237213b8bfaSStefano Zampini 238213b8bfaSStefano Zampini ierr = MatGetRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 239213b8bfaSStefano Zampini if (nc > 1) matis->sf_rootdata[i-rst] = 1; 240213b8bfaSStefano Zampini ierr = MatRestoreRow(pcbddc->discretegradient,i,&nc,NULL,NULL);CHKERRQ(ierr); 241213b8bfaSStefano Zampini } 242213b8bfaSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 243213b8bfaSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 244213b8bfaSStefano Zampini ierr = PetscMalloc1(n,&idx);CHKERRQ(ierr); 245213b8bfaSStefano Zampini for (i=0,ne=0;i<n;i++) if (matis->sf_leafdata[i]) idx[ne++] = i; 246213b8bfaSStefano Zampini ierr = ISCreateGeneral(comm,ne,idx,PETSC_OWN_POINTER,&nedfieldlocal);CHKERRQ(ierr); 247213b8bfaSStefano Zampini } else { 248213b8bfaSStefano Zampini SETERRQ(comm,PETSC_ERR_USER,"When multiple fields are present, the Nedelec field has to be specified"); 249213b8bfaSStefano Zampini } 250213b8bfaSStefano Zampini 251213b8bfaSStefano Zampini /* Sanity checks */ 252213b8bfaSStefano Zampini if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time"); 253213b8bfaSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis"); 254213b8bfaSStefano 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); 255213b8bfaSStefano Zampini 256213b8bfaSStefano Zampini /* Just set primal dofs and return */ 2571e0482f5SStefano Zampini if (setprimal) { 258eee23b56SStefano Zampini IS enedfieldlocal; 259eee23b56SStefano Zampini PetscInt *eidxs; 260eee23b56SStefano Zampini 261eee23b56SStefano Zampini ierr = PetscMalloc1(ne,&eidxs);CHKERRQ(ierr); 262eee23b56SStefano Zampini ierr = VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 263213b8bfaSStefano Zampini if (nedfieldlocal) { 264213b8bfaSStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 265eee23b56SStefano Zampini for (i=0,cum=0;i<ne;i++) { 266eee23b56SStefano Zampini if (PetscRealPart(vals[idxs[i]]) > 2.) { 267eee23b56SStefano Zampini eidxs[cum++] = idxs[i]; 268eee23b56SStefano Zampini } 269eee23b56SStefano Zampini } 270eee23b56SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 271213b8bfaSStefano Zampini } else { 272213b8bfaSStefano Zampini for (i=0,cum=0;i<ne;i++) { 273213b8bfaSStefano Zampini if (PetscRealPart(vals[i]) > 2.) { 274213b8bfaSStefano Zampini eidxs[cum++] = i; 275213b8bfaSStefano Zampini } 276213b8bfaSStefano Zampini } 277213b8bfaSStefano Zampini } 278213b8bfaSStefano Zampini ierr = VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);CHKERRQ(ierr); 279eee23b56SStefano Zampini ierr = ISCreateGeneral(comm,cum,eidxs,PETSC_COPY_VALUES,&enedfieldlocal);CHKERRQ(ierr); 280eee23b56SStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,enedfieldlocal);CHKERRQ(ierr); 281eee23b56SStefano Zampini ierr = PetscFree(eidxs);CHKERRQ(ierr); 2821e0482f5SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 283eee23b56SStefano Zampini ierr = ISDestroy(&enedfieldlocal);CHKERRQ(ierr); 2841e0482f5SStefano Zampini PetscFunctionReturn(0); 2851e0482f5SStefano Zampini } 286a13144ffSStefano Zampini 287213b8bfaSStefano Zampini /* Compute some l2g maps */ 288213b8bfaSStefano Zampini if (nedfieldlocal) { 289c2151214SStefano Zampini IS is; 290c2151214SStefano Zampini 291c2151214SStefano Zampini /* need to map from the local Nedelec field to local numbering */ 292c2151214SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);CHKERRQ(ierr); 2931e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering for the whole dofs*/ 2941e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(al2g,nedfieldlocal,&is);CHKERRQ(ierr); 2951e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is,&al2g);CHKERRQ(ierr); 2961e0482f5SStefano Zampini /* need to map from the local Nedelec field to global numbering (for Nedelec only) */ 2971e0482f5SStefano Zampini if (global) { 2981e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 2991e0482f5SStefano Zampini el2g = al2g; 3001e0482f5SStefano Zampini } else { 3011e0482f5SStefano Zampini IS gis; 3021e0482f5SStefano Zampini 3031e0482f5SStefano Zampini ierr = ISRenumber(is,NULL,NULL,&gis);CHKERRQ(ierr); 3041e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(gis,&el2g);CHKERRQ(ierr); 3051e0482f5SStefano Zampini ierr = ISDestroy(&gis);CHKERRQ(ierr); 3061e0482f5SStefano Zampini } 307c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 308c2151214SStefano Zampini } else { 3091e0482f5SStefano Zampini /* restore default */ 3101e0482f5SStefano Zampini pcbddc->nedfield = -1; 3111e0482f5SStefano Zampini /* one ref for the destruction of al2g, one for el2g */ 3121e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3131e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)al2g);CHKERRQ(ierr); 3141e0482f5SStefano Zampini el2g = al2g; 315c2151214SStefano Zampini fl2g = NULL; 316c2151214SStefano Zampini } 317a13144ffSStefano Zampini 318213b8bfaSStefano Zampini /* Start communication to drop connections for interior edges (for cc analysis only) */ 319c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscInt));CHKERRQ(ierr); 320c2151214SStefano Zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 321c2151214SStefano Zampini if (nedfieldlocal) { 322c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 323c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1; 324c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 325c2151214SStefano Zampini } else { 326c2151214SStefano Zampini for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1; 327c2151214SStefano Zampini } 328c2151214SStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 329c2151214SStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);CHKERRQ(ierr); 330213b8bfaSStefano Zampini 331213b8bfaSStefano Zampini if (!singular) { /* drop connections with interior edges to avoid unneeded communications and memory movements */ 332213b8bfaSStefano Zampini ierr = MatDuplicate(pcbddc->discretegradient,MAT_COPY_VALUES,&G);CHKERRQ(ierr); 333213b8bfaSStefano Zampini ierr = MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 3341e0482f5SStefano Zampini if (global) { 3351e0482f5SStefano Zampini PetscInt rst; 3361e0482f5SStefano Zampini 337c2151214SStefano Zampini ierr = MatGetOwnershipRange(G,&rst,NULL);CHKERRQ(ierr); 338c2151214SStefano Zampini for (i=0,cum=0;i<pc->pmat->rmap->n;i++) { 339c2151214SStefano Zampini if (matis->sf_rootdata[i] < 2) { 340c2151214SStefano Zampini matis->sf_rootdata[cum++] = i + rst; 341c2151214SStefano Zampini } 342c2151214SStefano Zampini } 343a13144ffSStefano Zampini ierr = MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);CHKERRQ(ierr); 344c2151214SStefano Zampini ierr = MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);CHKERRQ(ierr); 3451e0482f5SStefano Zampini } else { 3461e0482f5SStefano Zampini PetscInt *tbz; 3471e0482f5SStefano Zampini 3481e0482f5SStefano Zampini ierr = PetscMalloc1(ne,&tbz);CHKERRQ(ierr); 3491e0482f5SStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 3501e0482f5SStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 3511e0482f5SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3521e0482f5SStefano Zampini for (i=0,cum=0;i<ne;i++) 3531e0482f5SStefano Zampini if (matis->sf_leafdata[idxs[i]] == 1) 3541e0482f5SStefano Zampini tbz[cum++] = i; 3551e0482f5SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 3561e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApply(el2g,cum,tbz,tbz);CHKERRQ(ierr); 3571e0482f5SStefano Zampini ierr = MatZeroRows(G,cum,tbz,0.,NULL,NULL);CHKERRQ(ierr); 3581e0482f5SStefano Zampini ierr = PetscFree(tbz);CHKERRQ(ierr); 3591e0482f5SStefano Zampini } 360213b8bfaSStefano Zampini } else { /* we need the entire G to infer the nullspace */ 361213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->discretegradient);CHKERRQ(ierr); 362213b8bfaSStefano Zampini G = pcbddc->discretegradient; 363213b8bfaSStefano Zampini } 364a13144ffSStefano Zampini 365a13144ffSStefano Zampini /* Extract subdomain relevant rows of G */ 366a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(el2g,&idxs);CHKERRQ(ierr); 367a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);CHKERRQ(ierr); 3687dae84e0SHong Zhang ierr = MatCreateSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);CHKERRQ(ierr); 369a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);CHKERRQ(ierr); 370a13144ffSStefano Zampini ierr = ISDestroy(&lned);CHKERRQ(ierr); 371a13144ffSStefano Zampini ierr = MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);CHKERRQ(ierr); 372a13144ffSStefano Zampini ierr = MatDestroy(&lGall);CHKERRQ(ierr); 373a13144ffSStefano Zampini ierr = MatISGetLocalMat(lGis,&lG);CHKERRQ(ierr); 374a13144ffSStefano Zampini 375213b8bfaSStefano Zampini /* SF for nodal dofs communications */ 376c2151214SStefano Zampini ierr = MatGetLocalSize(G,NULL,&Lv);CHKERRQ(ierr); 377a13144ffSStefano Zampini ierr = MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);CHKERRQ(ierr); 378a13144ffSStefano Zampini ierr = PetscObjectReference((PetscObject)vl2g);CHKERRQ(ierr); 379a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(vl2g,&nv);CHKERRQ(ierr); 380a13144ffSStefano Zampini ierr = PetscSFCreate(comm,&sfv);CHKERRQ(ierr); 381a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(vl2g,&idxs);CHKERRQ(ierr); 382a13144ffSStefano Zampini ierr = PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);CHKERRQ(ierr); 383a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);CHKERRQ(ierr); 384213b8bfaSStefano Zampini i = singular ? 2 : 1; 385213b8bfaSStefano Zampini ierr = PetscMalloc2(i*nv,&sfvleaves,i*Lv,&sfvroots);CHKERRQ(ierr); 386a13144ffSStefano Zampini 3871e0482f5SStefano Zampini /* Destroy temporary G created in MATIS format and modified G */ 388213b8bfaSStefano Zampini ierr = PetscObjectReference((PetscObject)lG);CHKERRQ(ierr); 389a13144ffSStefano Zampini ierr = MatDestroy(&lGis);CHKERRQ(ierr); 390213b8bfaSStefano Zampini ierr = MatDestroy(&G);CHKERRQ(ierr); 391a13144ffSStefano Zampini 392213b8bfaSStefano Zampini if (print) { 393213b8bfaSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"initial_lG");CHKERRQ(ierr); 394213b8bfaSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 395213b8bfaSStefano Zampini } 396213b8bfaSStefano Zampini 397213b8bfaSStefano Zampini /* Save lG for values insertion in change of basis */ 3980569b399SStefano Zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);CHKERRQ(ierr); 3990569b399SStefano Zampini 400a13144ffSStefano Zampini /* Analyze the edge-nodes connections (duplicate lG) */ 4014e64d54eSstefano_zampini ierr = MatDuplicate(lG,MAT_COPY_VALUES,&lGe);CHKERRQ(ierr); 4024e64d54eSstefano_zampini ierr = MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 403a13144ffSStefano Zampini ierr = PetscBTCreate(nv,&btv);CHKERRQ(ierr); 404a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&bte);CHKERRQ(ierr); 405a13144ffSStefano Zampini ierr = PetscBTCreate(ne,&btb);CHKERRQ(ierr); 4067d871cd7SStefano Zampini ierr = PetscBTCreate(ne,&btbd);CHKERRQ(ierr); 407c2151214SStefano Zampini ierr = PetscBTCreate(nv,&btvcand);CHKERRQ(ierr); 408a13144ffSStefano Zampini /* need to import the boundary specification to ensure the 409a13144ffSStefano Zampini proper detection of coarse edges' endpoints */ 410a13144ffSStefano Zampini if (pcbddc->DirichletBoundariesLocal) { 411c2151214SStefano Zampini IS is; 412c2151214SStefano Zampini 413c2151214SStefano Zampini if (fl2g) { 414c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);CHKERRQ(ierr); 415c2151214SStefano Zampini } else { 416c2151214SStefano Zampini is = pcbddc->DirichletBoundariesLocal; 417c2151214SStefano Zampini } 418c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 419c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 420a13144ffSStefano Zampini for (i=0;i<cum;i++) { 421a13144ffSStefano Zampini if (idxs[i] >= 0) { 422a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 4237d871cd7SStefano Zampini ierr = PetscBTSet(btbd,idxs[i]);CHKERRQ(ierr); 424a13144ffSStefano Zampini } 425a13144ffSStefano Zampini } 426c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 427c2151214SStefano Zampini if (fl2g) { 428c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 429c2151214SStefano Zampini } 430a13144ffSStefano Zampini } 431a13144ffSStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 432c2151214SStefano Zampini IS is; 433c2151214SStefano Zampini 434c2151214SStefano Zampini if (fl2g) { 435c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);CHKERRQ(ierr); 436c2151214SStefano Zampini } else { 437c2151214SStefano Zampini is = pcbddc->NeumannBoundariesLocal; 438c2151214SStefano Zampini } 439c2151214SStefano Zampini ierr = ISGetLocalSize(is,&cum);CHKERRQ(ierr); 440c2151214SStefano Zampini ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 441a13144ffSStefano Zampini for (i=0;i<cum;i++) { 442a13144ffSStefano Zampini if (idxs[i] >= 0) { 443a13144ffSStefano Zampini ierr = PetscBTSet(btb,idxs[i]);CHKERRQ(ierr); 444a13144ffSStefano Zampini } 445a13144ffSStefano Zampini } 446c2151214SStefano Zampini ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 447c2151214SStefano Zampini if (fl2g) { 448c2151214SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 449a13144ffSStefano Zampini } 450c2151214SStefano Zampini } 451c2151214SStefano Zampini 452213b8bfaSStefano Zampini /* Count neighs per dof */ 453637e8532SStefano Zampini ierr = PetscCalloc1(ne,&ecount);CHKERRQ(ierr); 454637e8532SStefano Zampini ierr = PetscMalloc1(ne,&eneighs);CHKERRQ(ierr); 455637e8532SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 456637e8532SStefano Zampini for (i=1,cum=0;i<n_neigh;i++) { 457637e8532SStefano Zampini cum += n_shared[i]; 458637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 459637e8532SStefano Zampini ecount[shared[i][j]]++; 460637e8532SStefano Zampini } 461637e8532SStefano Zampini } 462637e8532SStefano Zampini if (ne) { 463637e8532SStefano Zampini ierr = PetscMalloc1(cum,&eneighs[0]);CHKERRQ(ierr); 464637e8532SStefano Zampini } 465637e8532SStefano Zampini for (i=1;i<ne;i++) eneighs[i] = eneighs[i-1] + ecount[i-1]; 466637e8532SStefano Zampini ierr = PetscMemzero(ecount,ne*sizeof(PetscInt));CHKERRQ(ierr); 467637e8532SStefano Zampini for (i=1;i<n_neigh;i++) { 468637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 469637e8532SStefano Zampini PetscInt k = shared[i][j]; 470637e8532SStefano Zampini eneighs[k][ecount[k]] = neigh[i]; 471637e8532SStefano Zampini ecount[k]++; 472637e8532SStefano Zampini } 473637e8532SStefano Zampini } 474637e8532SStefano Zampini for (i=0;i<ne;i++) { 475637e8532SStefano Zampini ierr = PetscSortRemoveDupsInt(&ecount[i],eneighs[i]);CHKERRQ(ierr); 476637e8532SStefano Zampini } 477637e8532SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(el2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 478637e8532SStefano Zampini ierr = PetscCalloc1(nv,&vcount);CHKERRQ(ierr); 479637e8532SStefano Zampini ierr = PetscMalloc1(nv,&vneighs);CHKERRQ(ierr); 480637e8532SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 481637e8532SStefano Zampini for (i=1,cum=0;i<n_neigh;i++) { 482637e8532SStefano Zampini cum += n_shared[i]; 483637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 484637e8532SStefano Zampini vcount[shared[i][j]]++; 485637e8532SStefano Zampini } 486637e8532SStefano Zampini } 487637e8532SStefano Zampini if (nv) { 488637e8532SStefano Zampini ierr = PetscMalloc1(cum,&vneighs[0]);CHKERRQ(ierr); 489637e8532SStefano Zampini } 490637e8532SStefano Zampini for (i=1;i<nv;i++) vneighs[i] = vneighs[i-1] + vcount[i-1]; 491637e8532SStefano Zampini ierr = PetscMemzero(vcount,nv*sizeof(PetscInt));CHKERRQ(ierr); 492637e8532SStefano Zampini for (i=1;i<n_neigh;i++) { 493637e8532SStefano Zampini for (j=0;j<n_shared[i];j++) { 494637e8532SStefano Zampini PetscInt k = shared[i][j]; 495637e8532SStefano Zampini vneighs[k][vcount[k]] = neigh[i]; 496637e8532SStefano Zampini vcount[k]++; 497637e8532SStefano Zampini } 498637e8532SStefano Zampini } 499637e8532SStefano Zampini for (i=0;i<nv;i++) { 500637e8532SStefano Zampini ierr = PetscSortRemoveDupsInt(&vcount[i],vneighs[i]);CHKERRQ(ierr); 501637e8532SStefano Zampini } 502637e8532SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 503637e8532SStefano Zampini 5047d871cd7SStefano Zampini /* need to remove coarse faces' dofs and coarse edges' dirichlet dofs 5057d871cd7SStefano Zampini for proper detection of coarse edges' endpoints */ 50662b0c6f7SStefano Zampini ierr = PetscBTCreate(ne,&btee);CHKERRQ(ierr); 50762b0c6f7SStefano Zampini for (i=0;i<ne;i++) { 5087d871cd7SStefano Zampini if ((ecount[i] > 1 && !PetscBTLookup(btbd,i)) || (ecount[i] == 1 && PetscBTLookup(btb,i))) { 50962b0c6f7SStefano Zampini ierr = PetscBTSet(btee,i);CHKERRQ(ierr); 51062b0c6f7SStefano Zampini } 51162b0c6f7SStefano Zampini } 512637e8532SStefano Zampini ierr = PetscMalloc1(ne,&marks);CHKERRQ(ierr); 51362b0c6f7SStefano Zampini if (!conforming) { 51462b0c6f7SStefano Zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 51562b0c6f7SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 51662b0c6f7SStefano Zampini } 5174e64d54eSstefano_zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 518dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGe,&vals);CHKERRQ(ierr); 51962b0c6f7SStefano Zampini cum = 0; 520a13144ffSStefano Zampini for (i=0;i<ne;i++) { 521dec27d64SStefano Zampini /* eliminate rows corresponding to edge dofs belonging to coarse faces */ 52262b0c6f7SStefano Zampini if (!PetscBTLookup(btee,i)) { 523a13144ffSStefano Zampini marks[cum++] = i; 524dec27d64SStefano Zampini continue; 525dec27d64SStefano Zampini } 526dec27d64SStefano Zampini /* set badly connected edge dofs as primal */ 52762b0c6f7SStefano Zampini if (!conforming) { 52862b0c6f7SStefano Zampini if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */ 529a13144ffSStefano Zampini marks[cum++] = i; 530a13144ffSStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 531a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 532a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 533a13144ffSStefano Zampini } 53462b0c6f7SStefano Zampini } else { 53562b0c6f7SStefano Zampini /* every edge dofs should be connected trough a certain number of nodal dofs 53662b0c6f7SStefano Zampini to other edge dofs belonging to coarse edges 53762b0c6f7SStefano Zampini - at most 2 endpoints 53862b0c6f7SStefano Zampini - order-1 interior nodal dofs 53962b0c6f7SStefano Zampini - no undefined nodal dofs (nconn < order) 54062b0c6f7SStefano Zampini */ 54162b0c6f7SStefano Zampini PetscInt ends = 0,ints = 0, undef = 0; 54262b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 54362b0c6f7SStefano Zampini PetscInt v = jj[j],k; 54462b0c6f7SStefano Zampini PetscInt nconn = iit[v+1]-iit[v]; 54562b0c6f7SStefano Zampini for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--; 54662b0c6f7SStefano Zampini if (nconn > order) ends++; 54762b0c6f7SStefano Zampini else if (nconn == order) ints++; 54862b0c6f7SStefano Zampini else undef++; 54962b0c6f7SStefano Zampini } 55062b0c6f7SStefano Zampini if (undef || ends > 2 || ints != order -1) { 55162b0c6f7SStefano Zampini marks[cum++] = i; 55262b0c6f7SStefano Zampini ierr = PetscBTSet(bte,i);CHKERRQ(ierr); 55362b0c6f7SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 55462b0c6f7SStefano Zampini ierr = PetscBTSet(btv,jj[j]);CHKERRQ(ierr); 55562b0c6f7SStefano Zampini } 55662b0c6f7SStefano Zampini } 55762b0c6f7SStefano Zampini } 558a13144ffSStefano Zampini } 559dec27d64SStefano Zampini /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */ 560dec27d64SStefano Zampini if (!order && ii[i+1] != ii[i]) { 561dec27d64SStefano Zampini PetscScalar val = 1./(ii[i+1]-ii[i]-1); 562dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vals[j] = val; 563a13144ffSStefano Zampini } 564dec27d64SStefano Zampini } 56562b0c6f7SStefano Zampini ierr = PetscBTDestroy(&btee);CHKERRQ(ierr); 566dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGe,&vals);CHKERRQ(ierr); 5674e64d54eSstefano_zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 56862b0c6f7SStefano Zampini if (!conforming) { 56962b0c6f7SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 57062b0c6f7SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 57162b0c6f7SStefano Zampini } 5724e64d54eSstefano_zampini ierr = MatZeroRows(lGe,cum,marks,0.,NULL,NULL);CHKERRQ(ierr); 573637e8532SStefano Zampini 574b03ebc13SStefano Zampini /* identify splitpoints and corner candidates */ 5754e64d54eSstefano_zampini ierr = MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 576a13144ffSStefano Zampini if (print) { 5774e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");CHKERRQ(ierr); 5784e64d54eSstefano_zampini ierr = MatView(lGe,NULL);CHKERRQ(ierr); 5794e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");CHKERRQ(ierr); 580a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 581a13144ffSStefano Zampini } 582a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 583dec27d64SStefano Zampini ierr = MatSeqAIJGetArray(lGt,&vals);CHKERRQ(ierr); 584a13144ffSStefano Zampini for (i=0;i<nv;i++) { 585637e8532SStefano Zampini PetscInt ord = order, test = ii[i+1]-ii[i], vc = vcount[i]; 5867d871cd7SStefano Zampini PetscBool sneighs = PETSC_TRUE, bdir = PETSC_FALSE; 587b03ebc13SStefano Zampini if (!order) { /* variable order */ 588dec27d64SStefano Zampini PetscReal vorder = 0.; 589dec27d64SStefano Zampini 590dec27d64SStefano Zampini for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]); 591dec27d64SStefano Zampini test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON); 592dec27d64SStefano Zampini if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%d)",vorder,test); 593dec27d64SStefano Zampini ord = 1; 594dec27d64SStefano Zampini } 595a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 596dec27d64SStefano 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); 597a13144ffSStefano Zampini #endif 598637e8532SStefano Zampini for (j=ii[i];j<ii[i+1] && sneighs;j++) { 5997d871cd7SStefano Zampini if (PetscBTLookup(btbd,jj[j])) { 6007d871cd7SStefano Zampini bdir = PETSC_TRUE; 6017d871cd7SStefano Zampini break; 6027d871cd7SStefano Zampini } 603637e8532SStefano Zampini if (vc != ecount[jj[j]]) { 604637e8532SStefano Zampini sneighs = PETSC_FALSE; 605637e8532SStefano Zampini } else { 606637e8532SStefano Zampini PetscInt k,*vn = vneighs[i], *en = eneighs[jj[j]]; 607637e8532SStefano Zampini for (k=0;k<vc;k++) { 608637e8532SStefano Zampini if (vn[k] != en[k]) { 609637e8532SStefano Zampini sneighs = PETSC_FALSE; 610637e8532SStefano Zampini break; 611637e8532SStefano Zampini } 612637e8532SStefano Zampini } 613637e8532SStefano Zampini } 614637e8532SStefano Zampini } 6157d871cd7SStefano Zampini if (!sneighs || test >= 3*ord || bdir) { /* splitpoints */ 6167d871cd7SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %d (%d %d %d)\n",i,!sneighs,test >= 3*ord,bdir); 617a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 618dec27d64SStefano Zampini } else if (test == ord) { 619b03ebc13SStefano Zampini if (order == 1 || (!order && ii[i+1]-ii[i] == 1)) { 620a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %d\n",i); 621a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 622a13144ffSStefano Zampini } else { 623a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %d\n",i); 624a13144ffSStefano Zampini ierr = PetscBTSet(btvcand,i);CHKERRQ(ierr); 625a13144ffSStefano Zampini } 626a13144ffSStefano Zampini } 627a13144ffSStefano Zampini } 628637e8532SStefano Zampini ierr = PetscFree(ecount);CHKERRQ(ierr); 629637e8532SStefano Zampini ierr = PetscFree(vcount);CHKERRQ(ierr); 630637e8532SStefano Zampini if (ne) { 631637e8532SStefano Zampini ierr = PetscFree(eneighs[0]);CHKERRQ(ierr); 632637e8532SStefano Zampini } 633637e8532SStefano Zampini if (nv) { 634637e8532SStefano Zampini ierr = PetscFree(vneighs[0]);CHKERRQ(ierr); 635637e8532SStefano Zampini } 636637e8532SStefano Zampini ierr = PetscFree(eneighs);CHKERRQ(ierr); 637637e8532SStefano Zampini ierr = PetscFree(vneighs);CHKERRQ(ierr); 6387d871cd7SStefano Zampini ierr = PetscBTDestroy(&btbd);CHKERRQ(ierr); 639b03ebc13SStefano Zampini 640b03ebc13SStefano Zampini /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */ 641b03ebc13SStefano Zampini if (order != 1) { 642b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"INSPECTING CANDIDATES\n"); 643b03ebc13SStefano Zampini ierr = MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 644b03ebc13SStefano Zampini for (i=0;i<nv;i++) { 645b03ebc13SStefano Zampini if (PetscBTLookup(btvcand,i)) { 646b03ebc13SStefano Zampini PetscBool found = PETSC_FALSE; 647b03ebc13SStefano Zampini for (j=ii[i];j<ii[i+1] && !found;j++) { 648b03ebc13SStefano Zampini PetscInt k,e = jj[j]; 649b03ebc13SStefano Zampini if (PetscBTLookup(bte,e)) continue; 650b03ebc13SStefano Zampini for (k=iit[e];k<iit[e+1];k++) { 651b03ebc13SStefano Zampini PetscInt v = jjt[k]; 652b03ebc13SStefano Zampini if (v != i && PetscBTLookup(btvcand,v)) { 653b03ebc13SStefano Zampini found = PETSC_TRUE; 654b03ebc13SStefano Zampini break; 655b03ebc13SStefano Zampini } 656b03ebc13SStefano Zampini } 657b03ebc13SStefano Zampini } 658b03ebc13SStefano Zampini if (!found) { 659b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d CLEARED\n",i); 660b03ebc13SStefano Zampini ierr = PetscBTClear(btvcand,i);CHKERRQ(ierr); 661b03ebc13SStefano Zampini } else { 662b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %d ACCEPTED\n",i); 663b03ebc13SStefano Zampini } 664b03ebc13SStefano Zampini } 665b03ebc13SStefano Zampini } 666b03ebc13SStefano Zampini ierr = MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 667b03ebc13SStefano Zampini } 668dec27d64SStefano Zampini ierr = MatSeqAIJRestoreArray(lGt,&vals);CHKERRQ(ierr); 669a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 670b03ebc13SStefano Zampini ierr = MatDestroy(&lGe);CHKERRQ(ierr); 671a13144ffSStefano Zampini 672a13144ffSStefano Zampini /* Get the local G^T explicitly */ 6730569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 674a13144ffSStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 6754e64d54eSstefano_zampini ierr = MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 676a13144ffSStefano Zampini 6774e64d54eSstefano_zampini /* Mark interior nodal dofs */ 678a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 6794e64d54eSstefano_zampini ierr = PetscBTCreate(nv,&btvi);CHKERRQ(ierr); 680a13144ffSStefano Zampini for (i=1;i<n_neigh;i++) { 681a13144ffSStefano Zampini for (j=0;j<n_shared[i];j++) { 6824e64d54eSstefano_zampini ierr = PetscBTSet(btvi,shared[i][j]);CHKERRQ(ierr); 683a13144ffSStefano Zampini } 684a13144ffSStefano Zampini } 685a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 686a13144ffSStefano Zampini 687a13144ffSStefano Zampini /* communicate corners and splitpoints */ 688a13144ffSStefano Zampini ierr = PetscMalloc1(nv,&vmarks);CHKERRQ(ierr); 689a13144ffSStefano Zampini ierr = PetscMemzero(sfvleaves,nv*sizeof(PetscInt));CHKERRQ(ierr); 690a13144ffSStefano Zampini ierr = PetscMemzero(sfvroots,Lv*sizeof(PetscInt));CHKERRQ(ierr); 691a13144ffSStefano Zampini for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1; 692a13144ffSStefano Zampini 693a13144ffSStefano Zampini if (print) { 694a13144ffSStefano Zampini IS tbz; 695a13144ffSStefano Zampini 696a13144ffSStefano Zampini cum = 0; 697a13144ffSStefano Zampini for (i=0;i<nv;i++) 698a13144ffSStefano Zampini if (sfvleaves[i]) 699a13144ffSStefano Zampini vmarks[cum++] = i; 700a13144ffSStefano Zampini 701a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 702a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");CHKERRQ(ierr); 703a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 704a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 705a13144ffSStefano Zampini } 706a13144ffSStefano Zampini 707a13144ffSStefano Zampini ierr = PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 708a13144ffSStefano Zampini ierr = PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);CHKERRQ(ierr); 709a13144ffSStefano Zampini ierr = PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 710a13144ffSStefano Zampini ierr = PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves);CHKERRQ(ierr); 711a13144ffSStefano Zampini 7124e64d54eSstefano_zampini /* Zero rows of lGt corresponding to identified corners 7134e64d54eSstefano_zampini and interior nodal dofs */ 714a13144ffSStefano Zampini cum = 0; 715a13144ffSStefano Zampini for (i=0;i<nv;i++) { 716a13144ffSStefano Zampini if (sfvleaves[i]) { 717a13144ffSStefano Zampini vmarks[cum++] = i; 718a13144ffSStefano Zampini ierr = PetscBTSet(btv,i);CHKERRQ(ierr); 719a13144ffSStefano Zampini } 7204e64d54eSstefano_zampini if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i; 721a13144ffSStefano Zampini } 7224e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvi);CHKERRQ(ierr); 723a13144ffSStefano Zampini if (print) { 724a13144ffSStefano Zampini IS tbz; 725a13144ffSStefano Zampini 726a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);CHKERRQ(ierr); 7274e64d54eSstefano_zampini ierr = PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");CHKERRQ(ierr); 728a13144ffSStefano Zampini ierr = ISView(tbz,NULL);CHKERRQ(ierr); 729a13144ffSStefano Zampini ierr = ISDestroy(&tbz);CHKERRQ(ierr); 730a13144ffSStefano Zampini } 731a13144ffSStefano Zampini ierr = MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);CHKERRQ(ierr); 732a13144ffSStefano Zampini ierr = PetscFree(vmarks);CHKERRQ(ierr); 733a13144ffSStefano Zampini ierr = PetscSFDestroy(&sfv);CHKERRQ(ierr); 734a13144ffSStefano Zampini ierr = PetscFree2(sfvleaves,sfvroots);CHKERRQ(ierr); 735a13144ffSStefano Zampini 736a13144ffSStefano Zampini /* Recompute G */ 737a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 738a13144ffSStefano Zampini ierr = MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);CHKERRQ(ierr); 739a13144ffSStefano Zampini if (print) { 740a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lG,"used_lG");CHKERRQ(ierr); 741a13144ffSStefano Zampini ierr = MatView(lG,NULL);CHKERRQ(ierr); 742a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)lGt,"used_lGt");CHKERRQ(ierr); 743a13144ffSStefano Zampini ierr = MatView(lGt,NULL);CHKERRQ(ierr); 744a13144ffSStefano Zampini } 745a13144ffSStefano Zampini 746a13144ffSStefano Zampini /* Get primal dofs (if any) */ 747a13144ffSStefano Zampini cum = 0; 748a13144ffSStefano Zampini for (i=0;i<ne;i++) { 749a13144ffSStefano Zampini if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i; 750a13144ffSStefano Zampini } 751c2151214SStefano Zampini if (fl2g) { 752c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);CHKERRQ(ierr); 753c2151214SStefano Zampini } 754a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 755a13144ffSStefano Zampini if (print) { 756a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");CHKERRQ(ierr); 757a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 758a13144ffSStefano Zampini } 759a13144ffSStefano Zampini ierr = PetscBTDestroy(&bte);CHKERRQ(ierr); 760c2151214SStefano Zampini /* TODO: what if the user passed in some of them ? */ 761a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 762a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 763a13144ffSStefano Zampini 764a13144ffSStefano Zampini /* Compute edge connectivity */ 765a13144ffSStefano Zampini ierr = PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");CHKERRQ(ierr); 766a13144ffSStefano Zampini ierr = MatMatMultSymbolic(lG,lGt,PETSC_DEFAULT,&conn);CHKERRQ(ierr); 767a13144ffSStefano Zampini ierr = MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 768c2151214SStefano Zampini if (fl2g) { 769c2151214SStefano Zampini PetscBT btf; 770c2151214SStefano Zampini PetscInt *iia,*jja,*iiu,*jju; 771c2151214SStefano Zampini PetscBool rest = PETSC_FALSE,free = PETSC_FALSE; 772c2151214SStefano Zampini 773c2151214SStefano Zampini /* create CSR for all local dofs */ 774c2151214SStefano Zampini ierr = PetscMalloc1(n+1,&iia);CHKERRQ(ierr); 775c2151214SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */ 776c2151214SStefano 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); 777c2151214SStefano Zampini iiu = pcbddc->mat_graph->xadj; 778c2151214SStefano Zampini jju = pcbddc->mat_graph->adjncy; 779c2151214SStefano Zampini } else if (pcbddc->use_local_adj) { 780c2151214SStefano Zampini rest = PETSC_TRUE; 781c2151214SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 782c2151214SStefano Zampini } else { 783c2151214SStefano Zampini free = PETSC_TRUE; 784c2151214SStefano Zampini ierr = PetscMalloc2(n+1,&iiu,n,&jju);CHKERRQ(ierr); 785c2151214SStefano Zampini iiu[0] = 0; 786c2151214SStefano Zampini for (i=0;i<n;i++) { 787c2151214SStefano Zampini iiu[i+1] = i+1; 788c2151214SStefano Zampini jju[i] = -1; 789d904f53bSStefano Zampini } 790c2151214SStefano Zampini } 791c2151214SStefano Zampini 792c2151214SStefano Zampini /* import sizes of CSR */ 793c2151214SStefano Zampini iia[0] = 0; 794c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i]; 795c2151214SStefano Zampini 796c2151214SStefano Zampini /* overwrite entries corresponding to the Nedelec field */ 797c2151214SStefano Zampini ierr = PetscBTCreate(n,&btf);CHKERRQ(ierr); 798c2151214SStefano Zampini ierr = ISGetIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 799c2151214SStefano Zampini for (i=0;i<ne;i++) { 800c2151214SStefano Zampini ierr = PetscBTSet(btf,idxs[i]);CHKERRQ(ierr); 801c2151214SStefano Zampini iia[idxs[i]+1] = ii[i+1]-ii[i]; 802c2151214SStefano Zampini } 803c2151214SStefano Zampini 804c2151214SStefano Zampini /* iia in CSR */ 805c2151214SStefano Zampini for (i=0;i<n;i++) iia[i+1] += iia[i]; 806c2151214SStefano Zampini 807c2151214SStefano Zampini /* jja in CSR */ 808c2151214SStefano Zampini ierr = PetscMalloc1(iia[n],&jja);CHKERRQ(ierr); 809c2151214SStefano Zampini for (i=0;i<n;i++) 810c2151214SStefano Zampini if (!PetscBTLookup(btf,i)) 811c2151214SStefano Zampini for (j=0;j<iiu[i+1]-iiu[i];j++) 812c2151214SStefano Zampini jja[iia[i]+j] = jju[iiu[i]+j]; 813c2151214SStefano Zampini 814c2151214SStefano Zampini /* map edge dofs connectivity */ 8151e0482f5SStefano Zampini if (jj) { 816c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);CHKERRQ(ierr); 817c2151214SStefano Zampini for (i=0;i<ne;i++) { 818c2151214SStefano Zampini PetscInt e = idxs[i]; 819c2151214SStefano Zampini for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j]; 820c2151214SStefano Zampini } 8211e0482f5SStefano Zampini } 822c2151214SStefano Zampini ierr = ISRestoreIndices(nedfieldlocal,&idxs);CHKERRQ(ierr); 823c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);CHKERRQ(ierr); 824c2151214SStefano Zampini if (rest) { 825c2151214SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);CHKERRQ(ierr); 826c2151214SStefano Zampini } 827c2151214SStefano Zampini if (free) { 828c2151214SStefano Zampini ierr = PetscFree2(iiu,jju);CHKERRQ(ierr); 829c2151214SStefano Zampini } 830c2151214SStefano Zampini ierr = PetscBTDestroy(&btf);CHKERRQ(ierr); 831c2151214SStefano Zampini } else { 832c2151214SStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);CHKERRQ(ierr); 833c2151214SStefano Zampini } 834c2151214SStefano Zampini 835a13144ffSStefano Zampini /* Analyze interface for edge dofs */ 836a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 837213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 838a13144ffSStefano Zampini 839a13144ffSStefano Zampini /* Get coarse edges in the edge space */ 840c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 841a13144ffSStefano Zampini ierr = MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 842a13144ffSStefano Zampini 843c2151214SStefano Zampini if (fl2g) { 844c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 845c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 846c2151214SStefano Zampini for (i=0;i<nee;i++) { 847c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 848c2151214SStefano Zampini } 849c2151214SStefano Zampini } else { 850c2151214SStefano Zampini eedges = alleedges; 851c2151214SStefano Zampini primals = allprimals; 852c2151214SStefano Zampini } 853c2151214SStefano Zampini 854a13144ffSStefano Zampini /* Mark fine edge dofs with their coarse edge id */ 855a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 856c2151214SStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 857c2151214SStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 858c2151214SStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = nee+1; 859c2151214SStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 860c2151214SStefano Zampini if (print) { 861c2151214SStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");CHKERRQ(ierr); 862c2151214SStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 863c2151214SStefano Zampini } 864c2151214SStefano Zampini 865c2151214SStefano Zampini maxsize = 0; 866a13144ffSStefano Zampini for (i=0;i<nee;i++) { 867a13144ffSStefano Zampini PetscInt size,mark = i+1; 868a13144ffSStefano Zampini 869a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 870a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 871a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 872a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 873a13144ffSStefano Zampini maxsize = PetscMax(maxsize,size); 874a13144ffSStefano Zampini } 875a13144ffSStefano Zampini 876a13144ffSStefano Zampini /* Find coarse edge endpoints */ 877a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 878a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 879a13144ffSStefano Zampini for (i=0;i<nee;i++) { 880a13144ffSStefano Zampini PetscInt mark = i+1,size; 881a13144ffSStefano Zampini 882a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 8831e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 8841e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 885a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 886a13144ffSStefano Zampini if (print) { 887a13144ffSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %d\n",i); 888a13144ffSStefano Zampini ISView(eedges[i],NULL); 889a13144ffSStefano Zampini } 890a13144ffSStefano Zampini for (j=0;j<size;j++) { 891a13144ffSStefano Zampini PetscInt k, ee = idxs[j]; 892a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," idx %d\n",ee); 893a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 894a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," inspect %d\n",jj[k]); 895a13144ffSStefano Zampini if (PetscBTLookup(btv,jj[k])) { 896a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %d\n",jj[k]); 897a13144ffSStefano Zampini } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */ 898a13144ffSStefano Zampini PetscInt k2; 899a13144ffSStefano Zampini PetscBool corner = PETSC_FALSE; 900a13144ffSStefano Zampini for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) { 901c2151214SStefano 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])); 902c2151214SStefano Zampini /* it's a corner if either is connected with an edge dof belonging to a different cc or 903c2151214SStefano Zampini if the edge dof lie on the natural part of the boundary */ 904c2151214SStefano Zampini if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) { 905a13144ffSStefano Zampini corner = PETSC_TRUE; 906a13144ffSStefano Zampini break; 907a13144ffSStefano Zampini } 908a13144ffSStefano Zampini } 909a13144ffSStefano Zampini if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */ 910a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," corner found %d\n",jj[k]); 911a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 912a13144ffSStefano Zampini } else { 913a13144ffSStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n"); 914a13144ffSStefano Zampini } 915a13144ffSStefano Zampini } 916a13144ffSStefano Zampini } 917a13144ffSStefano Zampini } 918a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 919a13144ffSStefano Zampini } 920a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 921a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 922c2151214SStefano Zampini ierr = PetscBTDestroy(&btb);CHKERRQ(ierr); 923a13144ffSStefano Zampini 924a13144ffSStefano Zampini /* Reset marked primal dofs */ 925a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 926a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 927a13144ffSStefano Zampini for (i=0;i<cum;i++) marks[idxs[i]] = 0; 928a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 929a13144ffSStefano Zampini 9300569b399SStefano Zampini /* Now use the initial lG */ 9310569b399SStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 9320569b399SStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 9330569b399SStefano Zampini lG = lGinit; 9340569b399SStefano Zampini ierr = MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);CHKERRQ(ierr); 9350569b399SStefano Zampini 936a13144ffSStefano Zampini /* Compute extended cols indices */ 937b03ebc13SStefano Zampini ierr = PetscBTCreate(nv,&btvc);CHKERRQ(ierr); 938b03ebc13SStefano Zampini ierr = PetscBTCreate(nee,&bter);CHKERRQ(ierr); 939a13144ffSStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 940a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lG,&i);CHKERRQ(ierr); 941a13144ffSStefano Zampini i *= maxsize; 942b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 943a13144ffSStefano Zampini ierr = PetscMalloc2(i,&extrow,i,&gidxs);CHKERRQ(ierr); 944a13144ffSStefano Zampini eerr = PETSC_FALSE; 945a13144ffSStefano Zampini for (i=0;i<nee;i++) { 946b03ebc13SStefano Zampini PetscInt size,found = 0; 947a13144ffSStefano Zampini 948a13144ffSStefano Zampini cum = 0; 949a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 9501e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 9511e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 952a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 953b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 954a13144ffSStefano Zampini for (j=0;j<size;j++) { 955a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 956b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 957b03ebc13SStefano Zampini PetscInt vv = jj[k]; 958b03ebc13SStefano Zampini if (!PetscBTLookup(btv,vv)) extrow[cum++] = vv; 959b03ebc13SStefano Zampini else if (!PetscBTLookupSet(btvc,vv)) found++; 960b03ebc13SStefano Zampini } 961a13144ffSStefano Zampini } 962a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 963a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 964a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 965a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 966a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 967a13144ffSStefano Zampini /* it may happen that endpoints are not defined at this point 968a13144ffSStefano Zampini if it is the case, mark this edge for a second pass */ 969b03ebc13SStefano Zampini if (cum != size -1 || found != 2) { 970b03ebc13SStefano Zampini ierr = PetscBTSet(bter,i);CHKERRQ(ierr); 971a13144ffSStefano Zampini if (print) { 972a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge");CHKERRQ(ierr); 973a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 974a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol");CHKERRQ(ierr); 975a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 976a13144ffSStefano Zampini } 977a13144ffSStefano Zampini eerr = PETSC_TRUE; 978a13144ffSStefano Zampini } 979a13144ffSStefano Zampini } 9804e64d54eSstefano_zampini /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */ 981a13144ffSStefano Zampini ierr = MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);CHKERRQ(ierr); 982a13144ffSStefano Zampini if (done) { 983a13144ffSStefano Zampini PetscInt *newprimals; 984a13144ffSStefano Zampini 985a13144ffSStefano Zampini ierr = PetscMalloc1(ne,&newprimals);CHKERRQ(ierr); 986a13144ffSStefano Zampini ierr = ISGetLocalSize(primals,&cum);CHKERRQ(ierr); 987a13144ffSStefano Zampini ierr = ISGetIndices(primals,&idxs);CHKERRQ(ierr); 988a13144ffSStefano Zampini ierr = PetscMemcpy(newprimals,idxs,cum*sizeof(PetscInt));CHKERRQ(ierr); 989a13144ffSStefano Zampini ierr = ISRestoreIndices(primals,&idxs);CHKERRQ(ierr); 9900569b399SStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 991b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"DOING SECOND PASS (eerr %d)\n",eerr); 992a13144ffSStefano Zampini for (i=0;i<nee;i++) { 993b03ebc13SStefano Zampini PetscBool has_candidates = PETSC_FALSE; 994b03ebc13SStefano Zampini if (PetscBTLookup(bter,i)) { 995a13144ffSStefano Zampini PetscInt size,mark = i+1; 996a13144ffSStefano Zampini 997a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 998a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 999c2151214SStefano Zampini /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */ 1000a13144ffSStefano Zampini for (j=0;j<size;j++) { 1001a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 1002b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF,"Inspecting edge dof %d [%d %d)\n",ee,ii[ee],ii[ee+1]); 1003a13144ffSStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 1004a13144ffSStefano Zampini /* set all candidates located on the edge as corners */ 1005a13144ffSStefano Zampini if (PetscBTLookup(btvcand,jj[k])) { 1006a13144ffSStefano Zampini PetscInt k2,vv = jj[k]; 1007b03ebc13SStefano Zampini has_candidates = PETSC_TRUE; 1008b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Candidate set to vertex %d\n",vv); 1009a13144ffSStefano Zampini ierr = PetscBTSet(btv,vv);CHKERRQ(ierr); 1010a13144ffSStefano Zampini /* set all edge dofs connected to candidate as primals */ 1011a13144ffSStefano Zampini for (k2=iit[vv];k2<iit[vv+1];k2++) { 1012a13144ffSStefano Zampini if (marks[jjt[k2]] == mark) { 1013a13144ffSStefano Zampini PetscInt k3,ee2 = jjt[k2]; 1014b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected edge dof set to primal %d\n",ee2); 1015a13144ffSStefano Zampini newprimals[cum++] = ee2; 1016a13144ffSStefano Zampini /* finally set the new corners */ 1017a13144ffSStefano Zampini for (k3=ii[ee2];k3<ii[ee2+1];k3++) { 1018b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Connected nodal dof set to vertex %d\n",jj[k3]); 1019a13144ffSStefano Zampini ierr = PetscBTSet(btv,jj[k3]);CHKERRQ(ierr); 1020a13144ffSStefano Zampini } 1021a13144ffSStefano Zampini } 1022a13144ffSStefano Zampini } 1023b03ebc13SStefano Zampini } else { 1024b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Not a candidate vertex %d\n",jj[k]); 1025a13144ffSStefano Zampini } 1026a13144ffSStefano Zampini } 1027a13144ffSStefano Zampini } 1028b03ebc13SStefano Zampini if (!has_candidates) { /* circular edge */ 1029b03ebc13SStefano Zampini PetscInt k, ee = idxs[0],*tmarks; 1030b03ebc13SStefano Zampini 1031b03ebc13SStefano Zampini ierr = PetscCalloc1(ne,&tmarks);CHKERRQ(ierr); 1032b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Circular edge %d\n",i); 1033b03ebc13SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) { 1034b03ebc13SStefano Zampini PetscInt k2; 1035b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Set to corner %d\n",jj[k]); 1036b03ebc13SStefano Zampini ierr = PetscBTSet(btv,jj[k]);CHKERRQ(ierr); 1037b03ebc13SStefano Zampini for (k2=iit[jj[k]];k2<iit[jj[k]+1];k2++) tmarks[jjt[k2]]++; 1038b03ebc13SStefano Zampini } 1039b03ebc13SStefano Zampini for (j=0;j<size;j++) { 1040b03ebc13SStefano Zampini if (tmarks[idxs[j]] > 1) { 1041b03ebc13SStefano Zampini if (print) PetscPrintf(PETSC_COMM_SELF," Edge dof set to primal %d\n",idxs[j]); 1042b03ebc13SStefano Zampini newprimals[cum++] = idxs[j]; 1043b03ebc13SStefano Zampini } 1044b03ebc13SStefano Zampini } 1045b03ebc13SStefano Zampini ierr = PetscFree(tmarks);CHKERRQ(ierr); 1046b03ebc13SStefano Zampini } 1047a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1048a13144ffSStefano Zampini } 1049a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 1050a13144ffSStefano Zampini } 1051b03ebc13SStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 10520569b399SStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);CHKERRQ(ierr); 1053a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,newprimals);CHKERRQ(ierr); 1054c2151214SStefano Zampini if (fl2g) { 1055c2151214SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);CHKERRQ(ierr); 1056c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1057c2151214SStefano Zampini for (i=0;i<nee;i++) { 1058c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1059c2151214SStefano Zampini } 1060c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1061c2151214SStefano Zampini } 1062c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1063a13144ffSStefano Zampini ierr = ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);CHKERRQ(ierr); 1064a13144ffSStefano Zampini ierr = PetscFree(newprimals);CHKERRQ(ierr); 1065a13144ffSStefano Zampini ierr = PCBDDCSetPrimalVerticesLocalIS(pc,primals);CHKERRQ(ierr); 1066a13144ffSStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1067a13144ffSStefano Zampini ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr); 1068213b8bfaSStefano Zampini pcbddc->mat_graph->twodim = PETSC_FALSE; 1069c2151214SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1070c2151214SStefano Zampini if (fl2g) { 1071c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);CHKERRQ(ierr); 1072c2151214SStefano Zampini ierr = PetscMalloc1(nee,&eedges);CHKERRQ(ierr); 1073c2151214SStefano Zampini for (i=0;i<nee;i++) { 1074c2151214SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);CHKERRQ(ierr); 1075c2151214SStefano Zampini } 1076c2151214SStefano Zampini } else { 1077c2151214SStefano Zampini eedges = alleedges; 1078c2151214SStefano Zampini primals = allprimals; 1079c2151214SStefano Zampini } 1080b03ebc13SStefano Zampini ierr = PetscCalloc1(nee,&extcols);CHKERRQ(ierr); 1081a13144ffSStefano Zampini 1082a13144ffSStefano Zampini /* Mark again */ 1083a13144ffSStefano Zampini ierr = PetscMemzero(marks,ne*sizeof(PetscInt));CHKERRQ(ierr); 1084a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1085a13144ffSStefano Zampini PetscInt size,mark = i+1; 1086a13144ffSStefano Zampini 1087a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 1088a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1089a13144ffSStefano Zampini for (j=0;j<size;j++) marks[idxs[j]] = mark; 1090a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1091a13144ffSStefano Zampini } 1092a13144ffSStefano Zampini if (print) { 1093a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");CHKERRQ(ierr); 1094a13144ffSStefano Zampini ierr = ISView(primals,NULL);CHKERRQ(ierr); 1095a13144ffSStefano Zampini } 1096a13144ffSStefano Zampini 1097a13144ffSStefano Zampini /* Recompute extended cols */ 1098a13144ffSStefano Zampini eerr = PETSC_FALSE; 1099a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1100a13144ffSStefano Zampini PetscInt size; 1101a13144ffSStefano Zampini 1102a13144ffSStefano Zampini cum = 0; 1103a13144ffSStefano Zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 11041e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 11051e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 1106a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1107a13144ffSStefano Zampini for (j=0;j<size;j++) { 1108a13144ffSStefano Zampini PetscInt k,ee = idxs[j]; 11091e0482f5SStefano Zampini for (k=ii[ee];k<ii[ee+1];k++) if (!PetscBTLookup(btv,jj[k])) extrow[cum++] = jj[k]; 1110a13144ffSStefano Zampini } 1111a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1112a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&cum,extrow);CHKERRQ(ierr); 1113a13144ffSStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);CHKERRQ(ierr); 1114a13144ffSStefano Zampini ierr = PetscSortIntWithArray(cum,gidxs,extrow);CHKERRQ(ierr); 1115a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);CHKERRQ(ierr); 1116a13144ffSStefano Zampini if (cum != size -1) { 1117a13144ffSStefano Zampini if (print) { 1118a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");CHKERRQ(ierr); 1119a13144ffSStefano Zampini ierr = ISView(eedges[i],NULL);CHKERRQ(ierr); 1120a13144ffSStefano Zampini ierr = PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");CHKERRQ(ierr); 1121a13144ffSStefano Zampini ierr = ISView(extcols[i],NULL);CHKERRQ(ierr); 1122a13144ffSStefano Zampini } 1123a13144ffSStefano Zampini eerr = PETSC_TRUE; 1124a13144ffSStefano Zampini } 1125a13144ffSStefano Zampini } 1126a13144ffSStefano Zampini } 1127a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1128a13144ffSStefano Zampini ierr = PetscFree2(extrow,gidxs);CHKERRQ(ierr); 1129b03ebc13SStefano Zampini ierr = PetscBTDestroy(&bter);CHKERRQ(ierr); 11307d871cd7SStefano Zampini if (print) { ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); } 1131a13144ffSStefano Zampini /* an error should not occur at this point */ 1132a13144ffSStefano Zampini if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS"); 1133a13144ffSStefano Zampini 11344e64d54eSstefano_zampini /* Check the number of endpoints */ 11350569b399SStefano Zampini ierr = MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1136b03ebc13SStefano Zampini ierr = PetscMalloc1(2*nee,&corners);CHKERRQ(ierr); 1137b03ebc13SStefano Zampini ierr = PetscMalloc1(nee,&cedges);CHKERRQ(ierr); 11384e64d54eSstefano_zampini for (i=0;i<nee;i++) { 1139b03ebc13SStefano Zampini PetscInt size, found = 0, gc[2]; 11404e64d54eSstefano_zampini 1141b03ebc13SStefano Zampini /* init with defaults */ 1142b03ebc13SStefano Zampini cedges[i] = corners[i*2] = corners[i*2+1] = -1; 11434e64d54eSstefano_zampini ierr = ISGetLocalSize(eedges[i],&size);CHKERRQ(ierr); 11441e0482f5SStefano Zampini if (!size && nedfieldlocal) continue; 11451e0482f5SStefano Zampini if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %d",i); 11464e64d54eSstefano_zampini ierr = ISGetIndices(eedges[i],&idxs);CHKERRQ(ierr); 1147b03ebc13SStefano Zampini ierr = PetscBTMemzero(nv,btvc);CHKERRQ(ierr); 11484e64d54eSstefano_zampini for (j=0;j<size;j++) { 11494e64d54eSstefano_zampini PetscInt k,ee = idxs[j]; 11504e64d54eSstefano_zampini for (k=ii[ee];k<ii[ee+1];k++) { 11514e64d54eSstefano_zampini PetscInt vv = jj[k]; 11524e64d54eSstefano_zampini if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) { 1153b03ebc13SStefano Zampini if (found == 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found more then two corners for edge %d\n",i); 1154b03ebc13SStefano Zampini corners[i*2+found++] = vv; 11554e64d54eSstefano_zampini } 11564e64d54eSstefano_zampini } 11574e64d54eSstefano_zampini } 1158b03ebc13SStefano Zampini if (found != 2) { 1159b03ebc13SStefano Zampini PetscInt e; 1160b03ebc13SStefano Zampini if (fl2g) { 1161b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,1,idxs,&e);CHKERRQ(ierr); 1162b03ebc13SStefano Zampini } else { 1163b03ebc13SStefano Zampini e = idxs[0]; 1164b03ebc13SStefano Zampini } 1165b03ebc13SStefano Zampini SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %d corners for edge %d (astart %d, estart %d)\n",found,i,e,idxs[0]); 1166b03ebc13SStefano Zampini } 1167eee23b56SStefano Zampini 1168eee23b56SStefano Zampini /* get primal dof index on this coarse edge */ 1169b03ebc13SStefano Zampini ierr = ISLocalToGlobalMappingApply(vl2g,2,corners+2*i,gc);CHKERRQ(ierr); 1170b03ebc13SStefano Zampini if (gc[0] > gc[1]) { 1171b03ebc13SStefano Zampini PetscInt swap = corners[2*i]; 1172b03ebc13SStefano Zampini corners[2*i] = corners[2*i+1]; 1173b03ebc13SStefano Zampini corners[2*i+1] = swap; 1174b03ebc13SStefano Zampini } 1175eee23b56SStefano Zampini cedges[i] = idxs[size-1]; 11764e64d54eSstefano_zampini ierr = ISRestoreIndices(eedges[i],&idxs);CHKERRQ(ierr); 1177b03ebc13SStefano 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]); 11784e64d54eSstefano_zampini } 11790569b399SStefano Zampini ierr = MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 11804e64d54eSstefano_zampini ierr = PetscBTDestroy(&btvc);CHKERRQ(ierr); 11814e64d54eSstefano_zampini 1182a13144ffSStefano Zampini #if defined(PETSC_USE_DEBUG) 1183a13144ffSStefano Zampini /* Inspects columns of lG (rows of lGt) and make sure the change of basis will 1184a13144ffSStefano Zampini not interfere with neighbouring coarse edges */ 1185a13144ffSStefano Zampini ierr = PetscMalloc1(nee+1,&emarks);CHKERRQ(ierr); 1186a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1187a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1188a13144ffSStefano Zampini PetscInt emax = 0,eemax = 0; 1189a13144ffSStefano Zampini 1190a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1191a13144ffSStefano Zampini ierr = PetscMemzero(emarks,(nee+1)*sizeof(PetscInt));CHKERRQ(ierr); 1192a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++; 1193a13144ffSStefano Zampini for (j=1;j<nee+1;j++) { 1194a13144ffSStefano Zampini if (emax < emarks[j]) { 1195a13144ffSStefano Zampini emax = emarks[j]; 1196a13144ffSStefano Zampini eemax = j; 1197a13144ffSStefano Zampini } 1198a13144ffSStefano Zampini } 1199a13144ffSStefano Zampini /* not relevant for edges */ 1200a13144ffSStefano Zampini if (!eemax) continue; 1201a13144ffSStefano Zampini 1202a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 1203a13144ffSStefano Zampini if (marks[jj[j]] && marks[jj[j]] != eemax) { 1204c2151214SStefano 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]); 1205a13144ffSStefano Zampini } 1206a13144ffSStefano Zampini } 1207a13144ffSStefano Zampini } 1208a13144ffSStefano Zampini ierr = PetscFree(emarks);CHKERRQ(ierr); 1209a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1210a13144ffSStefano Zampini #endif 1211a13144ffSStefano Zampini 1212a13144ffSStefano Zampini /* Compute extended rows indices for edge blocks of the change of basis */ 1213a13144ffSStefano Zampini ierr = MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1214a13144ffSStefano Zampini ierr = MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);CHKERRQ(ierr); 1215a13144ffSStefano Zampini extmem *= maxsize; 1216a13144ffSStefano Zampini ierr = PetscMalloc1(extmem*nee,&extrow);CHKERRQ(ierr); 1217a13144ffSStefano Zampini ierr = PetscMalloc1(nee,&extrows);CHKERRQ(ierr); 1218a13144ffSStefano Zampini ierr = PetscCalloc1(nee,&extrowcum);CHKERRQ(ierr); 1219a13144ffSStefano Zampini for (i=0;i<nv;i++) { 1220a13144ffSStefano Zampini PetscInt mark = 0,size,start; 1221213b8bfaSStefano Zampini 1222a13144ffSStefano Zampini if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue; 1223a13144ffSStefano Zampini for (j=ii[i];j<ii[i+1];j++) 1224a13144ffSStefano Zampini if (marks[jj[j]] && !mark) 1225a13144ffSStefano Zampini mark = marks[jj[j]]; 1226a13144ffSStefano Zampini 1227a13144ffSStefano Zampini /* not relevant */ 1228a13144ffSStefano Zampini if (!mark) continue; 1229a13144ffSStefano Zampini 1230a13144ffSStefano Zampini /* import extended row */ 1231a13144ffSStefano Zampini mark--; 1232a13144ffSStefano Zampini start = mark*extmem+extrowcum[mark]; 1233a13144ffSStefano Zampini size = ii[i+1]-ii[i]; 1234a13144ffSStefano Zampini if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %d > %d",extrowcum[mark] + size,extmem); 1235a13144ffSStefano Zampini ierr = PetscMemcpy(extrow+start,jj+ii[i],size*sizeof(PetscInt));CHKERRQ(ierr); 1236a13144ffSStefano Zampini extrowcum[mark] += size; 1237a13144ffSStefano Zampini } 1238a13144ffSStefano Zampini ierr = MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);CHKERRQ(ierr); 1239213b8bfaSStefano Zampini ierr = MatDestroy(&lGt);CHKERRQ(ierr); 1240213b8bfaSStefano Zampini ierr = PetscFree(marks);CHKERRQ(ierr); 1241213b8bfaSStefano Zampini 1242213b8bfaSStefano Zampini /* Compress extrows */ 1243a13144ffSStefano Zampini cum = 0; 1244a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1245a13144ffSStefano Zampini PetscInt size = extrowcum[i],*start = extrow + i*extmem; 1246a13144ffSStefano Zampini ierr = PetscSortRemoveDupsInt(&size,start);CHKERRQ(ierr); 1247a13144ffSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);CHKERRQ(ierr); 1248a13144ffSStefano Zampini cum = PetscMax(cum,size); 1249a13144ffSStefano Zampini } 1250a13144ffSStefano Zampini ierr = PetscFree(extrowcum);CHKERRQ(ierr); 1251a13144ffSStefano Zampini ierr = PetscBTDestroy(&btv);CHKERRQ(ierr); 1252a13144ffSStefano Zampini ierr = PetscBTDestroy(&btvcand);CHKERRQ(ierr); 1253a13144ffSStefano Zampini 1254a13144ffSStefano Zampini /* Workspace for lapack inner calls and VecSetValues */ 1255a13144ffSStefano Zampini ierr = PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);CHKERRQ(ierr); 1256a13144ffSStefano Zampini 1257a13144ffSStefano Zampini /* Create change of basis matrix (preallocation can be improved) */ 1258a13144ffSStefano Zampini ierr = MatCreate(comm,&T);CHKERRQ(ierr); 1259c2151214SStefano Zampini ierr = MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n, 1260c2151214SStefano Zampini pc->pmat->rmap->N,pc->pmat->rmap->N);CHKERRQ(ierr); 1261a13144ffSStefano Zampini ierr = MatSetType(T,MATAIJ);CHKERRQ(ierr); 1262a13144ffSStefano Zampini ierr = MatSeqAIJSetPreallocation(T,10,NULL);CHKERRQ(ierr); 1263a13144ffSStefano Zampini ierr = MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);CHKERRQ(ierr); 12641e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(T,al2g,al2g);CHKERRQ(ierr); 1265a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1266a13144ffSStefano Zampini ierr = MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 1267213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&al2g);CHKERRQ(ierr); 1268a13144ffSStefano Zampini 1269a13144ffSStefano Zampini /* Defaults to identity */ 1270c2151214SStefano Zampini ierr = MatCreateVecs(pc->pmat,&tvec,NULL);CHKERRQ(ierr); 1271a13144ffSStefano Zampini ierr = VecSet(tvec,1.0);CHKERRQ(ierr); 1272a13144ffSStefano Zampini ierr = MatDiagonalSet(T,tvec,INSERT_VALUES);CHKERRQ(ierr); 1273a13144ffSStefano Zampini ierr = VecDestroy(&tvec);CHKERRQ(ierr); 1274a13144ffSStefano Zampini 12751e0482f5SStefano Zampini /* Create discrete gradient for the coarser level if needed */ 12761e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 12771e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 12781e0482f5SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 12791e0482f5SStefano Zampini ISLocalToGlobalMapping cel2g,cvl2g; 12801e0482f5SStefano Zampini IS wis,gwis; 12811e0482f5SStefano Zampini PetscInt cnv,cne; 12821e0482f5SStefano Zampini 12831e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,nee,cedges,PETSC_COPY_VALUES,&wis);CHKERRQ(ierr); 12841e0482f5SStefano Zampini if (fl2g) { 12851e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(fl2g,wis,&pcbddc->nedclocal);CHKERRQ(ierr); 12861e0482f5SStefano Zampini } else { 12871e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)wis);CHKERRQ(ierr); 12881e0482f5SStefano Zampini pcbddc->nedclocal = wis; 12891e0482f5SStefano Zampini } 12901e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(el2g,wis,&gwis);CHKERRQ(ierr); 12911e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12921e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cne,&wis);CHKERRQ(ierr); 12931e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cel2g);CHKERRQ(ierr); 12941e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 12951e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 12961e0482f5SStefano Zampini 12971e0482f5SStefano Zampini ierr = ISCreateGeneral(comm,2*nee,corners,PETSC_USE_POINTER,&wis);CHKERRQ(ierr); 12981e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(vl2g,wis,&gwis);CHKERRQ(ierr); 12991e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13001e0482f5SStefano Zampini ierr = ISRenumber(gwis,NULL,&cnv,&wis);CHKERRQ(ierr); 13011e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(wis,&cvl2g);CHKERRQ(ierr); 13021e0482f5SStefano Zampini ierr = ISDestroy(&wis);CHKERRQ(ierr); 13031e0482f5SStefano Zampini ierr = ISDestroy(&gwis);CHKERRQ(ierr); 13041e0482f5SStefano Zampini 13051e0482f5SStefano Zampini ierr = MatCreate(comm,&pcbddc->nedcG);CHKERRQ(ierr); 13061e0482f5SStefano Zampini ierr = MatSetSizes(pcbddc->nedcG,PETSC_DECIDE,PETSC_DECIDE,cne,cnv);CHKERRQ(ierr); 13071e0482f5SStefano Zampini ierr = MatSetType(pcbddc->nedcG,MATAIJ);CHKERRQ(ierr); 13081e0482f5SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->nedcG,2,NULL);CHKERRQ(ierr); 13091e0482f5SStefano Zampini ierr = MatMPIAIJSetPreallocation(pcbddc->nedcG,2,NULL,2,NULL);CHKERRQ(ierr); 13101e0482f5SStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->nedcG,cel2g,cvl2g);CHKERRQ(ierr); 13111e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cel2g);CHKERRQ(ierr); 13121e0482f5SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cvl2g);CHKERRQ(ierr); 13131e0482f5SStefano Zampini } 1314213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&vl2g);CHKERRQ(ierr); 13151e0482f5SStefano Zampini 13161e0482f5SStefano Zampini #if defined(PRINT_GDET) 13171e0482f5SStefano Zampini inc = 0; 13181e0482f5SStefano Zampini lev = pcbddc->current_level; 13191e0482f5SStefano Zampini #endif 1320213b8bfaSStefano Zampini 1321213b8bfaSStefano Zampini /* Insert values in the change of basis matrix */ 1322a13144ffSStefano Zampini for (i=0;i<nee;i++) { 1323a13144ffSStefano Zampini Mat Gins = NULL, GKins = NULL; 13241e0482f5SStefano Zampini IS cornersis = NULL; 13251e0482f5SStefano Zampini PetscScalar cvals[2]; 1326a13144ffSStefano Zampini 13271e0482f5SStefano Zampini if (pcbddc->nedcG) { 13281e0482f5SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,2,corners+2*i,PETSC_USE_POINTER,&cornersis);CHKERRQ(ierr); 13291e0482f5SStefano Zampini } 13301e0482f5SStefano Zampini ierr = PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],cornersis,&Gins,&GKins,cvals,work,rwork);CHKERRQ(ierr); 1331a13144ffSStefano Zampini if (Gins && GKins) { 1332a13144ffSStefano Zampini PetscScalar *data; 1333a13144ffSStefano Zampini const PetscInt *rows,*cols; 1334a13144ffSStefano Zampini PetscInt nrh,nch,nrc,ncc; 1335a13144ffSStefano Zampini 1336a13144ffSStefano Zampini ierr = ISGetIndices(eedges[i],&cols);CHKERRQ(ierr); 1337a13144ffSStefano Zampini /* H1 */ 1338a13144ffSStefano Zampini ierr = ISGetIndices(extrows[i],&rows);CHKERRQ(ierr); 1339a13144ffSStefano Zampini ierr = MatGetSize(Gins,&nrh,&nch);CHKERRQ(ierr); 1340a13144ffSStefano Zampini ierr = MatDenseGetArray(Gins,&data);CHKERRQ(ierr); 1341a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);CHKERRQ(ierr); 1342a13144ffSStefano Zampini ierr = MatDenseRestoreArray(Gins,&data);CHKERRQ(ierr); 1343a13144ffSStefano Zampini ierr = ISRestoreIndices(extrows[i],&rows);CHKERRQ(ierr); 1344a13144ffSStefano Zampini /* complement */ 1345a13144ffSStefano Zampini ierr = MatGetSize(GKins,&nrc,&ncc);CHKERRQ(ierr); 13461e0482f5SStefano Zampini if (!ncc) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Constant function has not been generated for coarse edge %d",i); 1347213b8bfaSStefano 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); 1348213b8bfaSStefano 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); 1349a13144ffSStefano Zampini ierr = MatDenseGetArray(GKins,&data);CHKERRQ(ierr); 1350a13144ffSStefano Zampini ierr = MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);CHKERRQ(ierr); 1351a13144ffSStefano Zampini ierr = MatDenseRestoreArray(GKins,&data);CHKERRQ(ierr); 13521e0482f5SStefano Zampini 13531e0482f5SStefano Zampini /* coarse discrete gradient */ 13541e0482f5SStefano Zampini if (pcbddc->nedcG) { 13551e0482f5SStefano Zampini PetscInt cols[2]; 13561e0482f5SStefano Zampini 13571e0482f5SStefano Zampini cols[0] = 2*i; 13581e0482f5SStefano Zampini cols[1] = 2*i+1; 13591e0482f5SStefano Zampini ierr = MatSetValuesLocal(pcbddc->nedcG,1,&i,2,cols,cvals,INSERT_VALUES);CHKERRQ(ierr); 13601e0482f5SStefano Zampini } 1361a13144ffSStefano Zampini ierr = ISRestoreIndices(eedges[i],&cols);CHKERRQ(ierr); 1362a13144ffSStefano Zampini } 1363a13144ffSStefano Zampini ierr = ISDestroy(&extrows[i]);CHKERRQ(ierr); 1364a13144ffSStefano Zampini ierr = ISDestroy(&extcols[i]);CHKERRQ(ierr); 13651e0482f5SStefano Zampini ierr = ISDestroy(&cornersis);CHKERRQ(ierr); 1366a13144ffSStefano Zampini ierr = MatDestroy(&Gins);CHKERRQ(ierr); 1367a13144ffSStefano Zampini ierr = MatDestroy(&GKins);CHKERRQ(ierr); 1368a13144ffSStefano Zampini } 1369213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&el2g);CHKERRQ(ierr); 1370a13144ffSStefano Zampini 1371a13144ffSStefano Zampini /* Start assembling */ 1372a13144ffSStefano Zampini ierr = MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13731e0482f5SStefano Zampini if (pcbddc->nedcG) { 13741e0482f5SStefano Zampini ierr = MatAssemblyBegin(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13751e0482f5SStefano Zampini } 1376a13144ffSStefano Zampini 1377a13144ffSStefano Zampini /* Free */ 1378c2151214SStefano Zampini if (fl2g) { 1379c2151214SStefano Zampini ierr = ISDestroy(&primals);CHKERRQ(ierr); 1380c2151214SStefano Zampini for (i=0;i<nee;i++) { 1381c2151214SStefano Zampini ierr = ISDestroy(&eedges[i]);CHKERRQ(ierr); 1382c2151214SStefano Zampini } 1383c2151214SStefano Zampini ierr = PetscFree(eedges);CHKERRQ(ierr); 1384c2151214SStefano Zampini } 1385eee23b56SStefano Zampini 1386eee23b56SStefano Zampini /* hack mat_graph with primal dofs on the coarse edges */ 1387eee23b56SStefano Zampini { 1388eee23b56SStefano Zampini PCBDDCGraph graph = pcbddc->mat_graph; 1389eee23b56SStefano Zampini PetscInt *oqueue = graph->queue; 1390eee23b56SStefano Zampini PetscInt *ocptr = graph->cptr; 1391eee23b56SStefano Zampini PetscInt ncc,*idxs; 1392eee23b56SStefano Zampini 1393eee23b56SStefano Zampini /* find first primal edge */ 1394eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1395eee23b56SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1396eee23b56SStefano Zampini } else { 1397eee23b56SStefano Zampini if (fl2g) { 1398eee23b56SStefano Zampini ierr = ISLocalToGlobalMappingApply(fl2g,nee,cedges,cedges);CHKERRQ(ierr); 1399eee23b56SStefano Zampini } 1400eee23b56SStefano Zampini idxs = cedges; 1401eee23b56SStefano Zampini } 1402eee23b56SStefano Zampini cum = 0; 1403eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1404eee23b56SStefano Zampini 1405eee23b56SStefano Zampini /* adapt connected components */ 1406eee23b56SStefano Zampini ierr = PetscMalloc2(graph->nvtxs+1,&graph->cptr,ocptr[graph->ncc],&graph->queue);CHKERRQ(ierr); 1407eee23b56SStefano Zampini graph->cptr[0] = 0; 1408eee23b56SStefano Zampini for (i=0,ncc=0;i<graph->ncc;i++) { 1409eee23b56SStefano Zampini PetscInt lc = ocptr[i+1]-ocptr[i]; 1410eee23b56SStefano Zampini if (cum != nee && oqueue[ocptr[i+1]-1] == cedges[cum]) { /* this cc has a primal dof */ 1411eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc]+1; 1412eee23b56SStefano Zampini graph->queue[graph->cptr[ncc]] = cedges[cum]; 1413eee23b56SStefano Zampini ncc++; 1414eee23b56SStefano Zampini lc--; 1415eee23b56SStefano Zampini cum++; 1416eee23b56SStefano Zampini while (cum < nee && cedges[cum] < 0) cum++; 1417eee23b56SStefano Zampini } 1418eee23b56SStefano Zampini graph->cptr[ncc+1] = graph->cptr[ncc] + lc; 1419eee23b56SStefano Zampini for (j=0;j<lc;j++) graph->queue[graph->cptr[ncc]+j] = oqueue[ocptr[i]+j]; 1420eee23b56SStefano Zampini ncc++; 1421eee23b56SStefano Zampini } 1422eee23b56SStefano Zampini graph->ncc = ncc; 1423eee23b56SStefano Zampini if (pcbddc->nedclocal) { 1424eee23b56SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);CHKERRQ(ierr); 1425eee23b56SStefano Zampini } 1426eee23b56SStefano Zampini ierr = PetscFree2(ocptr,oqueue);CHKERRQ(ierr); 1427eee23b56SStefano Zampini } 1428213b8bfaSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&fl2g);CHKERRQ(ierr); 1429c2151214SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);CHKERRQ(ierr); 1430c2151214SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1431213b8bfaSStefano Zampini ierr = MatDestroy(&conn);CHKERRQ(ierr); 1432eee23b56SStefano Zampini 1433c2151214SStefano Zampini ierr = ISDestroy(&nedfieldlocal);CHKERRQ(ierr); 1434a13144ffSStefano Zampini ierr = PetscFree(extrow);CHKERRQ(ierr); 1435a13144ffSStefano Zampini ierr = PetscFree2(work,rwork);CHKERRQ(ierr); 1436b03ebc13SStefano Zampini ierr = PetscFree(corners);CHKERRQ(ierr); 1437b03ebc13SStefano Zampini ierr = PetscFree(cedges);CHKERRQ(ierr); 1438a13144ffSStefano Zampini ierr = PetscFree(extrows);CHKERRQ(ierr); 1439a13144ffSStefano Zampini ierr = PetscFree(extcols);CHKERRQ(ierr); 1440a13144ffSStefano Zampini ierr = MatDestroy(&lG);CHKERRQ(ierr); 1441a13144ffSStefano Zampini 1442a13144ffSStefano Zampini /* Complete assembling */ 1443a13144ffSStefano Zampini ierr = MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14441e0482f5SStefano Zampini if (pcbddc->nedcG) { 14451e0482f5SStefano Zampini ierr = MatAssemblyEnd(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14461e0482f5SStefano Zampini #if 0 14471e0482f5SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->nedcG,"coarse_G");CHKERRQ(ierr); 14481e0482f5SStefano Zampini ierr = MatView(pcbddc->nedcG,NULL);CHKERRQ(ierr); 14491e0482f5SStefano Zampini #endif 14501e0482f5SStefano Zampini } 1451a13144ffSStefano Zampini 1452a13144ffSStefano Zampini /* set change of basis */ 1453213b8bfaSStefano Zampini ierr = PCBDDCSetChangeOfBasisMat(pc,T,singular);CHKERRQ(ierr); 1454a13144ffSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1455a13144ffSStefano Zampini 1456a13144ffSStefano Zampini PetscFunctionReturn(0); 1457a13144ffSStefano Zampini } 1458a13144ffSStefano Zampini 1459d8203eabSStefano Zampini /* the near-null space of BDDC carries information on quadrature weights, 1460d8203eabSStefano Zampini and these can be collinear -> so cheat with MatNullSpaceCreate 1461d8203eabSStefano Zampini and create a suitable set of basis vectors first */ 1462d8203eabSStefano Zampini PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp) 1463d8203eabSStefano Zampini { 1464d8203eabSStefano Zampini PetscErrorCode ierr; 1465d8203eabSStefano Zampini PetscInt i; 1466d8203eabSStefano Zampini 1467d8203eabSStefano Zampini PetscFunctionBegin; 1468d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { 1469d8203eabSStefano Zampini PetscInt first,last; 1470d8203eabSStefano Zampini 1471d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 147286fa73c5SStefano Zampini if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented"); 1473d8203eabSStefano Zampini if (i>=first && i < last) { 1474d8203eabSStefano Zampini PetscScalar *data; 1475d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1476d8203eabSStefano Zampini if (!has_const) { 1477d8203eabSStefano Zampini data[i-first] = 1.; 1478d8203eabSStefano Zampini } else { 147986fa73c5SStefano Zampini data[2*i-first] = 1./PetscSqrtReal(2.); 148086fa73c5SStefano Zampini data[2*i-first+1] = -1./PetscSqrtReal(2.); 1481d8203eabSStefano Zampini } 1482d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1483d8203eabSStefano Zampini } 1484d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1485d8203eabSStefano Zampini } 1486d8203eabSStefano Zampini ierr = MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);CHKERRQ(ierr); 1487d8203eabSStefano Zampini for (i=0;i<nvecs;i++) { /* reset vectors */ 1488d8203eabSStefano Zampini PetscInt first,last; 1489ddc40e2cSstefano_zampini ierr = VecLockPop(quad_vecs[i]);CHKERRQ(ierr); 1490d8203eabSStefano Zampini ierr = VecGetOwnershipRange(quad_vecs[i],&first,&last);CHKERRQ(ierr); 1491d8203eabSStefano Zampini if (i>=first && i < last) { 1492d8203eabSStefano Zampini PetscScalar *data; 1493d8203eabSStefano Zampini ierr = VecGetArray(quad_vecs[i],&data);CHKERRQ(ierr); 1494d8203eabSStefano Zampini if (!has_const) { 1495d8203eabSStefano Zampini data[i-first] = 0.; 1496d8203eabSStefano Zampini } else { 149786fa73c5SStefano Zampini data[2*i-first] = 0.; 149886fa73c5SStefano Zampini data[2*i-first+1] = 0.; 1499d8203eabSStefano Zampini } 1500d8203eabSStefano Zampini ierr = VecRestoreArray(quad_vecs[i],&data);CHKERRQ(ierr); 1501d8203eabSStefano Zampini } 1502d8203eabSStefano Zampini ierr = PetscObjectStateIncrease((PetscObject)quad_vecs[i]);CHKERRQ(ierr); 1503ddc40e2cSstefano_zampini ierr = VecLockPush(quad_vecs[i]);CHKERRQ(ierr); 1504d8203eabSStefano Zampini } 1505d8203eabSStefano Zampini PetscFunctionReturn(0); 1506d8203eabSStefano Zampini } 1507d8203eabSStefano Zampini 15088ae0ca82SStefano Zampini PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp) 1509669cc0f4SStefano Zampini { 1510a198735bSStefano Zampini Mat loc_divudotp; 1511fa23a32eSStefano Zampini Vec p,v,vins,quad_vec,*quad_vecs; 15128ae0ca82SStefano Zampini ISLocalToGlobalMapping map; 1513669cc0f4SStefano Zampini IS *faces,*edges; 1514669cc0f4SStefano Zampini PetscScalar *vals; 1515669cc0f4SStefano Zampini const PetscScalar *array; 1516669cc0f4SStefano Zampini PetscInt i,maxneighs,lmaxneighs,maxsize,nf,ne; 15171ae86dd6SStefano Zampini PetscMPIInt rank; 1518a198735bSStefano Zampini PetscErrorCode ierr; 1519669cc0f4SStefano Zampini 1520669cc0f4SStefano Zampini PetscFunctionBegin; 1521669cc0f4SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1522669cc0f4SStefano Zampini if (graph->twodim) { 1523669cc0f4SStefano Zampini lmaxneighs = 2; 1524669cc0f4SStefano Zampini } else { 1525669cc0f4SStefano Zampini lmaxneighs = 1; 1526669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1527669cc0f4SStefano Zampini const PetscInt *idxs; 1528669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1529669cc0f4SStefano Zampini lmaxneighs = PetscMax(lmaxneighs,graph->count[idxs[0]]); 1530669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1531669cc0f4SStefano Zampini } 1532669cc0f4SStefano Zampini lmaxneighs++; /* graph count does not include self */ 1533669cc0f4SStefano Zampini } 1534669cc0f4SStefano Zampini ierr = MPIU_Allreduce(&lmaxneighs,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1535669cc0f4SStefano Zampini maxsize = 0; 1536669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1537669cc0f4SStefano Zampini PetscInt nn; 1538669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1539669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1540669cc0f4SStefano Zampini } 1541669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1542669cc0f4SStefano Zampini PetscInt nn; 1543669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1544669cc0f4SStefano Zampini maxsize = PetscMax(maxsize,nn); 1545669cc0f4SStefano Zampini } 1546669cc0f4SStefano Zampini ierr = PetscMalloc1(maxsize,&vals);CHKERRQ(ierr); 1547669cc0f4SStefano Zampini /* create vectors to hold quadrature weights */ 1548669cc0f4SStefano Zampini ierr = MatCreateVecs(A,&quad_vec,NULL);CHKERRQ(ierr); 15498ae0ca82SStefano Zampini if (!transpose) { 15508ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,&map,NULL);CHKERRQ(ierr); 15518ae0ca82SStefano Zampini } else { 15528ae0ca82SStefano Zampini ierr = MatGetLocalToGlobalMapping(A,NULL,&map);CHKERRQ(ierr); 15538ae0ca82SStefano Zampini } 1554669cc0f4SStefano Zampini ierr = VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);CHKERRQ(ierr); 15551ae86dd6SStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 1556d8203eabSStefano Zampini ierr = PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);CHKERRQ(ierr); 1557669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1558ddc40e2cSstefano_zampini ierr = VecLockPop(quad_vecs[i]);CHKERRQ(ierr); 15598ae0ca82SStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vecs[i],map);CHKERRQ(ierr); 1560669cc0f4SStefano Zampini } 1561d8203eabSStefano Zampini 1562669cc0f4SStefano Zampini /* compute local quad vec */ 1563a198735bSStefano Zampini ierr = MatISGetLocalMat(divudotp,&loc_divudotp);CHKERRQ(ierr); 15648ae0ca82SStefano Zampini if (!transpose) { 1565a198735bSStefano Zampini ierr = MatCreateVecs(loc_divudotp,&v,&p);CHKERRQ(ierr); 15668ae0ca82SStefano Zampini } else { 15678ae0ca82SStefano Zampini ierr = MatCreateVecs(loc_divudotp,&p,&v);CHKERRQ(ierr); 15688ae0ca82SStefano Zampini } 1569669cc0f4SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 15708ae0ca82SStefano Zampini if (!transpose) { 1571a198735bSStefano Zampini ierr = MatMultTranspose(loc_divudotp,p,v);CHKERRQ(ierr); 15728ae0ca82SStefano Zampini } else { 15738ae0ca82SStefano Zampini ierr = MatMult(loc_divudotp,p,v);CHKERRQ(ierr); 15748ae0ca82SStefano Zampini } 1575fa23a32eSStefano Zampini if (vl2l) { 1576fa23a32eSStefano Zampini ierr = VecGetSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1577fa23a32eSStefano Zampini } else { 1578fa23a32eSStefano Zampini vins = v; 1579fa23a32eSStefano Zampini } 1580fa23a32eSStefano Zampini ierr = VecGetArrayRead(vins,&array);CHKERRQ(ierr); 1581669cc0f4SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 15829a962809SStefano Zampini 15831ae86dd6SStefano Zampini /* insert in global quadrature vecs */ 15841ae86dd6SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 1585669cc0f4SStefano Zampini for (i=0;i<nf;i++) { 1586669cc0f4SStefano Zampini const PetscInt *idxs; 1587669cc0f4SStefano Zampini PetscInt idx,nn,j; 1588669cc0f4SStefano Zampini 1589669cc0f4SStefano Zampini ierr = ISGetIndices(faces[i],&idxs);CHKERRQ(ierr); 1590669cc0f4SStefano Zampini ierr = ISGetLocalSize(faces[i],&nn);CHKERRQ(ierr); 1591669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 15921ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1593669cc0f4SStefano Zampini idx = -(idx+1); 1594669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1595669cc0f4SStefano Zampini ierr = ISRestoreIndices(faces[i],&idxs);CHKERRQ(ierr); 1596669cc0f4SStefano Zampini } 1597669cc0f4SStefano Zampini for (i=0;i<ne;i++) { 1598669cc0f4SStefano Zampini const PetscInt *idxs; 1599669cc0f4SStefano Zampini PetscInt idx,nn,j; 1600669cc0f4SStefano Zampini 1601669cc0f4SStefano Zampini ierr = ISGetIndices(edges[i],&idxs);CHKERRQ(ierr); 1602669cc0f4SStefano Zampini ierr = ISGetLocalSize(edges[i],&nn);CHKERRQ(ierr); 1603669cc0f4SStefano Zampini for (j=0;j<nn;j++) vals[j] = array[idxs[j]]; 16041ae86dd6SStefano Zampini ierr = PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);CHKERRQ(ierr); 1605669cc0f4SStefano Zampini idx = -(idx+1); 1606669cc0f4SStefano Zampini ierr = VecSetValuesLocal(quad_vecs[idx],nn,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 1607669cc0f4SStefano Zampini ierr = ISRestoreIndices(edges[i],&idxs);CHKERRQ(ierr); 1608669cc0f4SStefano Zampini } 1609c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nf,&faces,&ne,&edges,NULL);CHKERRQ(ierr); 1610fa23a32eSStefano Zampini ierr = VecRestoreArrayRead(vins,&array);CHKERRQ(ierr); 1611fa23a32eSStefano Zampini if (vl2l) { 1612fa23a32eSStefano Zampini ierr = VecRestoreSubVector(v,vl2l,&vins);CHKERRQ(ierr); 1613fa23a32eSStefano Zampini } 1614669cc0f4SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1615669cc0f4SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 1616669cc0f4SStefano Zampini 1617669cc0f4SStefano Zampini /* assemble near null space */ 1618669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1619669cc0f4SStefano Zampini ierr = VecAssemblyBegin(quad_vecs[i]);CHKERRQ(ierr); 1620669cc0f4SStefano Zampini } 1621669cc0f4SStefano Zampini for (i=0;i<maxneighs;i++) { 1622669cc0f4SStefano Zampini ierr = VecAssemblyEnd(quad_vecs[i]);CHKERRQ(ierr); 1623ddc40e2cSstefano_zampini ierr = VecLockPush(quad_vecs[i]);CHKERRQ(ierr); 1624669cc0f4SStefano Zampini } 1625669cc0f4SStefano Zampini ierr = VecDestroyVecs(maxneighs,&quad_vecs);CHKERRQ(ierr); 1626669cc0f4SStefano Zampini PetscFunctionReturn(0); 1627669cc0f4SStefano Zampini } 1628669cc0f4SStefano Zampini 1629669cc0f4SStefano Zampini 16301f4df5f7SStefano Zampini PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc) 16311f4df5f7SStefano Zampini { 16321f4df5f7SStefano Zampini PetscErrorCode ierr; 16331f4df5f7SStefano Zampini Vec local,global; 16341f4df5f7SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 16351f4df5f7SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 16361f4df5f7SStefano Zampini 16371f4df5f7SStefano Zampini PetscFunctionBegin; 16381f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 16391f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 16401f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 16411f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 16421f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 16431f4df5f7SStefano Zampini PetscInt i; 16441f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16451f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 16461f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16471f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 16481f4df5f7SStefano Zampini } 16491f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 16501f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 16511f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 16521f4df5f7SStefano Zampini } 16531f4df5f7SStefano Zampini } else { 1654986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 16551f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1656986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 1657986cdee1SStefano Zampini if (i > 1) { 1658986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 16591f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16601f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16611f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16621f4df5f7SStefano Zampini } 16631f4df5f7SStefano Zampini } 16647a0e7b2cSstefano_zampini } else { 16657a0e7b2cSstefano_zampini PetscInt i; 16667a0e7b2cSstefano_zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16677a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16687a0e7b2cSstefano_zampini } 16691f4df5f7SStefano Zampini } 1670986cdee1SStefano Zampini } 16711f4df5f7SStefano Zampini 16721f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 16731f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16747a0e7b2cSstefano_zampini } else if (pcbddc->DirichletBoundariesLocal) { 16757a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16761f4df5f7SStefano Zampini } 16771f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 16781f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16797a0e7b2cSstefano_zampini } else if (pcbddc->NeumannBoundariesLocal) { 16807a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16811f4df5f7SStefano Zampini } 16821f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 16831f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 16841f4df5f7SStefano Zampini } 16851f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 16861f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 16877a0e7b2cSstefano_zampini 16887a0e7b2cSstefano_zampini PetscFunctionReturn(0); 16897a0e7b2cSstefano_zampini } 16907a0e7b2cSstefano_zampini 16917a0e7b2cSstefano_zampini PetscErrorCode PCBDDCConsistencyCheckIS(PC pc, MPI_Op mop, IS *is) 16927a0e7b2cSstefano_zampini { 16937a0e7b2cSstefano_zampini Mat_IS *matis = (Mat_IS*)(pc->pmat->data); 16947a0e7b2cSstefano_zampini PetscErrorCode ierr; 16957a0e7b2cSstefano_zampini IS nis; 16967a0e7b2cSstefano_zampini const PetscInt *idxs; 16977a0e7b2cSstefano_zampini PetscInt i,nd,n = matis->A->rmap->n,*nidxs,nnd; 16987a0e7b2cSstefano_zampini PetscBool *ld; 16997a0e7b2cSstefano_zampini 17007a0e7b2cSstefano_zampini PetscFunctionBegin; 17017a0e7b2cSstefano_zampini if (mop != MPI_LAND && mop != MPI_LOR) SETERRQ(PetscObjectComm((PetscObject)(pc)),PETSC_ERR_SUP,"Supported are MPI_LAND and MPI_LOR"); 17027a0e7b2cSstefano_zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 17037a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 17047a0e7b2cSstefano_zampini /* init rootdata with true */ 17057a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_rootdata; 17067a0e7b2cSstefano_zampini for (i=0;i<pc->pmat->rmap->n;i++) ld[i] = PETSC_TRUE; 17077a0e7b2cSstefano_zampini } else { 17087a0e7b2cSstefano_zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscBool));CHKERRQ(ierr); 17097a0e7b2cSstefano_zampini } 17107a0e7b2cSstefano_zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscBool));CHKERRQ(ierr); 17117a0e7b2cSstefano_zampini ierr = ISGetLocalSize(*is,&nd);CHKERRQ(ierr); 17127a0e7b2cSstefano_zampini ierr = ISGetIndices(*is,&idxs);CHKERRQ(ierr); 17137a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_leafdata; 17147a0e7b2cSstefano_zampini for (i=0;i<nd;i++) 17157a0e7b2cSstefano_zampini if (-1 < idxs[i] && idxs[i] < n) 17167a0e7b2cSstefano_zampini ld[idxs[i]] = PETSC_TRUE; 17177a0e7b2cSstefano_zampini ierr = ISRestoreIndices(*is,&idxs);CHKERRQ(ierr); 17187a0e7b2cSstefano_zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 17197a0e7b2cSstefano_zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 17207a0e7b2cSstefano_zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 17217a0e7b2cSstefano_zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 17227a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 17237a0e7b2cSstefano_zampini ierr = PetscMalloc1(nd,&nidxs);CHKERRQ(ierr); 17247a0e7b2cSstefano_zampini } else { 17257a0e7b2cSstefano_zampini ierr = PetscMalloc1(n,&nidxs);CHKERRQ(ierr); 17267a0e7b2cSstefano_zampini } 17277a0e7b2cSstefano_zampini for (i=0,nnd=0;i<n;i++) 17287a0e7b2cSstefano_zampini if (ld[i]) 17297a0e7b2cSstefano_zampini nidxs[nnd++] = i; 17307a0e7b2cSstefano_zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(*is)),nnd,nidxs,PETSC_OWN_POINTER,&nis);CHKERRQ(ierr); 17317a0e7b2cSstefano_zampini ierr = ISDestroy(is);CHKERRQ(ierr); 17327a0e7b2cSstefano_zampini *is = nis; 17331f4df5f7SStefano Zampini PetscFunctionReturn(0); 17341f4df5f7SStefano Zampini } 17351f4df5f7SStefano Zampini 17363e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 17373e589ea0SStefano Zampini { 17383e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 17393e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 17403e589ea0SStefano Zampini PetscErrorCode ierr; 17413e589ea0SStefano Zampini 17423e589ea0SStefano Zampini PetscFunctionBegin; 17433e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 17443e589ea0SStefano Zampini PetscFunctionReturn(0); 17453e589ea0SStefano Zampini } 17463e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 17473e589ea0SStefano Zampini Vec swap; 17483e589ea0SStefano Zampini 17493e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 17503e589ea0SStefano Zampini swap = pcbddc->work_change; 17513e589ea0SStefano Zampini pcbddc->work_change = r; 17523e589ea0SStefano Zampini r = swap; 17533e589ea0SStefano Zampini } 17543e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 17553e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 17563e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 17573e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 17583e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 17593e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 17603e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1761f913dca9SStefano Zampini pcbddc->work_change = r; 17623e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 17633e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 17643e589ea0SStefano Zampini } 17653e589ea0SStefano Zampini PetscFunctionReturn(0); 17663e589ea0SStefano Zampini } 17673e589ea0SStefano Zampini 1768a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1769a3df083aSStefano Zampini { 1770a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1771a3df083aSStefano Zampini PetscErrorCode ierr; 1772a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1773a3df083aSStefano Zampini 1774a3df083aSStefano Zampini PetscFunctionBegin; 1775a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1776a3df083aSStefano Zampini if (transpose) { 1777a3df083aSStefano Zampini apply_right = ctx->apply_left; 1778a3df083aSStefano Zampini apply_left = ctx->apply_right; 1779a3df083aSStefano Zampini } else { 1780a3df083aSStefano Zampini apply_right = ctx->apply_right; 1781a3df083aSStefano Zampini apply_left = ctx->apply_left; 1782a3df083aSStefano Zampini } 1783a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1784a3df083aSStefano Zampini if (apply_right) { 1785a3df083aSStefano Zampini const PetscScalar *ax; 1786a3df083aSStefano Zampini PetscInt nl,i; 1787a3df083aSStefano Zampini 1788a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1789a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1790a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 1791a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1792a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1793a3df083aSStefano Zampini PetscScalar sum,val; 1794a3df083aSStefano Zampini const PetscInt *idxs; 1795a3df083aSStefano Zampini PetscInt nz,j; 1796a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1797a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1798a3df083aSStefano Zampini sum = 0.; 1799a3df083aSStefano Zampini if (ctx->apply_p0) { 1800a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1801a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1802a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1803a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1804a3df083aSStefano Zampini } 1805a3df083aSStefano Zampini } else { 1806a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1807a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1808a3df083aSStefano Zampini } 1809a3df083aSStefano Zampini } 1810a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1811a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1812a3df083aSStefano Zampini } 1813a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1814a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1815a3df083aSStefano Zampini } 1816a3df083aSStefano Zampini if (transpose) { 1817a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1818a3df083aSStefano Zampini } else { 1819a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1820a3df083aSStefano Zampini } 1821a3df083aSStefano Zampini if (reset_x) { 1822a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1823a3df083aSStefano Zampini } 1824a3df083aSStefano Zampini if (apply_left) { 1825a3df083aSStefano Zampini PetscScalar *ay; 1826a3df083aSStefano Zampini PetscInt i; 1827a3df083aSStefano Zampini 1828a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1829a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1830a3df083aSStefano Zampini PetscScalar sum,val; 1831a3df083aSStefano Zampini const PetscInt *idxs; 1832a3df083aSStefano Zampini PetscInt nz,j; 1833a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1834a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1835a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 1836a3df083aSStefano Zampini if (ctx->apply_p0) { 1837a3df083aSStefano Zampini sum = 0.; 1838a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1839a3df083aSStefano Zampini sum += ay[idxs[j]]; 1840a3df083aSStefano Zampini ay[idxs[j]] += val; 1841a3df083aSStefano Zampini } 1842a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 1843a3df083aSStefano Zampini } else { 1844a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1845a3df083aSStefano Zampini ay[idxs[j]] += val; 1846a3df083aSStefano Zampini } 1847a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 1848a3df083aSStefano Zampini } 1849a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1850a3df083aSStefano Zampini } 1851a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 1852a3df083aSStefano Zampini } 1853a3df083aSStefano Zampini PetscFunctionReturn(0); 1854a3df083aSStefano Zampini } 1855a3df083aSStefano Zampini 1856a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 1857a3df083aSStefano Zampini { 1858a3df083aSStefano Zampini PetscErrorCode ierr; 1859a3df083aSStefano Zampini 1860a3df083aSStefano Zampini PetscFunctionBegin; 1861a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 1862a3df083aSStefano Zampini PetscFunctionReturn(0); 1863a3df083aSStefano Zampini } 1864a3df083aSStefano Zampini 1865a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 1866a3df083aSStefano Zampini { 1867a3df083aSStefano Zampini PetscErrorCode ierr; 1868a3df083aSStefano Zampini 1869a3df083aSStefano Zampini PetscFunctionBegin; 1870a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 1871a3df083aSStefano Zampini PetscFunctionReturn(0); 1872a3df083aSStefano Zampini } 1873a3df083aSStefano Zampini 1874a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 1875a3df083aSStefano Zampini { 1876a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1877a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1878a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1879a3df083aSStefano Zampini PetscErrorCode ierr; 1880a3df083aSStefano Zampini 1881a3df083aSStefano Zampini PetscFunctionBegin; 1882a3df083aSStefano Zampini if (!restore) { 18831dd7afcfSStefano Zampini Mat A_IB,A_BI; 1884a3df083aSStefano Zampini PetscScalar *work; 1885b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 1886a3df083aSStefano Zampini 18879a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 18889a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 1889a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 1890a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 1891a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1892a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 1893a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 1894a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 1895a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 1896a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 1897a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 1898a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 1899a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 1900a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 1901059032f7SStefano Zampini if (reuse) { 1902a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 19031dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 1904059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 1905059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 1906059032f7SStefano Zampini PetscInt i; 1907059032f7SStefano Zampini 1908059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 1909059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1910059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1911059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1912059032f7SStefano Zampini } 1913059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 19141dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 1915059032f7SStefano Zampini } 1916a3df083aSStefano Zampini ctx->A = pcis->A_IB; 1917a3df083aSStefano Zampini ctx->work = work; 1918a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 1919a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1920a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1921a3df083aSStefano Zampini pcis->A_IB = A_IB; 1922a3df083aSStefano Zampini 1923a3df083aSStefano Zampini /* A_BI as A_IB^T */ 1924a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 1925a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 1926a3df083aSStefano Zampini pcis->A_BI = A_BI; 1927a3df083aSStefano Zampini } else { 19281dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 19291dd7afcfSStefano Zampini PetscFunctionReturn(0); 19301dd7afcfSStefano Zampini } 1931a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 1932a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 1933a3df083aSStefano Zampini pcis->A_IB = ctx->A; 19341dd7afcfSStefano Zampini ctx->A = NULL; 19351dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 19361dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 19371dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 19381dd7afcfSStefano Zampini if (ctx->free) { 1939059032f7SStefano Zampini PetscInt i; 19401dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1941059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1942059032f7SStefano Zampini } 1943059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1944059032f7SStefano Zampini } 1945a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 1946a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 1947a3df083aSStefano Zampini } 1948a3df083aSStefano Zampini PetscFunctionReturn(0); 1949a3df083aSStefano Zampini } 1950a3df083aSStefano Zampini 1951a3df083aSStefano Zampini /* used just in bddc debug mode */ 1952a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 1953a3df083aSStefano Zampini { 1954a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1955a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1956a3df083aSStefano Zampini Mat An; 1957a3df083aSStefano Zampini PetscErrorCode ierr; 1958a3df083aSStefano Zampini 1959a3df083aSStefano Zampini PetscFunctionBegin; 1960a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 1961a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 1962a3df083aSStefano Zampini if (is1) { 19637dae84e0SHong Zhang ierr = MatCreateSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 1964a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 1965a3df083aSStefano Zampini } else { 1966a3df083aSStefano Zampini *B = An; 1967a3df083aSStefano Zampini } 1968a3df083aSStefano Zampini PetscFunctionReturn(0); 1969a3df083aSStefano Zampini } 1970a3df083aSStefano Zampini 19711cf9b237SStefano Zampini /* TODO: add reuse flag */ 19721cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 19731cf9b237SStefano Zampini { 19741cf9b237SStefano Zampini Mat Bt; 19751cf9b237SStefano Zampini PetscScalar *a,*bdata; 19761cf9b237SStefano Zampini const PetscInt *ii,*ij; 19771cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 19781cf9b237SStefano Zampini PetscBool flg_row; 19791cf9b237SStefano Zampini PetscErrorCode ierr; 19801cf9b237SStefano Zampini 19811cf9b237SStefano Zampini PetscFunctionBegin; 19821cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 19831cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 19841cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 19851cf9b237SStefano Zampini nnz = n; 19861cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 19871cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 19881cf9b237SStefano Zampini } 19891cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 19901cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 19911cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 19921cf9b237SStefano Zampini nnz = 0; 19931cf9b237SStefano Zampini bii[0] = 0; 19941cf9b237SStefano Zampini for (i=0;i<n;i++) { 19951cf9b237SStefano Zampini PetscInt j; 19961cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 19971cf9b237SStefano Zampini PetscScalar entry = a[j]; 19981cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 19991cf9b237SStefano Zampini bij[nnz] = ij[j]; 20001cf9b237SStefano Zampini bdata[nnz] = entry; 20011cf9b237SStefano Zampini nnz++; 20021cf9b237SStefano Zampini } 20031cf9b237SStefano Zampini } 20041cf9b237SStefano Zampini bii[i+1] = nnz; 20051cf9b237SStefano Zampini } 20061cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 20071cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 20081cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 20091cf9b237SStefano Zampini { 20101cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 20111cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 20121cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 20131cf9b237SStefano Zampini } 20141cf9b237SStefano Zampini *B = Bt; 20151cf9b237SStefano Zampini PetscFunctionReturn(0); 20161cf9b237SStefano Zampini } 20171cf9b237SStefano Zampini 20184f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 20194f1b2e48SStefano Zampini { 20204f1b2e48SStefano Zampini Mat B; 20214f1b2e48SStefano Zampini IS is_dummy,*cc_n; 20224f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 20234f1b2e48SStefano Zampini PCBDDCGraph graph; 20244f1b2e48SStefano Zampini PetscInt i,n; 20254f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 20264f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 20274f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 20284f1b2e48SStefano Zampini PetscErrorCode ierr; 20294f1b2e48SStefano Zampini 20304f1b2e48SStefano Zampini PetscFunctionBegin; 203163c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 203263c961adSStefano Zampini *ncc = 0; 203363c961adSStefano Zampini *cc = NULL; 203463c961adSStefano Zampini PetscFunctionReturn(0); 203563c961adSStefano Zampini } 20364f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 20374f1b2e48SStefano Zampini if (!isseqaij && filter) { 20381cf9b237SStefano Zampini PetscBool isseqdense; 20391cf9b237SStefano Zampini 20401cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 20411cf9b237SStefano Zampini if (!isseqdense) { 20424f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 20431cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 20441cf9b237SStefano Zampini PetscScalar *array; 20451cf9b237SStefano Zampini PetscReal chop=1.e-6; 20461cf9b237SStefano Zampini 20471cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 20481cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 20491cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 20501cf9b237SStefano Zampini for (i=0;i<n;i++) { 20511cf9b237SStefano Zampini PetscInt j; 20521cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 20531cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 20541cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 20551cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 20561cf9b237SStefano Zampini } 20571cf9b237SStefano Zampini } 20581cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 20599d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 20601cf9b237SStefano Zampini } 20614f1b2e48SStefano Zampini } else { 20624f1b2e48SStefano Zampini B = A; 20634f1b2e48SStefano Zampini } 20644f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 20654f1b2e48SStefano Zampini 20664f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 20674f1b2e48SStefano Zampini if (filter) { 20684f1b2e48SStefano Zampini PetscScalar *data; 20694f1b2e48SStefano Zampini PetscInt j,cum; 20704f1b2e48SStefano Zampini 20714f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 20724f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 20734f1b2e48SStefano Zampini cum = 0; 20744f1b2e48SStefano Zampini for (i=0;i<n;i++) { 20754f1b2e48SStefano Zampini PetscInt t; 20764f1b2e48SStefano Zampini 20774f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 20784f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 20794f1b2e48SStefano Zampini continue; 20804f1b2e48SStefano Zampini } 20814f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 20824f1b2e48SStefano Zampini } 20834f1b2e48SStefano Zampini t = xadj_filtered[i]; 20844f1b2e48SStefano Zampini xadj_filtered[i] = cum; 20854f1b2e48SStefano Zampini cum += t; 20864f1b2e48SStefano Zampini } 20874f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 20884f1b2e48SStefano Zampini } else { 20894f1b2e48SStefano Zampini xadj_filtered = NULL; 20904f1b2e48SStefano Zampini adjncy_filtered = NULL; 20914f1b2e48SStefano Zampini } 20924f1b2e48SStefano Zampini 20934f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 20944f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 20954f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 20964f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 20974f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 2098be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 20994f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 21004f1b2e48SStefano Zampini if (xadj_filtered) { 21014f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 21024f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 21034f1b2e48SStefano Zampini } else { 21044f1b2e48SStefano Zampini graph->xadj = xadj; 21054f1b2e48SStefano Zampini graph->adjncy = adjncy; 21064f1b2e48SStefano Zampini } 21074f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 21084f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 21094f1b2e48SStefano Zampini /* partial clean up */ 21104f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 21114f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 21121cf9b237SStefano Zampini if (A != B) { 21134f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 21144f1b2e48SStefano Zampini } 21154f1b2e48SStefano Zampini 21164f1b2e48SStefano Zampini /* get back data */ 21171cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 21181cf9b237SStefano Zampini if (cc) { 21194f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 21204f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 21214f1b2e48SStefano 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); 21224f1b2e48SStefano Zampini } 21234f1b2e48SStefano Zampini *cc = cc_n; 21241cf9b237SStefano Zampini } 21254f1b2e48SStefano Zampini /* clean up graph */ 21264f1b2e48SStefano Zampini graph->xadj = 0; 21274f1b2e48SStefano Zampini graph->adjncy = 0; 21284f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 21294f1b2e48SStefano Zampini PetscFunctionReturn(0); 21304f1b2e48SStefano Zampini } 21314f1b2e48SStefano Zampini 21325408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 21335408967cSStefano Zampini { 21345408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 21355408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2136dee84bffSStefano Zampini IS dirIS = NULL; 21374f1b2e48SStefano Zampini PetscInt i; 21385408967cSStefano Zampini PetscErrorCode ierr; 21395408967cSStefano Zampini 21405408967cSStefano Zampini PetscFunctionBegin; 2141dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 21425408967cSStefano Zampini if (zerodiag) { 21435408967cSStefano Zampini Mat A; 21445408967cSStefano Zampini Vec vec3_N; 21455408967cSStefano Zampini PetscScalar *vals; 21465408967cSStefano Zampini const PetscInt *idxs; 2147d12d3064SStefano Zampini PetscInt nz,*count; 21485408967cSStefano Zampini 21495408967cSStefano Zampini /* p0 */ 21505408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 21515408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 21525408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 21535408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 21544f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 21555408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21565408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 21575408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 21585408967cSStefano Zampini /* v_I */ 21595408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 21605408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 21615408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21625408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 21635408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 21645408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 21655408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21665408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 21675408967cSStefano Zampini if (dirIS) { 21685408967cSStefano Zampini PetscInt n; 21695408967cSStefano Zampini 21705408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 21715408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 21725408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 21735408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21745408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 21755408967cSStefano Zampini } 21765408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 21775408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 21785408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 21795408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 2180669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 21815408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 21825408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 21839a962809SStefano 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])); 21845408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 21855408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 2186d12d3064SStefano Zampini 2187d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 2188d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 2189d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2190d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 2191d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2192d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 21939a962809SStefano 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]); 2194d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 2195d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 21965408967cSStefano Zampini } 2197dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 21985408967cSStefano Zampini 21995408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 22005408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 22014f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 22025408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 22034f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 22045408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 2205f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2206f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 2207f2a566d8SStefano 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); 2208f2a566d8SStefano Zampini } 22095408967cSStefano Zampini PetscFunctionReturn(0); 22105408967cSStefano Zampini } 22115408967cSStefano Zampini 2212339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 2213339f8db1SStefano Zampini { 2214339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 22154edc6404Sstefano_zampini IS pressures,zerodiag,zerodiag_save,*zerodiag_subs; 2216b0f5fe93SStefano Zampini PetscInt nz,n; 22174edc6404Sstefano_zampini PetscInt *interior_dofs,n_interior_dofs,nneu; 22184edc6404Sstefano_zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag,checkb; 2219339f8db1SStefano Zampini PetscErrorCode ierr; 2220339f8db1SStefano Zampini 2221339f8db1SStefano Zampini PetscFunctionBegin; 22229f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 22239f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 2224a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 2225a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 2226a3df083aSStefano Zampini } 2227a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 2228a3df083aSStefano Zampini pcbddc->benign_n = 0; 222928b8efb1Sstefano_zampini 223028b8efb1Sstefano_zampini /* if a local info on dofs is present, uses the last field for "pressures" (or fid by command line) 22314f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 22324f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 22334f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 22341ae86dd6SStefano Zampini since the local Schur complements are already SPD 22354f1b2e48SStefano Zampini */ 22364f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 22374f1b2e48SStefano Zampini have_null = PETSC_TRUE; 223840fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 22394f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 22404f1b2e48SStefano Zampini 224128b8efb1Sstefano_zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC benign options","PC");CHKERRQ(ierr); 224228b8efb1Sstefano_zampini ierr = PetscOptionsInt ("-pc_bddc_pressure_field","Field id for pressures",NULL,p,&p,NULL);CHKERRQ(ierr); 224328b8efb1Sstefano_zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 224428b8efb1Sstefano_zampini if (p < 0 || p > pcbddc->n_ISForDofsLocal-1) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Invalid field id for pressures %D",p); 22454f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 22464f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 22474f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 22484f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 2249ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 225040fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 225140fa8d13SStefano Zampini if (!sorted) { 225240fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 225340fa8d13SStefano Zampini } 225440fa8d13SStefano Zampini } else { 225540fa8d13SStefano Zampini pressures = NULL; 225640fa8d13SStefano Zampini } 225797d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 225897d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 225927b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 226097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 2261339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 2262339f8db1SStefano Zampini if (!sorted) { 2263339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 2264339f8db1SStefano Zampini } 22654edc6404Sstefano_zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 22664edc6404Sstefano_zampini zerodiag_save = zerodiag; 2267339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 22684f1b2e48SStefano Zampini if (!nz) { 22694f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 22704f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 227140fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 227240fa8d13SStefano Zampini } 22734f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 22744f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 22754f1b2e48SStefano Zampini zerodiag_subs = NULL; 22764f1b2e48SStefano Zampini pcbddc->benign_n = 0; 22771f4df5f7SStefano Zampini n_interior_dofs = 0; 22781f4df5f7SStefano Zampini interior_dofs = NULL; 22794edc6404Sstefano_zampini nneu = 0; 22804edc6404Sstefano_zampini if (pcbddc->NeumannBoundariesLocal) { 22814edc6404Sstefano_zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nneu);CHKERRQ(ierr); 22824edc6404Sstefano_zampini } 22833369cb78Sstefano_zampini checkb = (PetscBool)(!pcbddc->NeumannBoundariesLocal || pcbddc->current_level); 22844edc6404Sstefano_zampini if (checkb) { /* need to compute interior nodes */ 22851f4df5f7SStefano Zampini PetscInt n,i,j; 22861f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 22871f4df5f7SStefano Zampini PetscInt *iwork; 22881f4df5f7SStefano Zampini 22891f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 22901f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 22911f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 22921f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 229390648384SStefano Zampini for (i=1;i<n_neigh;i++) 22941f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 22951f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 22961f4df5f7SStefano Zampini for (i=0;i<n;i++) 22971f4df5f7SStefano Zampini if (!iwork[i]) 22981f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 22991f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 23001f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 23011f4df5f7SStefano Zampini } 23024f1b2e48SStefano Zampini if (has_null_pressures) { 23034f1b2e48SStefano Zampini IS *subs; 23044edc6404Sstefano_zampini PetscInt nsubs,i,j,nl; 23051f4df5f7SStefano Zampini const PetscInt *idxs; 23061f4df5f7SStefano Zampini PetscScalar *array; 23071f4df5f7SStefano Zampini Vec *work; 23081f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 23094f1b2e48SStefano Zampini 23104f1b2e48SStefano Zampini subs = pcbddc->local_subs; 23114f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 23121f4df5f7SStefano 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) */ 23134edc6404Sstefano_zampini if (checkb) { 23141f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 23151f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 23161f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 23171f4df5f7SStefano Zampini /* work[0] = 1_p */ 23181f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 23191f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 23201f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 23211f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 23221f4df5f7SStefano Zampini /* work[0] = 1_v */ 23231f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 23241f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 23251f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 23261f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 23271f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 23281f4df5f7SStefano Zampini } 23294f1b2e48SStefano Zampini if (nsubs > 1) { 23304f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 23314f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 23324f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 23334f1b2e48SStefano Zampini IS t_zerodiag_subs; 23344f1b2e48SStefano Zampini PetscInt nl; 23354f1b2e48SStefano Zampini 23364f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 23374f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 23384f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 23394f1b2e48SStefano Zampini if (nl) { 23404f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 23414f1b2e48SStefano Zampini 23424edc6404Sstefano_zampini if (checkb) { 23431f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 23441f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 23451f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 23461f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 23471f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 23481f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 23491f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 23501f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 23511f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 23521f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 23531f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 23541f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 23551f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 23561f4df5f7SStefano Zampini valid = PETSC_FALSE; 23571f4df5f7SStefano Zampini break; 23581f4df5f7SStefano Zampini } 23591f4df5f7SStefano Zampini } 23601f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 23611f4df5f7SStefano Zampini } 23626632bad2Sstefano_zampini if (valid && nneu) { 23636632bad2Sstefano_zampini const PetscInt *idxs; 23641f4df5f7SStefano Zampini PetscInt nzb; 23651f4df5f7SStefano Zampini 23666632bad2Sstefano_zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 23676632bad2Sstefano_zampini ierr = ISGlobalToLocalMappingApply(l2g,IS_GTOLM_DROP,nneu,idxs,&nzb,NULL);CHKERRQ(ierr); 23686632bad2Sstefano_zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 23691f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 23701f4df5f7SStefano Zampini } 23711f4df5f7SStefano Zampini if (valid && pressures) { 23724f1b2e48SStefano Zampini IS t_pressure_subs; 23734f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 23744f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 23754f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 23764f1b2e48SStefano Zampini } 23774f1b2e48SStefano Zampini if (valid) { 23784f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 23794f1b2e48SStefano Zampini pcbddc->benign_n++; 23804f1b2e48SStefano Zampini } else { 23814f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 23824f1b2e48SStefano Zampini } 23834f1b2e48SStefano Zampini } 23844f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 23854f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 23864f1b2e48SStefano Zampini } 23874f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 23884f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 23891f4df5f7SStefano Zampini 23906632bad2Sstefano_zampini if (nneu) valid = PETSC_FALSE; 23911f4df5f7SStefano Zampini if (valid && pressures) { 23924f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 23934f1b2e48SStefano Zampini } 23944edc6404Sstefano_zampini if (valid && checkb) { 23951f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 23961f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 23971f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 23981f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 23991f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 24001f4df5f7SStefano Zampini valid = PETSC_FALSE; 24011f4df5f7SStefano Zampini break; 24021f4df5f7SStefano Zampini } 24031f4df5f7SStefano Zampini } 24041f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 24051f4df5f7SStefano Zampini } 24064f1b2e48SStefano Zampini if (valid) { 24074f1b2e48SStefano Zampini pcbddc->benign_n = 1; 2408ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 24094f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 24104f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 24114f1b2e48SStefano Zampini } 24124f1b2e48SStefano Zampini } 24134edc6404Sstefano_zampini if (checkb) { 24141f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 24154f1b2e48SStefano Zampini } 24161f4df5f7SStefano Zampini } 24171f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 24184f1b2e48SStefano Zampini 24194f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2420b9b0e38cSStefano Zampini PetscInt n; 2421b9b0e38cSStefano Zampini 24224f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 24234f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2424b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2425b9b0e38cSStefano Zampini if (n) { 24264f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 24274f1b2e48SStefano Zampini have_null = PETSC_FALSE; 24284f1b2e48SStefano Zampini } 2429b9b0e38cSStefano Zampini } 24304f1b2e48SStefano Zampini 24314f1b2e48SStefano Zampini /* final check for null pressures */ 24324f1b2e48SStefano Zampini if (zerodiag && pressures) { 24334f1b2e48SStefano Zampini PetscInt nz,np; 24344f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 24354f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 24364f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 24374f1b2e48SStefano Zampini } 24384f1b2e48SStefano Zampini 24394f1b2e48SStefano Zampini if (recompute_zerodiag) { 24404f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 24414f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 24424f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 24434f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 24444f1b2e48SStefano Zampini } else { 24454f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 24464f1b2e48SStefano Zampini 24474f1b2e48SStefano Zampini nzn = 0; 24484f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 24494f1b2e48SStefano Zampini PetscInt ns; 24504f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 24514f1b2e48SStefano Zampini nzn += ns; 24524f1b2e48SStefano Zampini } 24534f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 24544f1b2e48SStefano Zampini nzn = 0; 24554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 24564f1b2e48SStefano Zampini PetscInt ns,*idxs; 24574f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 24584f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 24594f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 24604f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 24614f1b2e48SStefano Zampini nzn += ns; 24624f1b2e48SStefano Zampini } 24634f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 24644f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 24654f1b2e48SStefano Zampini } 24664f1b2e48SStefano Zampini have_null = PETSC_FALSE; 24674f1b2e48SStefano Zampini } 24684f1b2e48SStefano Zampini 2469669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2470a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2471a198735bSStefano Zampini Mat A,loc_divudotp; 2472a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2473a198735bSStefano Zampini IS row,col,isused = NULL; 2474a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2475a198735bSStefano Zampini 24761f4df5f7SStefano Zampini if (pressures) { 24771f4df5f7SStefano Zampini isused = pressures; 24781f4df5f7SStefano Zampini } else { 24794edc6404Sstefano_zampini isused = zerodiag_save; 24801f4df5f7SStefano Zampini } 2481a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2482669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 24831ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 24841ae86dd6SStefano 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"); 2485a198735bSStefano Zampini n_isused = 0; 2486a198735bSStefano Zampini if (isused) { 2487a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2488a198735bSStefano Zampini } 2489a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2490a198735bSStefano Zampini st = st-n_isused; 24911ae86dd6SStefano Zampini if (n) { 2492a198735bSStefano Zampini const PetscInt *gidxs; 2493a198735bSStefano Zampini 24947dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2495a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2496a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2497a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2498a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2499a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 25001ae86dd6SStefano Zampini } else { 2501a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2502a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2503a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2504a198735bSStefano Zampini } 2505a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2506a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2507a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2508a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2509a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2510a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2511a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2512a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2513a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2514a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2515a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2516a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2517a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2518a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 25191ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 25201ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 25211ae86dd6SStefano Zampini } 25224edc6404Sstefano_zampini ierr = ISDestroy(&zerodiag_save);CHKERRQ(ierr); 2523b3afcdbeSStefano Zampini 2524b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 25254f1b2e48SStefano Zampini if (has_null_pressures) { 25264f1b2e48SStefano Zampini IS zerodiagc; 25274f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 25284f1b2e48SStefano Zampini PetscInt i,s,*nnz; 25294f1b2e48SStefano Zampini 25304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2531339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 2532339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 2533339f8db1SStefano Zampini /* local change of basis for pressures */ 2534339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 253597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2536339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2537339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2538339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 25394f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 25404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 25414f1b2e48SStefano Zampini PetscInt nzs,j; 25424f1b2e48SStefano Zampini 25434f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 25444f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 25454f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 25464f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 25474f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 25484f1b2e48SStefano Zampini } 2549339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2550339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2551339f8db1SStefano Zampini /* set identity on velocities */ 2552339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 2553339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 2554339f8db1SStefano Zampini } 25554f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 25564f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 25579f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 25584f1b2e48SStefano 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); 2559339f8db1SStefano Zampini /* set change on pressures */ 25604f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 25614f1b2e48SStefano Zampini PetscScalar *array; 25624f1b2e48SStefano Zampini PetscInt nzs; 25634f1b2e48SStefano Zampini 25644f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 25654f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 25664f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2567339f8db1SStefano Zampini PetscScalar vals[2]; 2568339f8db1SStefano Zampini PetscInt cols[2]; 2569339f8db1SStefano Zampini 2570339f8db1SStefano Zampini cols[0] = idxs[i]; 25714f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2572339f8db1SStefano Zampini vals[0] = 1.; 2573b0f5fe93SStefano Zampini vals[1] = 1.; 25744f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2575339f8db1SStefano Zampini } 25764f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 25774f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 25784f1b2e48SStefano Zampini array[nzs-1] = 1.; 25794f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 25804f1b2e48SStefano Zampini /* store local idxs for p0 */ 25814f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 25824f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2583339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 25844f1b2e48SStefano Zampini } 2585339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2586339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2587a3df083aSStefano Zampini /* project if needed */ 2588a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 25891dd7afcfSStefano Zampini Mat M; 25901dd7afcfSStefano Zampini 25911dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2592339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25931dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25941dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2595a3df083aSStefano Zampini } 25964f1b2e48SStefano Zampini /* store global idxs for p0 */ 25974f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2598339f8db1SStefano Zampini } 2599ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 26004f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 2601b0f5fe93SStefano Zampini 2602b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 2603b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 260427b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 260527b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2606339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2607339f8db1SStefano Zampini PetscFunctionReturn(0); 2608339f8db1SStefano Zampini } 2609339f8db1SStefano Zampini 2610015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 2611efc2fbd9SStefano Zampini { 2612efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2613de9d7bd0SStefano Zampini PetscScalar *array; 2614efc2fbd9SStefano Zampini PetscErrorCode ierr; 2615efc2fbd9SStefano Zampini 2616efc2fbd9SStefano Zampini PetscFunctionBegin; 2617efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 2618efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 26194f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2620efc2fbd9SStefano Zampini } 2621de9d7bd0SStefano Zampini if (get) { 2622efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 26234f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 26244f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 2625efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 2626de9d7bd0SStefano Zampini } else { 2627de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 2628de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2629de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2630de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 2631efc2fbd9SStefano Zampini } 2632efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2633efc2fbd9SStefano Zampini } 2634efc2fbd9SStefano Zampini 2635c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 2636c263805aSStefano Zampini { 2637c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2638c263805aSStefano Zampini PetscErrorCode ierr; 2639c263805aSStefano Zampini 2640c263805aSStefano Zampini PetscFunctionBegin; 2641c263805aSStefano Zampini /* TODO: add error checking 2642c263805aSStefano Zampini - avoid nested pop (or push) calls. 2643c263805aSStefano Zampini - cannot push before pop. 26441c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 2645c263805aSStefano Zampini */ 26464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2647efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2648efc2fbd9SStefano Zampini } 2649c263805aSStefano Zampini if (pop) { 2650a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 26514f1b2e48SStefano Zampini IS is_p0; 26524f1b2e48SStefano Zampini MatReuse reuse; 2653c263805aSStefano Zampini 2654c263805aSStefano Zampini /* extract B_0 */ 26554f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 26564f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 26574f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 26584f1b2e48SStefano Zampini } 26594f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 26607dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 2661c263805aSStefano Zampini /* remove rows and cols from local problem */ 2662c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 266397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 26644f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 26654f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 2666a3df083aSStefano Zampini } else { 2667a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2668a3df083aSStefano Zampini PetscScalar *vals; 2669a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 2670a3df083aSStefano Zampini 2671a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 2672a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 2673a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 26740b5adadeSStefano Zampini PetscInt *nnz; 2675a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 2676a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 2677a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2678331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 2679331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2680331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 2681331e053bSStefano Zampini nnz[i] = n - nnz[i]; 2682331e053bSStefano Zampini } 2683331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 2684331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2685331e053bSStefano Zampini } 2686a3df083aSStefano Zampini 2687a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2688a3df083aSStefano Zampini PetscScalar *array; 2689a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 2690a3df083aSStefano Zampini 2691a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 2692a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2693a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2694a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 2695a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2696a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 2697a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 2698a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 2699a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 2700a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 2701a3df083aSStefano Zampini cum = 0; 2702a3df083aSStefano Zampini for (j=0;j<n;j++) { 270322db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 2704a3df083aSStefano Zampini vals[cum] = array[j]; 2705a3df083aSStefano Zampini idxs_ins[cum] = j; 2706a3df083aSStefano Zampini cum++; 2707a3df083aSStefano Zampini } 2708a3df083aSStefano Zampini } 2709a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 2710a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 2711a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2712a3df083aSStefano Zampini } 2713a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2714a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2715a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 2716a3df083aSStefano Zampini } 2717c263805aSStefano Zampini } else { /* push */ 2718a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 27194f1b2e48SStefano Zampini PetscInt i; 27204f1b2e48SStefano Zampini 27214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 27224f1b2e48SStefano Zampini PetscScalar *B0_vals; 27234f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 27244f1b2e48SStefano Zampini 27254f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 27264f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 27277b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 27284f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 27294f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 27304f1b2e48SStefano Zampini } 2731c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2732c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2733a3df083aSStefano Zampini } else { 2734a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 2735a3df083aSStefano Zampini } 2736c263805aSStefano Zampini } 2737c263805aSStefano Zampini PetscFunctionReturn(0); 2738c263805aSStefano Zampini } 2739c263805aSStefano Zampini 274008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 2741b1b3d7a2SStefano Zampini { 2742b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 274308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 274408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 274508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 274608122e43SStefano Zampini PetscScalar *work,lwork; 274708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 274808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 274908122e43SStefano Zampini PetscReal *eigs,thresh; 27501b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 2751f6f667cfSStefano Zampini PetscBool allocated_S_St; 275208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 275308122e43SStefano Zampini PetscReal *rwork; 275408122e43SStefano Zampini #endif 2755b1b3d7a2SStefano Zampini PetscErrorCode ierr; 2756b1b3d7a2SStefano Zampini 2757b1b3d7a2SStefano Zampini PetscFunctionBegin; 2758b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 2759af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 2760d750e150Sstefano_zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\nRerun with -sub_schurs_hermitian 1 -sub_schurs_posdef 1 if the problem is SPD",sub_schurs->is_hermitian,sub_schurs->is_posdef); 276106a4e24aSStefano Zampini 2762fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2763fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2764fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2765fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 27661575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2767fd14bc51SStefano Zampini } 2768fd14bc51SStefano Zampini 2769e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 2770e496cd5dSStefano 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); 2771e496cd5dSStefano Zampini } 2772e496cd5dSStefano Zampini 277308122e43SStefano Zampini /* max size of subsets */ 277408122e43SStefano Zampini mss = 0; 277508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 277608122e43SStefano Zampini PetscInt subset_size; 2777862806e4SStefano Zampini 277808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 277908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 278008122e43SStefano Zampini } 278108122e43SStefano Zampini 278208122e43SStefano Zampini /* min/max and threshold */ 278308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 2784f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 278508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 2786f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 2787f6f667cfSStefano Zampini if (nmin) { 2788f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 2789f6f667cfSStefano Zampini } 279008122e43SStefano Zampini 279108122e43SStefano Zampini /* allocate lapack workspace */ 279208122e43SStefano Zampini cum = cum2 = 0; 279308122e43SStefano Zampini maxneigs = 0; 279408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 279508122e43SStefano Zampini PetscInt n,subset_size; 2796f6f667cfSStefano Zampini 279708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 279808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 27999162d606SStefano Zampini cum += subset_size; 28009162d606SStefano Zampini cum2 += subset_size*n; 280108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 280208122e43SStefano Zampini } 280308122e43SStefano Zampini if (mss) { 28049ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 280508122e43SStefano Zampini PetscBLASInt B_itype = 1; 280608122e43SStefano Zampini PetscBLASInt B_N = mss; 28074c6709b3SStefano Zampini PetscReal zero = 0.0; 28084c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 280908122e43SStefano Zampini 281008122e43SStefano Zampini B_lwork = -1; 281108122e43SStefano Zampini S = NULL; 281208122e43SStefano Zampini St = NULL; 2813a58a30b4SStefano Zampini eigs = NULL; 2814a58a30b4SStefano Zampini eigv = NULL; 2815a58a30b4SStefano Zampini B_iwork = NULL; 2816a58a30b4SStefano Zampini B_ifail = NULL; 2817d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2818d1710679SStefano Zampini rwork = NULL; 2819d1710679SStefano Zampini #endif 28208bec7fa6SStefano Zampini thresh = 1.0; 282108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 282208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 282308122e43SStefano 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)); 282408122e43SStefano Zampini #else 282508122e43SStefano 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)); 282608122e43SStefano Zampini #endif 282708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 282808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 282908122e43SStefano Zampini } else { 283008122e43SStefano Zampini /* TODO */ 283108122e43SStefano Zampini } 283208122e43SStefano Zampini } else { 283308122e43SStefano Zampini lwork = 0; 283408122e43SStefano Zampini } 283508122e43SStefano Zampini 283608122e43SStefano Zampini nv = 0; 2837d62866d3SStefano 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) */ 2838d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 283908122e43SStefano Zampini } 28404c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 2841f6f667cfSStefano Zampini if (allocated_S_St) { 2842f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 2843f6f667cfSStefano Zampini } 2844f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 284508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 284608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 284708122e43SStefano Zampini #endif 28489162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 28499162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 28509162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 285108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 28529162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 285308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 285408122e43SStefano Zampini 285508122e43SStefano Zampini maxneigs = 0; 285672b8c272SStefano Zampini cum = cumarray = 0; 28579162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 28589162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 2859d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 286008122e43SStefano Zampini const PetscInt *idxs; 286108122e43SStefano Zampini 2862d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 286308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 286408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 286508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 286608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 28679162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 28689162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 286908122e43SStefano Zampini } 2870d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 287108122e43SStefano Zampini } 287208122e43SStefano Zampini 287308122e43SStefano Zampini if (mss) { /* multilevel */ 287408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 287508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 287608122e43SStefano Zampini } 287708122e43SStefano Zampini 2878ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 287908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 288008122e43SStefano Zampini const PetscInt *idxs; 28819d54b7f4SStefano Zampini PetscReal upper,lower; 2882862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 288308122e43SStefano Zampini PetscBLASInt B_N; 2884aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 288508122e43SStefano Zampini 28869d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 28879d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 28889d54b7f4SStefano Zampini lower = thresh; 28899d54b7f4SStefano Zampini } else { 28909d54b7f4SStefano Zampini upper = 1./thresh; 28919d54b7f4SStefano Zampini lower = 0.; 28929d54b7f4SStefano Zampini } 2893862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 2894ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 2895f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 2896f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 28979ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 2898aff50787SStefano Zampini PetscInt j,k; 2899aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 2900aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 2901aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 290208122e43SStefano Zampini } 290308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 2904aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 2905aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 2906aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 2907aff50787SStefano Zampini } 290808122e43SStefano Zampini } 290908122e43SStefano Zampini } else { 291008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 291108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 291208122e43SStefano Zampini } 29138bec7fa6SStefano Zampini } else { 2914f6f667cfSStefano Zampini S = Sarray + cumarray; 2915f6f667cfSStefano Zampini St = Starray + cumarray; 29168bec7fa6SStefano Zampini } 2917aff50787SStefano Zampini /* see if we can save some work */ 2918b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 2919aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 2920aff50787SStefano Zampini } 2921aff50787SStefano Zampini 2922b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 2923aff50787SStefano Zampini B_neigs = 0; 2924aff50787SStefano Zampini } else { 29259ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 292608122e43SStefano Zampini PetscBLASInt B_itype = 1; 2927f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 29284c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 29299552c7c7SStefano Zampini PetscInt nmin_s; 2930b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 293108122e43SStefano Zampini 2932fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2933eee23b56SStefano 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]]); 2934fd14bc51SStefano Zampini } 2935d16cbb6bSStefano Zampini 2936b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 2937b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 2938b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 2939b7ab4a40SStefano Zampini } 2940b7ab4a40SStefano Zampini 294108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2942b7ab4a40SStefano Zampini if (compute_range) { 2943d16cbb6bSStefano Zampini 2944d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 294508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 29469d54b7f4SStefano 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)); 294708122e43SStefano Zampini #else 29489d54b7f4SStefano 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)); 294908122e43SStefano Zampini #endif 2950b7ab4a40SStefano Zampini } else if (!same_data) { 2951d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 2952d16cbb6bSStefano Zampini B_IL = 1; 2953d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 29549d54b7f4SStefano 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)); 2955d16cbb6bSStefano Zampini #else 29569d54b7f4SStefano 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)); 2957d16cbb6bSStefano Zampini #endif 2958b03ebc13SStefano Zampini } else { /* same_data is true, so just get the adaptive functional requested by the user */ 2959b7ab4a40SStefano Zampini PetscInt k; 2960b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 2961b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 2962b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 2963b7ab4a40SStefano Zampini nmin = nmax; 2964b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 2965b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 2966b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 2967b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 2968b7ab4a40SStefano Zampini } 2969d16cbb6bSStefano Zampini } 297008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 297108122e43SStefano Zampini if (B_ierr) { 29726c4ed002SBarry 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); 29736c4ed002SBarry 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); 29746c4ed002SBarry 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); 297508122e43SStefano Zampini } 297608122e43SStefano Zampini 297708122e43SStefano Zampini if (B_neigs > nmax) { 2978fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2979fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 2980fd14bc51SStefano Zampini } 29819d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 298208122e43SStefano Zampini B_neigs = nmax; 298308122e43SStefano Zampini } 298408122e43SStefano Zampini 29859552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 29869552c7c7SStefano Zampini if (B_neigs < nmin_s) { 298708122e43SStefano Zampini PetscBLASInt B_neigs2; 298808122e43SStefano Zampini 29899d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2990f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 29919d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 29929d54b7f4SStefano Zampini } else { 29939d54b7f4SStefano Zampini B_IL = B_neigs + 1; 29949d54b7f4SStefano Zampini B_IU = nmin_s; 29959d54b7f4SStefano Zampini } 2996fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2997fd14bc51SStefano 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); 2998fd14bc51SStefano Zampini } 29999ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 30001ae86dd6SStefano Zampini PetscInt j,k; 300108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 30021ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 30031ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 30041ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 300508122e43SStefano Zampini } 300608122e43SStefano Zampini } 300708122e43SStefano Zampini } else { 300808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 300908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 301008122e43SStefano Zampini } 301108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 301208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 30139d54b7f4SStefano 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)); 301408122e43SStefano Zampini #else 30159d54b7f4SStefano 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)); 301608122e43SStefano Zampini #endif 301708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 301808122e43SStefano Zampini B_neigs += B_neigs2; 301908122e43SStefano Zampini } 302008122e43SStefano Zampini if (B_ierr) { 30216c4ed002SBarry 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); 30226c4ed002SBarry 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); 30236c4ed002SBarry 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); 302408122e43SStefano Zampini } 3025fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3026ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 302708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 302808122e43SStefano Zampini if (eigs[j] == 0.0) { 3029ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 303008122e43SStefano Zampini } else { 30319d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3032ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 30339d54b7f4SStefano Zampini } else { 30349d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 30359d54b7f4SStefano Zampini } 3036fd14bc51SStefano Zampini } 303708122e43SStefano Zampini } 303808122e43SStefano Zampini } 303908122e43SStefano Zampini } else { 304008122e43SStefano Zampini /* TODO */ 304108122e43SStefano Zampini } 3042aff50787SStefano Zampini } 30436c3e6151SStefano Zampini /* change the basis back to the original one */ 30446c3e6151SStefano Zampini if (sub_schurs->change) { 304572b8c272SStefano Zampini Mat change,phi,phit; 30466c3e6151SStefano Zampini 30476c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 30486c3e6151SStefano Zampini PetscInt ii; 30496c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 30506c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 30516c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 3052684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3053684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 3054684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 3055684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3056684229deSStefano Zampini #else 30576c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 3058684229deSStefano Zampini #endif 30596c3e6151SStefano Zampini } 30606c3e6151SStefano Zampini } 30616c3e6151SStefano Zampini } 306272b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 30636c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 306472b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 30656c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 30666c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 30676c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 30686c3e6151SStefano Zampini } 30698bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 30708bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 30719162d606SStefano Zampini if (B_neigs) { 30729162d606SStefano 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); 3073fd14bc51SStefano Zampini 3074fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 30759552c7c7SStefano Zampini PetscInt ii; 30769552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 3077ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 30789552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 3079ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3080ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3081ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3082ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3083ac47001eSStefano Zampini #else 3084ac47001eSStefano 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); 3085ac47001eSStefano Zampini #endif 30869552c7c7SStefano Zampini } 30879552c7c7SStefano Zampini } 3088fd14bc51SStefano Zampini } 30899162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 30909162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 30919162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 30929162d606SStefano Zampini cum++; 309308122e43SStefano Zampini } 309408122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 309508122e43SStefano Zampini /* shift for next computation */ 309608122e43SStefano Zampini cumarray += subset_size*subset_size; 309708122e43SStefano Zampini } 3098fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3099fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3100fd14bc51SStefano Zampini } 310108122e43SStefano Zampini 310208122e43SStefano Zampini if (mss) { 310308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 310408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 3105f6f667cfSStefano Zampini /* destroy matrices (junk) */ 3106f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 3107f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 310808122e43SStefano Zampini } 3109f6f667cfSStefano Zampini if (allocated_S_St) { 3110f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 3111f6f667cfSStefano Zampini } 3112f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 311308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 311408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 311508122e43SStefano Zampini #endif 311608122e43SStefano Zampini if (pcbddc->dbg_flag) { 31171b968477SStefano Zampini PetscInt maxneigs_r; 3118b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 31199b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 312008122e43SStefano Zampini } 312108122e43SStefano Zampini PetscFunctionReturn(0); 312208122e43SStefano Zampini } 3123b1b3d7a2SStefano Zampini 3124c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 3125c8587f34SStefano Zampini { 31268629588bSStefano Zampini PetscScalar *coarse_submat_vals; 3127c8587f34SStefano Zampini PetscErrorCode ierr; 3128c8587f34SStefano Zampini 3129c8587f34SStefano Zampini PetscFunctionBegin; 3130f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 31315e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 3132c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 3133c8587f34SStefano Zampini 3134684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 31350fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 3136684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3137c8587f34SStefano Zampini 31388629588bSStefano Zampini /* 31398629588bSStefano Zampini Setup local correction and local part of coarse basis. 31408629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 31418629588bSStefano Zampini */ 314247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 31438629588bSStefano Zampini 31448629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 31458629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 31468629588bSStefano Zampini 31478629588bSStefano Zampini /* free */ 31488629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 3149c8587f34SStefano Zampini PetscFunctionReturn(0); 3150c8587f34SStefano Zampini } 3151c8587f34SStefano Zampini 3152674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 3153674ae819SStefano Zampini { 3154674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3155674ae819SStefano Zampini PetscErrorCode ierr; 3156674ae819SStefano Zampini 3157674ae819SStefano Zampini PetscFunctionBegin; 3158674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 315930368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3160674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 3161785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3162674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 3163f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3164f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3165785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 316663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 316763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 3168674ae819SStefano Zampini PetscFunctionReturn(0); 3169674ae819SStefano Zampini } 3170674ae819SStefano Zampini 3171674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 3172674ae819SStefano Zampini { 3173674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 31744f1b2e48SStefano Zampini PetscInt i; 3175674ae819SStefano Zampini PetscErrorCode ierr; 3176674ae819SStefano Zampini 3177674ae819SStefano Zampini PetscFunctionBegin; 31781e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 31791e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 3180a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 3181b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3182674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 318316909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 31841dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 3185674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3186669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 3187fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 31889326c5c6Sstefano_zampini ierr = PCBDDCGraphDestroy(&pcbddc->mat_graph);CHKERRQ(ierr); 31894f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 31904f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 31914f1b2e48SStefano Zampini } 3192e68a0315Sstefano_zampini pcbddc->n_local_subs = 0; 31934f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 3194e68a0315Sstefano_zampini ierr = PCBDDCSubSchursDestroy(&pcbddc->sub_schurs);CHKERRQ(ierr); 3195c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 31968af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 3197674ae819SStefano Zampini PetscFunctionReturn(0); 3198674ae819SStefano Zampini } 3199674ae819SStefano Zampini 3200674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 3201674ae819SStefano Zampini { 3202674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3203674ae819SStefano Zampini PetscErrorCode ierr; 3204674ae819SStefano Zampini 3205674ae819SStefano Zampini PetscFunctionBegin; 3206674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 320758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 3208ca92afb2SStefano Zampini PetscScalar *array; 320906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 321006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 321158da7f69SStefano Zampini } 3212674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3213674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 321415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 321515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3216674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 3217674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 3218674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 321906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 3220674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3221674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 32228ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 3223674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 3224674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 3225674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 32269326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_D);CHKERRQ(ierr); 32279326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 32289326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3229f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3230727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 32310e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3232f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 323370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 323481d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 32350369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 32361dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 32374f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 32388b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 3239ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 3240ca92afb2SStefano Zampini PetscInt i; 3241ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3242ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 3243ca92afb2SStefano Zampini } 3244ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 3245ca92afb2SStefano Zampini } 32464f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 3247674ae819SStefano Zampini PetscFunctionReturn(0); 3248674ae819SStefano Zampini } 3249674ae819SStefano Zampini 3250f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 32516bfb1811SStefano Zampini { 32526bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 32536bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 32546bfb1811SStefano Zampini VecType impVecType; 32554f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 32566bfb1811SStefano Zampini PetscErrorCode ierr; 32576bfb1811SStefano Zampini 32586bfb1811SStefano Zampini PetscFunctionBegin; 32594f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 3260b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 32616bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 3262e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 3263e7b262bdSStefano Zampini /* R nodes */ 3264e7b262bdSStefano Zampini old_size = -1; 3265e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 3266e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 3267e7b262bdSStefano Zampini } 3268e7b262bdSStefano Zampini if (n_R != old_size) { 3269e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3270e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 32716bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 32726bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 32736bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 32746bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 3275e7b262bdSStefano Zampini } 3276e7b262bdSStefano Zampini /* local primal dofs */ 3277e7b262bdSStefano Zampini old_size = -1; 3278e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 3279e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 3280e7b262bdSStefano Zampini } 3281e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 3282e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 328383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 3284e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 32856bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 3286e7b262bdSStefano Zampini } 3287e7b262bdSStefano Zampini /* local explicit constraints */ 3288e7b262bdSStefano Zampini old_size = -1; 3289e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 3290e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 3291e7b262bdSStefano Zampini } 3292e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 3293e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 329483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 329583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 329683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 329783b7ccabSStefano Zampini } 32986bfb1811SStefano Zampini PetscFunctionReturn(0); 32996bfb1811SStefano Zampini } 33006bfb1811SStefano Zampini 330147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 330288ebb749SStefano Zampini { 330325084f0cSStefano Zampini PetscErrorCode ierr; 330425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 330588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 330688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3307d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 330825084f0cSStefano Zampini /* submatrices of local problem */ 330980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 331006656605SStefano Zampini /* submatrices of local coarse problem */ 331106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 331225084f0cSStefano Zampini /* working matrices */ 331306656605SStefano Zampini Mat C_CR; 331425084f0cSStefano Zampini /* additional working stuff */ 331506656605SStefano Zampini PC pc_R; 3316*c58f9fdbSStefano Zampini Mat F,Brhs = NULL; 33175cbda25cSStefano Zampini Vec dummy_vec; 3318*c58f9fdbSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction,sparserhs; 331925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 332006656605SStefano Zampini PetscScalar *work; 332106656605SStefano Zampini PetscInt *idx_V_B; 3322ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 332306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 3324ffd830a3SStefano Zampini 332525084f0cSStefano Zampini /* some shortcuts to scalars */ 332606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 332788ebb749SStefano Zampini 332888ebb749SStefano Zampini PetscFunctionBegin; 33299a962809SStefano 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"); 3330ffd830a3SStefano Zampini 3331ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 3332b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 33334f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 3334b371cd4fSStefano Zampini n_B = pcis->n_B; 3335b371cd4fSStefano Zampini n_D = pcis->n - n_B; 333688ebb749SStefano Zampini n_R = pcis->n - n_vertices; 333788ebb749SStefano Zampini 333888ebb749SStefano Zampini /* vertices in boundary numbering */ 3339785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 33400e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 33416c4ed002SBarry 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); 334288ebb749SStefano Zampini 334306656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 3344019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 334506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 334606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 334706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 334806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 334906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 335006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 335106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 335206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 335306656605SStefano Zampini 335406656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 335506656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 335606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 335706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 335806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 3359ffd830a3SStefano Zampini lda_rhs = n_R; 3360a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 336106656605SStefano Zampini if (isLU || isILU || isCHOL) { 336206656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 3363b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 3364df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3365d62866d3SStefano Zampini MatFactorType type; 3366d62866d3SStefano Zampini 3367df4d28bfSStefano Zampini F = reuse_solver->F; 33686816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 3369d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 3370ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 337122db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 337206656605SStefano Zampini } else { 337306656605SStefano Zampini F = NULL; 337406656605SStefano Zampini } 337506656605SStefano Zampini 3376*c58f9fdbSStefano Zampini /* determine if we can use a sparse right-hand side */ 3377*c58f9fdbSStefano Zampini sparserhs = PETSC_FALSE; 3378*c58f9fdbSStefano Zampini if (F) { 3379*c58f9fdbSStefano Zampini const MatSolverPackage solver; 3380*c58f9fdbSStefano Zampini 3381*c58f9fdbSStefano Zampini ierr = MatFactorGetSolverPackage(F,&solver);CHKERRQ(ierr); 3382*c58f9fdbSStefano Zampini ierr = PetscStrcmp(solver,MATSOLVERMUMPS,&sparserhs);CHKERRQ(ierr); 3383*c58f9fdbSStefano Zampini } 3384*c58f9fdbSStefano Zampini 3385ffd830a3SStefano Zampini /* allocate workspace */ 3386ffd830a3SStefano Zampini n = 0; 3387ffd830a3SStefano Zampini if (n_constraints) { 3388ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 3389ffd830a3SStefano Zampini } 3390ffd830a3SStefano Zampini if (n_vertices) { 3391ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 3392ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 3393ffd830a3SStefano Zampini } 33942a3a6641Sstefano_zampini if (!pcbddc->symmetric_primal) { 33952a3a6641Sstefano_zampini n = PetscMax(2*lda_rhs*pcbddc->local_primal_size,n); 33962a3a6641Sstefano_zampini } 3397ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 3398ffd830a3SStefano Zampini 33995cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 34005cbda25cSStefano Zampini dummy_vec = NULL; 34015cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 34025cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 34035cbda25cSStefano Zampini } 34045cbda25cSStefano Zampini 340588ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 340688ebb749SStefano Zampini if (n_constraints) { 340772b8c272SStefano Zampini Mat M1,M2,M3,C_B; 340806656605SStefano Zampini IS is_aux; 340980677318SStefano Zampini PetscScalar *array,*array2; 341006656605SStefano Zampini 3411f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 341280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 341388ebb749SStefano Zampini 341425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 341525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 34167dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 34177dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 341888ebb749SStefano Zampini 341980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 342080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 3421*c58f9fdbSStefano Zampini if (!sparserhs) { 3422ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 342388ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 342406656605SStefano Zampini const PetscScalar *row_cmat_values; 342506656605SStefano Zampini const PetscInt *row_cmat_indices; 342606656605SStefano Zampini PetscInt size_of_constraint,j; 342788ebb749SStefano Zampini 342806656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 342906656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 3430ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 343106656605SStefano Zampini } 343206656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 343306656605SStefano Zampini } 3434*c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&Brhs);CHKERRQ(ierr); 3435*c58f9fdbSStefano Zampini } else { 3436*c58f9fdbSStefano Zampini Mat tC_CR; 3437*c58f9fdbSStefano Zampini 3438*c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 3439*c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 3440*c58f9fdbSStefano Zampini PetscScalar *aa; 3441*c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 3442*c58f9fdbSStefano Zampini PetscBool done; 3443*c58f9fdbSStefano Zampini 3444*c58f9fdbSStefano Zampini ierr = MatGetRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3445*c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed");CHKERRQ(ierr); 3446*c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(C_CR,&aa);CHKERRQ(ierr); 3447*c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_constraints,lda_rhs,ii,jj,aa,&tC_CR);CHKERRQ(ierr); 3448*c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3449*c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed");CHKERRQ(ierr); 3450*c58f9fdbSStefano Zampini } else { 3451*c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)C_CR);CHKERRQ(ierr); 3452*c58f9fdbSStefano Zampini tC_CR = C_CR; 3453*c58f9fdbSStefano Zampini } 3454*c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tC_CR,&Brhs);CHKERRQ(ierr); 3455*c58f9fdbSStefano Zampini ierr = MatDestroy(&tC_CR);CHKERRQ(ierr); 3456*c58f9fdbSStefano Zampini } 3457ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 345806656605SStefano Zampini if (F) { 3459a3df083aSStefano Zampini if (need_benign_correction) { 3460df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3461a3df083aSStefano Zampini 346272b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 346372b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 3464a3df083aSStefano Zampini } 3465*c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,local_auxmat2_R);CHKERRQ(ierr); 3466a3df083aSStefano Zampini if (need_benign_correction) { 3467a3df083aSStefano Zampini PetscScalar *marr; 3468df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3469a3df083aSStefano Zampini 3470a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 34715cbda25cSStefano Zampini if (lda_rhs != n_R) { 34725cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 34735cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 34745cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 34755cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 34765cbda25cSStefano Zampini } 34775cbda25cSStefano Zampini } else { 3478a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 3479a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 34805cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3481a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3482a3df083aSStefano Zampini } 34835cbda25cSStefano Zampini } 3484a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 3485a3df083aSStefano Zampini } 348606656605SStefano Zampini } else { 348780677318SStefano Zampini PetscScalar *marr; 348880677318SStefano Zampini 348980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 349006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 3491ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3492ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 349306656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 349406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 349506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 349606656605SStefano Zampini } 349780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 349806656605SStefano Zampini } 3499*c58f9fdbSStefano Zampini if (sparserhs) { 3500*c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 3501*c58f9fdbSStefano Zampini } 3502*c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 350380677318SStefano Zampini if (!pcbddc->switch_static) { 350480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 350580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 350680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 350780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 3508ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 350980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 351080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 351180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 351280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 351380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 351480677318SStefano Zampini } 351580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 351680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 351772b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 351880677318SStefano Zampini } else { 3519ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3520ffd830a3SStefano Zampini IS dummy; 3521ffd830a3SStefano Zampini 3522ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 35237dae84e0SHong Zhang ierr = MatCreateSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 3524ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 3525ffd830a3SStefano Zampini } else { 352680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 352780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 3528ffd830a3SStefano Zampini } 352925084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 353080677318SStefano Zampini } 353180677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 353280677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 353380677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 353406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 353506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 353680677318SStefano Zampini if (isCHOL) { 353780677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 353880677318SStefano Zampini } else { 353925084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 354080677318SStefano Zampini } 354180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 354206656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 354325084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 354425084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 354525084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 354680677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 354772b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 354872b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 354906656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 355006656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 3551f4ddd8eeSStefano Zampini } 3552fc227af8SStefano Zampini 3553fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 355488ebb749SStefano Zampini if (n_vertices) { 355506656605SStefano Zampini IS is_aux; 3556*c58f9fdbSStefano Zampini PetscBool isseqaij; 35573a50541eSStefano Zampini 3558b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 35596816873aSStefano Zampini IS tis; 35606816873aSStefano Zampini 35616816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 35626816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 35636816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 35646816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 35656816873aSStefano Zampini } else { 35663a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 35676816873aSStefano Zampini } 35687dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 35697dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3570*c58f9fdbSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3571*c58f9fdbSStefano Zampini if (!isseqaij) { /* MatMatMult(A_VR,A_RRmA_RV) below will raise an error */ 3572*c58f9fdbSStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 3573*c58f9fdbSStefano Zampini } 35747dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 357525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 357688ebb749SStefano Zampini } 357788ebb749SStefano Zampini 357888ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 3579f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 358006656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 358106656605SStefano Zampini if (pcbddc->coarse_phi_D) { 358206656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 358306656605SStefano Zampini } 3584f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 358506656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 358606656605SStefano Zampini PetscScalar *marray; 358706656605SStefano Zampini 358806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 358906656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 3590f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3591f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3592f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3593f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3594f4ddd8eeSStefano Zampini } 3595f4ddd8eeSStefano Zampini } 359606656605SStefano Zampini 3597f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 3598a6e023c1Sstefano_zampini PetscScalar *marr; 359988ebb749SStefano Zampini 3600a6e023c1Sstefano_zampini /* memory size */ 360106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 3602a6e023c1Sstefano_zampini if (pcbddc->switch_static || pcbddc->dbg_flag) n += n_D*pcbddc->local_primal_size; 3603a6e023c1Sstefano_zampini if (!pcbddc->symmetric_primal) n *= 2; 3604a6e023c1Sstefano_zampini ierr = PetscCalloc1(n,&marr);CHKERRQ(ierr); 3605a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3606a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 36078eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3608a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3609a6e023c1Sstefano_zampini marr += n_D*pcbddc->local_primal_size; 361088ebb749SStefano Zampini } 36113301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 3612a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3613a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 36148eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3615a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 361688ebb749SStefano Zampini } 361788ebb749SStefano Zampini } else { 3618c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 3619c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 36201b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3621c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 3622c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 3623c0553b1fSStefano Zampini } 362488ebb749SStefano Zampini } 362506656605SStefano Zampini } 3626019a44ceSStefano Zampini 362706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 36284f1b2e48SStefano Zampini p0_lidx_I = NULL; 36294f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 3630d12edf2fSStefano Zampini const PetscInt *idxs; 3631d12edf2fSStefano Zampini 3632d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 36334f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 36344f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 36354f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 36364f1b2e48SStefano Zampini } 3637d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 3638d12edf2fSStefano Zampini } 3639d16cbb6bSStefano Zampini 364006656605SStefano Zampini /* vertices */ 364106656605SStefano Zampini if (n_vertices) { 3642*c58f9fdbSStefano Zampini PetscBool restoreavr = PETSC_FALSE; 364316f15bc4SStefano Zampini 3644af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 364504708bb6SStefano Zampini 364616f15bc4SStefano Zampini if (n_R) { 364714393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 364806656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 364916f15bc4SStefano Zampini PetscScalar *x,*y; 365006656605SStefano Zampini 365121eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 365214393ed6SStefano Zampini if (need_benign_correction) { 365314393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 365414393ed6SStefano Zampini IS is_p0; 365514393ed6SStefano Zampini PetscInt *idxs_p0,n; 365614393ed6SStefano Zampini 365714393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 365814393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 365914393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 3660af25d912SStefano 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); 366114393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 366214393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 36637dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 366414393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 366514393ed6SStefano Zampini } 366614393ed6SStefano Zampini 3667*c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3668*c58f9fdbSStefano Zampini if (!sparserhs || need_benign_correction) { 3669ffd830a3SStefano Zampini if (lda_rhs == n_R) { 3670af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3671ffd830a3SStefano Zampini } else { 3672ca92afb2SStefano Zampini PetscScalar *av,*array; 3673ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 3674ca92afb2SStefano Zampini PetscInt n; 3675ca92afb2SStefano Zampini PetscBool flg_row; 3676ffd830a3SStefano Zampini 3677ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 3678ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 36799d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3680ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3681ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 3682ca92afb2SStefano Zampini for (i=0;i<n;i++) { 3683ca92afb2SStefano Zampini PetscInt j; 3684ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 3685ffd830a3SStefano Zampini } 3686ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3687ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3688ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 3689ffd830a3SStefano Zampini } 3690a3df083aSStefano Zampini if (need_benign_correction) { 3691df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3692a3df083aSStefano Zampini PetscScalar *marr; 3693a3df083aSStefano Zampini 3694a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 369514393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 369614393ed6SStefano Zampini 369714393ed6SStefano Zampini | 0 0 0 | (V) 369814393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 369914393ed6SStefano Zampini | 0 0 -1 | (p0) 370014393ed6SStefano Zampini 370114393ed6SStefano Zampini */ 3702df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 370314393ed6SStefano Zampini const PetscScalar *vals; 370414393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 370514393ed6SStefano Zampini PetscInt n,j,nz; 370614393ed6SStefano Zampini 3707df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3708df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 370914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 371014393ed6SStefano Zampini for (j=0;j<n;j++) { 371114393ed6SStefano Zampini PetscScalar val = vals[j]; 371214393ed6SStefano Zampini PetscInt k,col = idxs[j]; 371314393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 371414393ed6SStefano Zampini } 371514393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3716df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 371714393ed6SStefano Zampini } 371872b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 371972b8c272SStefano Zampini } 3720*c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RV);CHKERRQ(ierr); 3721*c58f9fdbSStefano Zampini Brhs = A_RV; 3722*c58f9fdbSStefano Zampini } else { 3723*c58f9fdbSStefano Zampini Mat tA_RVT,A_RVT; 3724*c58f9fdbSStefano Zampini 3725*c58f9fdbSStefano Zampini if (!pcbddc->symmetric_primal) { 3726*c58f9fdbSStefano Zampini ierr = MatTranspose(A_RV,MAT_INITIAL_MATRIX,&A_RVT);CHKERRQ(ierr); 3727*c58f9fdbSStefano Zampini } else { 3728*c58f9fdbSStefano Zampini restoreavr = PETSC_TRUE; 3729*c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 3730*c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_VR);CHKERRQ(ierr); 3731*c58f9fdbSStefano Zampini A_RVT = A_VR; 3732*c58f9fdbSStefano Zampini } 3733*c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 3734*c58f9fdbSStefano Zampini PetscScalar *aa; 3735*c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 3736*c58f9fdbSStefano Zampini PetscBool done; 3737*c58f9fdbSStefano Zampini 3738*c58f9fdbSStefano Zampini ierr = MatGetRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3739*c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed");CHKERRQ(ierr); 3740*c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(A_RVT,&aa);CHKERRQ(ierr); 3741*c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_vertices,lda_rhs,ii,jj,aa,&tA_RVT);CHKERRQ(ierr); 3742*c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3743*c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed");CHKERRQ(ierr); 3744*c58f9fdbSStefano Zampini } else { 3745*c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RVT);CHKERRQ(ierr); 3746*c58f9fdbSStefano Zampini tA_RVT = A_RVT; 3747*c58f9fdbSStefano Zampini } 3748*c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tA_RVT,&Brhs);CHKERRQ(ierr); 3749*c58f9fdbSStefano Zampini ierr = MatDestroy(&tA_RVT);CHKERRQ(ierr); 3750*c58f9fdbSStefano Zampini ierr = MatDestroy(&A_RVT);CHKERRQ(ierr); 3751*c58f9fdbSStefano Zampini } 375272b8c272SStefano Zampini if (F) { 375314393ed6SStefano Zampini /* need to correct the rhs */ 375472b8c272SStefano Zampini if (need_benign_correction) { 375572b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 375672b8c272SStefano Zampini PetscScalar *marr; 375772b8c272SStefano Zampini 3758*c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&marr);CHKERRQ(ierr); 37595cbda25cSStefano Zampini if (lda_rhs != n_R) { 37605cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 37615cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 37625cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 37635cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 37645cbda25cSStefano Zampini } 37655cbda25cSStefano Zampini } else { 3766a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3767a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 37685cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3769a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3770a3df083aSStefano Zampini } 37715cbda25cSStefano Zampini } 3772*c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&marr);CHKERRQ(ierr); 3773a3df083aSStefano Zampini } 3774*c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,A_RRmA_RV);CHKERRQ(ierr); 3775*c58f9fdbSStefano Zampini if (restoreavr) { 3776*c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 3777*c58f9fdbSStefano Zampini } 377814393ed6SStefano Zampini /* need to correct the solution */ 3779a3df083aSStefano Zampini if (need_benign_correction) { 3780df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3781a3df083aSStefano Zampini PetscScalar *marr; 3782a3df083aSStefano Zampini 3783a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 37845cbda25cSStefano Zampini if (lda_rhs != n_R) { 37855cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 37865cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 37875cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 37885cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 37895cbda25cSStefano Zampini } 37905cbda25cSStefano Zampini } else { 3791a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3792a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 37935cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3794a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3795a3df083aSStefano Zampini } 37965cbda25cSStefano Zampini } 3797a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 3798a3df083aSStefano Zampini } 379906656605SStefano Zampini } else { 3800*c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&y);CHKERRQ(ierr); 380106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 3802ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 3803ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 380406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 380506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 380606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 380706656605SStefano Zampini } 3808*c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&y);CHKERRQ(ierr); 380906656605SStefano Zampini } 381080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3811*c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 3812ffd830a3SStefano Zampini /* S_VV and S_CV */ 381306656605SStefano Zampini if (n_constraints) { 381406656605SStefano Zampini Mat B; 381580677318SStefano Zampini 3816ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 381780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 3818ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3819ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 382080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 382180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 382280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 382380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 382480677318SStefano Zampini } 3825ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 382680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 382780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3828ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 382980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 383006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 3831ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 3832ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 383306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 383406656605SStefano Zampini } 3835ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3836ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 3837ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3838ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 3839ffd830a3SStefano Zampini } 384006656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 384114393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 384214393ed6SStefano Zampini if (need_benign_correction) { 3843df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 384414393ed6SStefano Zampini PetscScalar *marr,*sums; 384514393ed6SStefano Zampini 384614393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 3847f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 3848df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 384914393ed6SStefano Zampini const PetscScalar *vals; 385014393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 385114393ed6SStefano Zampini PetscInt n,j,nz; 385214393ed6SStefano Zampini 3853df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3854df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 385514393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 385614393ed6SStefano Zampini PetscInt k; 385714393ed6SStefano Zampini sums[j] = 0.; 385814393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 385914393ed6SStefano Zampini } 386014393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 386114393ed6SStefano Zampini for (j=0;j<n;j++) { 386214393ed6SStefano Zampini PetscScalar val = vals[j]; 386314393ed6SStefano Zampini PetscInt k; 386414393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 386514393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 386614393ed6SStefano Zampini } 386714393ed6SStefano Zampini } 386814393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3869df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 387014393ed6SStefano Zampini } 387114393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 3872f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 387314393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 387414393ed6SStefano Zampini } 387580677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 387606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 387706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 387806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 387906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 388006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 388106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 388206656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3883d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 3884019a44ceSStefano Zampini } else { 3885d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3886d16cbb6bSStefano Zampini } 388721eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 3888d16cbb6bSStefano Zampini 388906656605SStefano Zampini /* coarse basis functions */ 389006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 389116f15bc4SStefano Zampini PetscScalar *y; 389216f15bc4SStefano Zampini 3893ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 389406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 389506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 389606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 389706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 389806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 389906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 390006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 390106656605SStefano Zampini 390206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 39034f1b2e48SStefano Zampini PetscInt j; 39044f1b2e48SStefano Zampini 390506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 390606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 390706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 390806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 390906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 39104f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 391106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 391206656605SStefano Zampini } 391306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 391406656605SStefano Zampini } 391504708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 391604708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 391706656605SStefano Zampini } 39185cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 391906656605SStefano Zampini 392006656605SStefano Zampini if (n_constraints) { 392106656605SStefano Zampini Mat B; 392206656605SStefano Zampini 3923ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 392406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 392580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 392606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 392706656605SStefano Zampini if (n_vertices) { 392880677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 392980677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 393080677318SStefano Zampini } else { 393180677318SStefano Zampini Mat S_VCt; 393280677318SStefano Zampini 3933ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3934ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 393572b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 3936ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 3937ffd830a3SStefano Zampini } 393880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 393980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 394080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 394180677318SStefano Zampini } 394206656605SStefano Zampini } 394306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 394406656605SStefano Zampini /* coarse basis functions */ 394506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 394606656605SStefano Zampini PetscScalar *y; 394706656605SStefano Zampini 3948ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 394906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 395006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 395106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 395206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 395306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 395406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 395506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 39564f1b2e48SStefano Zampini PetscInt j; 39574f1b2e48SStefano Zampini 395806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 395906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 396006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 396106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 396206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 39634f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 396406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 396506656605SStefano Zampini } 396606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 396706656605SStefano Zampini } 396806656605SStefano Zampini } 396980677318SStefano Zampini if (n_constraints) { 397080677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 397180677318SStefano Zampini } 39724f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 397372b8c272SStefano Zampini 397472b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 397572b8c272SStefano Zampini if (pcbddc->benign_n) { 397672b8c272SStefano Zampini Mat B0_B,B0_BPHI; 397772b8c272SStefano Zampini IS is_dummy; 397872b8c272SStefano Zampini PetscScalar *data; 397972b8c272SStefano Zampini PetscInt j; 398072b8c272SStefano Zampini 398172b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 39827dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 398372b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 398472b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 398586c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 398672b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 398772b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 398872b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 398972b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 399072b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 399172b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 399272b8c272SStefano Zampini } 399372b8c272SStefano Zampini } 399472b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 399572b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 399672b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 399772b8c272SStefano Zampini } 3998019a44ceSStefano Zampini 399906656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 40003301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 4001ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 4002ffd830a3SStefano Zampini PetscScalar *marray; 400306656605SStefano Zampini 400406656605SStefano Zampini if (n_constraints) { 4005ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 400606656605SStefano Zampini 4007abc8f43dSstefano_zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 400806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 4009ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 401016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 401106656605SStefano Zampini if (n_vertices) { 4012ffd830a3SStefano Zampini Mat S_VCT; 401306656605SStefano Zampini 401406656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 4015ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 401616f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 401706656605SStefano Zampini } 4018ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 40195b782168SStefano Zampini } else { 40205b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 402106656605SStefano Zampini } 402216f15bc4SStefano Zampini if (n_vertices && n_R) { 4023ffd830a3SStefano Zampini PetscScalar *av,*marray; 4024ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 4025ffd830a3SStefano Zampini PetscInt n; 4026ffd830a3SStefano Zampini PetscBool flg_row; 402706656605SStefano Zampini 4028ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 4029af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 4030ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4031ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 4032ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4033ffd830a3SStefano Zampini for (i=0;i<n;i++) { 4034ffd830a3SStefano Zampini PetscInt j; 4035ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 4036ffd830a3SStefano Zampini } 4037ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4038ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4039ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 404006656605SStefano Zampini } 404106656605SStefano Zampini 4042ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 4043abc8f43dSstefano_zampini if (n_vertices) { 4044ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4045ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 4046ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 4047ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 404806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 404906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 405006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 405106656605SStefano Zampini } 4052ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4053abc8f43dSstefano_zampini } 40545b782168SStefano Zampini if (B_C) { 4055ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 4056ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 4057ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 4058ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 4059ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4060ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4061ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 406206656605SStefano Zampini } 4063ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 40645b782168SStefano Zampini } 406506656605SStefano Zampini /* coarse basis functions */ 406606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 406706656605SStefano Zampini PetscScalar *y; 406806656605SStefano Zampini 4069ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 407006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 407106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 407206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 407306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 407406656605SStefano Zampini if (i<n_vertices) { 407506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 407606656605SStefano Zampini } 407706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 407806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 407906656605SStefano Zampini 408006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 408106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 408206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 408306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 408606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 408706656605SStefano Zampini } 408806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 408906656605SStefano Zampini } 4090ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 4091ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 409206656605SStefano Zampini } 4093a6e023c1Sstefano_zampini 4094d62866d3SStefano Zampini /* free memory */ 409588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 409606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 409706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 409806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 409906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 4100d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4101d62866d3SStefano Zampini if (n_vertices) { 4102d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 4103d62866d3SStefano Zampini } 4104d62866d3SStefano Zampini if (n_constraints) { 4105d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 4106d62866d3SStefano Zampini } 410788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 410888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 410988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 4110d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 411188ebb749SStefano Zampini Mat coarse_sub_mat; 411225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 411388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 411488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 411588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 41168bec7fa6SStefano Zampini Mat C_B,CPHI; 41178bec7fa6SStefano Zampini IS is_dummy; 41188bec7fa6SStefano Zampini Vec mones; 411988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 412088ebb749SStefano Zampini PetscReal real_value; 412188ebb749SStefano Zampini 4122a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 4123a3df083aSStefano Zampini Mat A; 4124a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 41257dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 41267dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 41277dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 41287dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4129a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 4130a3df083aSStefano Zampini } else { 413188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 413288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 413388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 413488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4135a3df083aSStefano Zampini } 413688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 413788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 4138ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 413988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 414088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 414188ebb749SStefano Zampini } 414288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 414388ebb749SStefano Zampini 414425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41453301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 414625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4147ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 414888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 414988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 415088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 415188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 415288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 415388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 415488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 415588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 415688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 415788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 415888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 415988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 416088ebb749SStefano Zampini } else { 416188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 416288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 416388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 416488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 416588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 416688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 416788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 416888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 416988ebb749SStefano Zampini } 417088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 417188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 417288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 4173511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 41744f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4175fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 4176d12edf2fSStefano Zampini PetscScalar *data,*data2; 41774f1b2e48SStefano Zampini PetscInt j; 4178d12edf2fSStefano Zampini 41794f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 41807dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 4181d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 418286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 4183d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 4184d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 41854f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 41864f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 4187d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 41884f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 41894f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 41904f1b2e48SStefano Zampini } 4191d12edf2fSStefano Zampini } 4192d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 4193d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 4194d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 4195d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 4196d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 4197d12edf2fSStefano Zampini } 4198d12edf2fSStefano Zampini #if 0 4199d12edf2fSStefano Zampini { 4200d12edf2fSStefano Zampini PetscViewer viewer; 4201d12edf2fSStefano Zampini char filename[256]; 4202ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4203d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4204d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4205ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 4206ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 4207ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 4208d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 420972b8c272SStefano Zampini if (save_change) { 421072b8c272SStefano Zampini Mat phi_B; 421172b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 421272b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 421372b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 421472b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 421572b8c272SStefano Zampini } else { 4216ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 4217ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 421872b8c272SStefano Zampini } 4219ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 4220ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 4221ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 4222ffd830a3SStefano Zampini } 4223ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 4224ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 4225ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 4226ffd830a3SStefano Zampini } 422772b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 4228ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 4229ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 4230ffd830a3SStefano Zampini } 4231d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4232d12edf2fSStefano Zampini } 4233d12edf2fSStefano Zampini #endif 423481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 42358bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 42361575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 423706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 42388bec7fa6SStefano Zampini 42398bec7fa6SStefano Zampini /* check constraints */ 4240a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 42417dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 42424f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 42438bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4244a00504b5SStefano Zampini } else { 4245a00504b5SStefano Zampini PetscScalar *data; 4246a00504b5SStefano Zampini Mat tmat; 4247a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4248a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 4249a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4250a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4251a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4252a00504b5SStefano Zampini } 42538bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 42548bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 42558bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 42568bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4257bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 4258ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 4259bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4260bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 4261bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 4262bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4263bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 426488ebb749SStefano Zampini } 42658bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 42668bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 42678bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 42688bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 426925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 427088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 427188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 427288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 427388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 427488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 427588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 427688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 427788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 427888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 427988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 4280ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 428188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 428288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 428388ebb749SStefano Zampini } 428488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 428588ebb749SStefano Zampini } 42868629588bSStefano Zampini /* get back data */ 42878629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 428888ebb749SStefano Zampini PetscFunctionReturn(0); 428988ebb749SStefano Zampini } 429088ebb749SStefano Zampini 42917dae84e0SHong Zhang PetscErrorCode MatCreateSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 4292aa0d41d4SStefano Zampini { 4293d65f70fdSStefano Zampini Mat *work_mat; 4294d65f70fdSStefano Zampini IS isrow_s,iscol_s; 4295d65f70fdSStefano Zampini PetscBool rsorted,csorted; 4296c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 4297aa0d41d4SStefano Zampini PetscErrorCode ierr; 4298aa0d41d4SStefano Zampini 4299aa0d41d4SStefano Zampini PetscFunctionBegin; 4300d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 4301d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 4302d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 4303d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 4304aa0d41d4SStefano Zampini 4305d65f70fdSStefano Zampini if (!rsorted) { 4306906d46d4SStefano Zampini const PetscInt *idxs; 4307906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 4308aa0d41d4SStefano Zampini 4309d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 4310d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 4311d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4312d65f70fdSStefano Zampini idxs_perm_r[i] = i; 4313aa0d41d4SStefano Zampini } 4314d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 4315d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 4316d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4317d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 4318aa0d41d4SStefano Zampini } 4319d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 4320d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 4321d65f70fdSStefano Zampini } else { 4322d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 4323d65f70fdSStefano Zampini isrow_s = isrow; 4324aa0d41d4SStefano Zampini } 4325906d46d4SStefano Zampini 4326d65f70fdSStefano Zampini if (!csorted) { 4327d65f70fdSStefano Zampini if (isrow == iscol) { 4328d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 4329d65f70fdSStefano Zampini iscol_s = isrow_s; 4330d65f70fdSStefano Zampini } else { 4331d65f70fdSStefano Zampini const PetscInt *idxs; 4332d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 4333906d46d4SStefano Zampini 4334d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 4335d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 4336d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4337d65f70fdSStefano Zampini idxs_perm_c[i] = i; 4338d65f70fdSStefano Zampini } 4339d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 4340d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 4341d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4342d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 4343d65f70fdSStefano Zampini } 4344d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 4345d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 4346d65f70fdSStefano Zampini } 4347d65f70fdSStefano Zampini } else { 4348d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 4349d65f70fdSStefano Zampini iscol_s = iscol; 4350d65f70fdSStefano Zampini } 4351d65f70fdSStefano Zampini 43527dae84e0SHong Zhang ierr = MatCreateSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4353d65f70fdSStefano Zampini 4354d65f70fdSStefano Zampini if (!rsorted || !csorted) { 4355906d46d4SStefano Zampini Mat new_mat; 4356d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 4357906d46d4SStefano Zampini 4358d65f70fdSStefano Zampini if (!rsorted) { 4359d65f70fdSStefano Zampini PetscInt *idxs_r,i; 4360d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 4361d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4362d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 4363906d46d4SStefano Zampini } 4364d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 4365d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 4366d65f70fdSStefano Zampini } else { 4367d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 4368906d46d4SStefano Zampini } 4369d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 4370d65f70fdSStefano Zampini 4371d65f70fdSStefano Zampini if (!csorted) { 4372d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 4373d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 4374d65f70fdSStefano Zampini is_perm_c = is_perm_r; 4375d65f70fdSStefano Zampini } else { 4376d65f70fdSStefano Zampini PetscInt *idxs_c,i; 4377f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 4378d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 4379d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4380d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 4381d65f70fdSStefano Zampini } 4382d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 4383d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 4384d65f70fdSStefano Zampini } 4385d65f70fdSStefano Zampini } else { 4386d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 4387d65f70fdSStefano Zampini } 4388d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 4389d65f70fdSStefano Zampini 4390d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 4391d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 4392d65f70fdSStefano Zampini work_mat[0] = new_mat; 4393d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 4394d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 4395d65f70fdSStefano Zampini } 4396d65f70fdSStefano Zampini 4397d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 4398d65f70fdSStefano Zampini *B = work_mat[0]; 4399d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 4400d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 4401d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 4402d65f70fdSStefano Zampini PetscFunctionReturn(0); 4403d65f70fdSStefano Zampini } 4404d65f70fdSStefano Zampini 44055e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 4406aa0d41d4SStefano Zampini { 4407aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 44085e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4409022d8d2bSstefano_zampini Mat new_mat,lA; 44105e8657edSStefano Zampini IS is_local,is_global; 4411d65f70fdSStefano Zampini PetscInt local_size; 4412d65f70fdSStefano Zampini PetscBool isseqaij; 4413aa0d41d4SStefano Zampini PetscErrorCode ierr; 4414aa0d41d4SStefano Zampini 4415aa0d41d4SStefano Zampini PetscFunctionBegin; 4416aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 44175e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 44185e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 4419b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 4420aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 44217dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 4422aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 4423906d46d4SStefano Zampini 4424906d46d4SStefano Zampini /* check */ 4425906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4426906d46d4SStefano Zampini Vec x,x_change; 4427906d46d4SStefano Zampini PetscReal error; 4428906d46d4SStefano Zampini 44295e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 4430906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 44315e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 4432e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4433e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4434d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 443588428137SStefano Zampini if (!pcbddc->change_interior) { 443688428137SStefano Zampini const PetscScalar *x,*y,*v; 443788428137SStefano Zampini PetscReal lerror = 0.; 443888428137SStefano Zampini PetscInt i; 443988428137SStefano Zampini 444088428137SStefano Zampini ierr = VecGetArrayRead(matis->x,&x);CHKERRQ(ierr); 444188428137SStefano Zampini ierr = VecGetArrayRead(matis->y,&y);CHKERRQ(ierr); 444288428137SStefano Zampini ierr = VecGetArrayRead(matis->counter,&v);CHKERRQ(ierr); 444388428137SStefano Zampini for (i=0;i<local_size;i++) 444488428137SStefano Zampini if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror) 444588428137SStefano Zampini lerror = PetscAbsScalar(x[i]-y[i]); 444688428137SStefano Zampini ierr = VecRestoreArrayRead(matis->x,&x);CHKERRQ(ierr); 444788428137SStefano Zampini ierr = VecRestoreArrayRead(matis->y,&y);CHKERRQ(ierr); 444888428137SStefano Zampini ierr = VecRestoreArrayRead(matis->counter,&v);CHKERRQ(ierr); 444988428137SStefano Zampini ierr = MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4450637e8532SStefano Zampini if (error > PETSC_SMALL) { 4451637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4452637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on I: %1.6e\n",error); 4453637e8532SStefano Zampini } else { 4454637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on I: %1.6e\n",error); 4455637e8532SStefano Zampini } 4456637e8532SStefano Zampini } 445788428137SStefano Zampini } 4458e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4459e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4460906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4461906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4462637e8532SStefano Zampini if (error > PETSC_SMALL) { 4463637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4464637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 4465637e8532SStefano Zampini } else { 4466637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on N: %1.6e\n",error); 4467637e8532SStefano Zampini } 4468637e8532SStefano Zampini } 4469906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4470906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4471906d46d4SStefano Zampini } 4472906d46d4SStefano Zampini 4473022d8d2bSstefano_zampini /* lA is present if we are setting up an inner BDDC for a saddle point FETI-DP */ 4474022d8d2bSstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject*)&lA);CHKERRQ(ierr); 4475022d8d2bSstefano_zampini 447622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 44779b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 447822d5777bSStefano Zampini if (isseqaij) { 4479a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4480a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 4481022d8d2bSstefano_zampini if (lA) { 4482022d8d2bSstefano_zampini Mat work; 4483022d8d2bSstefano_zampini ierr = MatPtAP(lA,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 4484022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 4485022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 4486022d8d2bSstefano_zampini } 4487aa0d41d4SStefano Zampini } else { 4488a00504b5SStefano Zampini Mat work_mat; 44891cf9b237SStefano Zampini 4490a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4491aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4492a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 44931d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 4494022d8d2bSstefano_zampini if (lA) { 4495022d8d2bSstefano_zampini Mat work; 4496022d8d2bSstefano_zampini ierr = MatConvert(lA,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4497022d8d2bSstefano_zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 4498022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 4499022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 4500022d8d2bSstefano_zampini } 4501aa0d41d4SStefano Zampini } 45023301b35fSStefano Zampini if (matis->A->symmetric_set) { 45033301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 4504e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 45053301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 4506e496cd5dSStefano Zampini #endif 45073301b35fSStefano Zampini } 4508d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 4509aa0d41d4SStefano Zampini PetscFunctionReturn(0); 4510aa0d41d4SStefano Zampini } 4511aa0d41d4SStefano Zampini 45128ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 4513a64d13efSStefano Zampini { 4514a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4515a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4516d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 451753892102SStefano Zampini PetscInt *idx_R_local=NULL; 45183a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 45193a50541eSStefano Zampini PetscInt vbs,bs; 45206816873aSStefano Zampini PetscBT bitmask=NULL; 4521a64d13efSStefano Zampini PetscErrorCode ierr; 4522a64d13efSStefano Zampini 4523a64d13efSStefano Zampini PetscFunctionBegin; 4524b23d619eSStefano Zampini /* 4525b23d619eSStefano Zampini No need to setup local scatters if 4526b23d619eSStefano Zampini - primal space is unchanged 4527b23d619eSStefano Zampini AND 4528b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 4529b23d619eSStefano Zampini AND 4530b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 4531b23d619eSStefano Zampini */ 4532b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 4533f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 4534f4ddd8eeSStefano Zampini } 4535f4ddd8eeSStefano Zampini /* destroy old objects */ 4536f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 4537f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 4538f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 4539a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 4540b371cd4fSStefano Zampini n_B = pcis->n_B; 4541b371cd4fSStefano Zampini n_D = pcis->n - n_B; 4542b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 45433a50541eSStefano Zampini 4544a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 45456816873aSStefano Zampini 454653892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 4547b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4548854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 4549a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 4550a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 45510e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 4552a64d13efSStefano Zampini } 4553a64d13efSStefano Zampini 4554a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 45554641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 45566816873aSStefano Zampini idx_R_local[n_R++] = i; 4557a64d13efSStefano Zampini } 4558a64d13efSStefano Zampini } 4559df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 4560df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 45616816873aSStefano Zampini 4562df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4563df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 45646816873aSStefano Zampini } 45653a50541eSStefano Zampini 45663a50541eSStefano Zampini /* Block code */ 45673a50541eSStefano Zampini vbs = 1; 45683a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 45693a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 45703a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 45713a50541eSStefano Zampini PetscInt *vary; 4572b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4573785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 45743a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 4575d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 4576d3df7717SStefano 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 */ 45770e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 4578d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 45793a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 45803a50541eSStefano Zampini is_blocked = PETSC_FALSE; 45813a50541eSStefano Zampini break; 45823a50541eSStefano Zampini } 45833a50541eSStefano Zampini } 4584d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 4585d3df7717SStefano Zampini } else { 4586d3df7717SStefano Zampini /* Verify directly the R set */ 4587d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 4588d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 4589d3df7717SStefano Zampini for (j=1; j<bs; j++) { 4590d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 4591d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 4592d3df7717SStefano Zampini break; 4593d3df7717SStefano Zampini } 4594d3df7717SStefano Zampini } 4595d3df7717SStefano Zampini } 4596d3df7717SStefano Zampini } 45973a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 45983a50541eSStefano Zampini vbs = bs; 45993a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 46003a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 46013a50541eSStefano Zampini } 46023a50541eSStefano Zampini } 46033a50541eSStefano Zampini } 46043a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 4605b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4606df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 460753892102SStefano Zampini 4608df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4609df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 461053892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 4611df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 461253892102SStefano Zampini } else { 46133a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 461453892102SStefano Zampini } 4615a64d13efSStefano Zampini 4616a64d13efSStefano Zampini /* print some info if requested */ 4617a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 4618a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4619a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46201575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4621a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 4622a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 46234f1b2e48SStefano 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); 4624a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4625a64d13efSStefano Zampini } 4626a64d13efSStefano Zampini 4627a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 4628b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 46296816873aSStefano Zampini IS is_aux1,is_aux2; 46306816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 46316816873aSStefano Zampini 46323a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4633854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 4634854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 4635a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 46364641a718SStefano Zampini for (i=0; i<n_D; i++) { 46374641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 46384641a718SStefano Zampini } 4639a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4640a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 46414641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 46424641a718SStefano Zampini aux_array1[j++] = i; 4643a64d13efSStefano Zampini } 4644a64d13efSStefano Zampini } 4645a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4646a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4647a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 46484641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 46494641a718SStefano Zampini aux_array2[j++] = i; 4650a64d13efSStefano Zampini } 4651a64d13efSStefano Zampini } 4652a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4653a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 4654a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 4655a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4656a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 4657a64d13efSStefano Zampini 46588eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4659785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 4660a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 46614641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 46624641a718SStefano Zampini aux_array1[j++] = i; 4663a64d13efSStefano Zampini } 4664a64d13efSStefano Zampini } 4665a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4666a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 4667a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4668a64d13efSStefano Zampini } 46694641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 46703a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4671d62866d3SStefano Zampini } else { 4672df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 46736816873aSStefano Zampini IS tis; 46746816873aSStefano Zampini PetscInt schur_size; 46756816873aSStefano Zampini 4676df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 46776816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 4678df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 46796816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 46806816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 46816816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 46826816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 46836816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 4684d62866d3SStefano Zampini } 4685d62866d3SStefano Zampini } 4686a64d13efSStefano Zampini PetscFunctionReturn(0); 4687a64d13efSStefano Zampini } 4688a64d13efSStefano Zampini 4689304d26faSStefano Zampini 4690684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 4691304d26faSStefano Zampini { 4692304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4693304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4694304d26faSStefano Zampini PC pc_temp; 4695304d26faSStefano Zampini Mat A_RR; 4696f4ddd8eeSStefano Zampini MatReuse reuse; 4697304d26faSStefano Zampini PetscScalar m_one = -1.0; 4698304d26faSStefano Zampini PetscReal value; 469904708bb6SStefano Zampini PetscInt n_D,n_R; 4700c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 4701304d26faSStefano Zampini PetscErrorCode ierr; 4702e604994aSStefano Zampini /* prefixes stuff */ 4703312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 4704e604994aSStefano Zampini size_t len; 4705304d26faSStefano Zampini 4706304d26faSStefano Zampini PetscFunctionBegin; 4707304d26faSStefano Zampini 4708e604994aSStefano Zampini /* compute prefixes */ 4709e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 4710e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 4711e604994aSStefano Zampini if (!pcbddc->current_level) { 4712e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4713e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4714e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4715e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4716e604994aSStefano Zampini } else { 4717e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4718312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4719e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4720e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 4721312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4722312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 472334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 472434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4725e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4726e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4727e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 4728e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 4729e604994aSStefano Zampini } 4730e604994aSStefano Zampini 4731304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 4732684f6988SStefano Zampini if (dirichlet) { 4733d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4734450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 47359a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 4736450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 4737a3df083aSStefano Zampini Mat A_IIn; 4738a3df083aSStefano Zampini 4739a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 4740a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 4741a3df083aSStefano Zampini pcis->A_II = A_IIn; 4742a3df083aSStefano Zampini } 4743450f8f5eSStefano Zampini } 47443301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 47453301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 4746964fefecSStefano Zampini } 4747ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 4748964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 4749304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 4750304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 4751304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 4752304d26faSStefano Zampini /* default */ 4753304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 4754e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 47559577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 4756304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 47579577ea80SStefano Zampini if (issbaij) { 47589577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 47599577ea80SStefano Zampini } else { 4760304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 47619577ea80SStefano Zampini } 4762304d26faSStefano Zampini /* Allow user's customization */ 4763304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 4764304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4765304d26faSStefano Zampini } 4766d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 4767b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4768df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4769d62866d3SStefano Zampini 4770df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 4771d5574798SStefano Zampini } 4772304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4773304d26faSStefano Zampini if (!n_D) { 4774304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 4775304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4776304d26faSStefano Zampini } 4777304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 4778304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 4779304d26faSStefano Zampini /* set ksp_D into pcis data */ 4780304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 4781304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 4782304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 4783684f6988SStefano Zampini } 4784304d26faSStefano Zampini 4785304d26faSStefano Zampini /* NEUMANN PROBLEM */ 4786684f6988SStefano Zampini A_RR = 0; 4787684f6988SStefano Zampini if (neumann) { 4788d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 478904708bb6SStefano Zampini PetscInt ibs,mbs; 479004708bb6SStefano Zampini PetscBool issbaij; 479104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4792f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 47938ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 4794f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 4795f4ddd8eeSStefano Zampini PetscInt nn_R; 479681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 4797f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4798f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 4799f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 4800f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 4801f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4802f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4803f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 4804727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 4805f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4806f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4807f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 4808f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 4809f4ddd8eeSStefano Zampini } 4810f4ddd8eeSStefano Zampini } 4811f4ddd8eeSStefano Zampini /* last check */ 4812d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 4813f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4814f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4815f4ddd8eeSStefano Zampini } 4816f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 4817f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4818f4ddd8eeSStefano Zampini } 4819a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 4820af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 4821af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 482204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 482304708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 482404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 482504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 482604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4827af732b37SStefano Zampini } else { 4828511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 48296816873aSStefano Zampini } 483004708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 483104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 483204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 483304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483404708bb6SStefano Zampini } else { 4835511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483604708bb6SStefano Zampini } 483704708bb6SStefano Zampini } 4838a00504b5SStefano Zampini /* extract A_RR */ 4839b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4840a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4841a00504b5SStefano Zampini 4842a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 484316e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4844a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 484516e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 484616e386b8SStefano Zampini } else { 48477dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 4848a00504b5SStefano Zampini } 4849a00504b5SStefano Zampini } else { 4850a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4851a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 4852a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4853a00504b5SStefano Zampini } 4854a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 48557dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 485616e386b8SStefano Zampini } 48573301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 48583301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 48596816873aSStefano Zampini } 4860f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 4861304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 4862304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 4863304d26faSStefano Zampini /* default */ 4864304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 4865e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 4866304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 48679577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 48689577ea80SStefano Zampini if (issbaij) { 48699577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 48709577ea80SStefano Zampini } else { 4871304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 48729577ea80SStefano Zampini } 4873304d26faSStefano Zampini /* Allow user's customization */ 4874304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 4875304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4876304d26faSStefano Zampini } 4877304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4878304d26faSStefano Zampini if (!n_R) { 4879304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 4880304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4881304d26faSStefano Zampini } 48825cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 4883df4d28bfSStefano Zampini /* Reuse solver if it is present */ 4884b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4885df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4886d62866d3SStefano Zampini 4887df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 4888d62866d3SStefano Zampini } 4889304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 4890304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 4891684f6988SStefano Zampini } 4892304d26faSStefano Zampini 4893684f6988SStefano Zampini if (pcbddc->dbg_flag) { 4894684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 48951575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4896684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4897684f6988SStefano Zampini } 4898c7017625SStefano Zampini 4899c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 4900c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 4901c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 4902c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 4903c7017625SStefano Zampini } 4904c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 4905c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 4906c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 4907c7017625SStefano Zampini } 4908c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 4909c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 4910c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 4911c7017625SStefano Zampini } 4912c7017625SStefano Zampini 4913c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 4914c7017625SStefano Zampini if (pcbddc->dbg_flag) { 4915684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 49160fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 49170fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 49180fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 49190fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 49200fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 4921e604994aSStefano 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); 4922c7017625SStefano Zampini if (check_corr[0]) { 4923c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 4924c7017625SStefano Zampini } 4925304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4926304d26faSStefano Zampini } 4927684f6988SStefano Zampini if (neumann) { /* Neumann */ 49280fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 49290fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 49300fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 49310fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 49320fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 4933e604994aSStefano 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); 4934c7017625SStefano Zampini if (check_corr[1]) { 4935c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 4936c7017625SStefano Zampini } 4937304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4938304d26faSStefano Zampini } 4939684f6988SStefano Zampini } 49405cbda25cSStefano Zampini /* free Neumann problem's matrix */ 49415cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4942304d26faSStefano Zampini PetscFunctionReturn(0); 4943304d26faSStefano Zampini } 4944304d26faSStefano Zampini 494580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 4946674ae819SStefano Zampini { 4947674ae819SStefano Zampini PetscErrorCode ierr; 4948674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4949be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4950b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 4951674ae819SStefano Zampini 4952674ae819SStefano Zampini PetscFunctionBegin; 4953b334f244SStefano Zampini if (!reuse_solver) { 495480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 495520c7b377SStefano Zampini } 495680677318SStefano Zampini if (!pcbddc->switch_static) { 495780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 495880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 495980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 496020c7b377SStefano Zampini } 4961b334f244SStefano Zampini if (!reuse_solver) { 496280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 496380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 496420c7b377SStefano Zampini } else { 4965df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4966be83ff47SStefano Zampini 4967df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4968df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 496920c7b377SStefano Zampini } 4970be83ff47SStefano Zampini } else { 497180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 497680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 497780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 497880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4980674ae819SStefano Zampini } 4981674ae819SStefano Zampini } 4982b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 498380677318SStefano Zampini if (applytranspose) { 498480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 498580677318SStefano Zampini } else { 498680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 498780677318SStefano Zampini } 4988be83ff47SStefano Zampini } else { 4989df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4990be83ff47SStefano Zampini 4991be83ff47SStefano Zampini if (applytranspose) { 4992df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4993be83ff47SStefano Zampini } else { 4994df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4995be83ff47SStefano Zampini } 4996be83ff47SStefano Zampini } 499780677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 499880677318SStefano Zampini if (!pcbddc->switch_static) { 4999b334f244SStefano Zampini if (!reuse_solver) { 500080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 500180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5002be83ff47SStefano Zampini } else { 5003df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5004be83ff47SStefano Zampini 5005df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5006df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5007be83ff47SStefano Zampini } 500880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 500980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 501080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 501180677318SStefano Zampini } 501280677318SStefano Zampini } else { 501380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 501480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 501580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 501680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 501780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 501880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 501980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 502080677318SStefano Zampini } 502180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5025674ae819SStefano Zampini } 5026674ae819SStefano Zampini PetscFunctionReturn(0); 5027674ae819SStefano Zampini } 5028674ae819SStefano Zampini 5029dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 5030dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 5031674ae819SStefano Zampini { 5032674ae819SStefano Zampini PetscErrorCode ierr; 5033674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 5034674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 5035674ae819SStefano Zampini const PetscScalar zero = 0.0; 5036674ae819SStefano Zampini 5037674ae819SStefano Zampini PetscFunctionBegin; 5038dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 50394fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5040dc359a40SStefano Zampini if (applytranspose) { 5041674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 50428eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 5043dc359a40SStefano Zampini } else { 5044674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 5045674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 504615aaf578SStefano Zampini } 50474fee134fSStefano Zampini } else { 50484fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 50494fee134fSStefano Zampini } 5050efc2fbd9SStefano Zampini 5051efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 50524f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5053efc2fbd9SStefano Zampini PetscScalar *array; 50544f1b2e48SStefano Zampini PetscInt j; 5055efc2fbd9SStefano Zampini 5056efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 50574f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 5058efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5059efc2fbd9SStefano Zampini } 5060efc2fbd9SStefano Zampini 506112edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 506212edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 506312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 506412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 506512edc857SStefano Zampini 50669f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 506712edc857SStefano Zampini if (pcbddc->coarse_ksp) { 506851694757SStefano Zampini Mat coarse_mat; 5069964fefecSStefano Zampini Vec rhs,sol; 507051694757SStefano Zampini MatNullSpace nullsp; 507127b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 5072964fefecSStefano Zampini 507327b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 507427b6a85dSStefano Zampini PC coarse_pc; 507527b6a85dSStefano Zampini 507627b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 507727b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 507827b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 507927b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 508027b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 508127b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 50823bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 508327b6a85dSStefano Zampini } 508427b6a85dSStefano Zampini } 5085964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 5086964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 508751694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 508851694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 508951694757SStefano Zampini if (nullsp) { 509051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 509151694757SStefano Zampini } 509212edc857SStefano Zampini if (applytranspose) { 50939a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 5094964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 50952701bc32SStefano Zampini } else { 50961f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 50972701bc32SStefano Zampini PC coarse_pc; 50982701bc32SStefano Zampini 50992701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 51002701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 51013e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 51022701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 510312edc857SStefano Zampini } else { 5104964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 510512edc857SStefano Zampini } 51062701bc32SStefano Zampini } 51071d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 510827b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 510927b6a85dSStefano Zampini PC coarse_pc; 511027b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 511127b6a85dSStefano Zampini 511227b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 511327b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 511427b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 51153bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 511627b6a85dSStefano Zampini } 511751694757SStefano Zampini if (nullsp) { 511851694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 511951694757SStefano Zampini } 512012edc857SStefano Zampini } 5121674ae819SStefano Zampini 5122674ae819SStefano Zampini /* Local solution on R nodes */ 51234fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 512480677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 51259f00e9b4SStefano Zampini } 51269f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 51279f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 512812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5129674ae819SStefano Zampini 51304fee134fSStefano Zampini /* Sum contributions from the two levels */ 51314fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5132dc359a40SStefano Zampini if (applytranspose) { 5133dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5134dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5135dc359a40SStefano Zampini } else { 5136674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 51378eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5138dc359a40SStefano Zampini } 5139efc2fbd9SStefano Zampini /* store p0 */ 51404f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5141efc2fbd9SStefano Zampini PetscScalar *array; 51424f1b2e48SStefano Zampini PetscInt j; 5143efc2fbd9SStefano Zampini 5144efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 51454f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 5146efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5147efc2fbd9SStefano Zampini } 51484fee134fSStefano Zampini } else { /* expand the coarse solution */ 51494fee134fSStefano Zampini if (applytranspose) { 51504fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 51514fee134fSStefano Zampini } else { 51524fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 51534fee134fSStefano Zampini } 51544fee134fSStefano Zampini } 5155674ae819SStefano Zampini PetscFunctionReturn(0); 5156674ae819SStefano Zampini } 5157674ae819SStefano Zampini 515812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 5159674ae819SStefano Zampini { 5160674ae819SStefano Zampini PetscErrorCode ierr; 5161674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 516258da7f69SStefano Zampini PetscScalar *array; 516312edc857SStefano Zampini Vec from,to; 5164674ae819SStefano Zampini 5165674ae819SStefano Zampini PetscFunctionBegin; 516612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 516712edc857SStefano Zampini from = pcbddc->coarse_vec; 516812edc857SStefano Zampini to = pcbddc->vec1_P; 516912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 517012edc857SStefano Zampini Vec tvec; 517158da7f69SStefano Zampini 517258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 517358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 517412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 517558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 517658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 517758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 517812edc857SStefano Zampini } 517912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 518012edc857SStefano Zampini from = pcbddc->vec1_P; 518112edc857SStefano Zampini to = pcbddc->coarse_vec; 518212edc857SStefano Zampini } 518312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 5184674ae819SStefano Zampini PetscFunctionReturn(0); 5185674ae819SStefano Zampini } 5186674ae819SStefano Zampini 518712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 5188674ae819SStefano Zampini { 5189674ae819SStefano Zampini PetscErrorCode ierr; 5190674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 519158da7f69SStefano Zampini PetscScalar *array; 519212edc857SStefano Zampini Vec from,to; 5193674ae819SStefano Zampini 5194674ae819SStefano Zampini PetscFunctionBegin; 519512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 519612edc857SStefano Zampini from = pcbddc->coarse_vec; 519712edc857SStefano Zampini to = pcbddc->vec1_P; 519812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 519912edc857SStefano Zampini from = pcbddc->vec1_P; 520012edc857SStefano Zampini to = pcbddc->coarse_vec; 520112edc857SStefano Zampini } 520212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 520312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 520412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 520512edc857SStefano Zampini Vec tvec; 520658da7f69SStefano Zampini 520712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 520858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 520958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 521058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 521158da7f69SStefano Zampini } 521258da7f69SStefano Zampini } else { 521358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 521458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 521512edc857SStefano Zampini } 521612edc857SStefano Zampini } 5217674ae819SStefano Zampini PetscFunctionReturn(0); 5218674ae819SStefano Zampini } 5219674ae819SStefano Zampini 5220984c4197SStefano Zampini /* uncomment for testing purposes */ 5221984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 5222674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 5223674ae819SStefano Zampini { 5224674ae819SStefano Zampini PetscErrorCode ierr; 5225674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 5226674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5227674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5228984c4197SStefano Zampini /* one and zero */ 5229984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 5230984c4197SStefano Zampini /* space to store constraints and their local indices */ 52319162d606SStefano Zampini PetscScalar *constraints_data; 52329162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 52339162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 52349162d606SStefano Zampini PetscInt *constraints_n; 5235984c4197SStefano Zampini /* iterators */ 5236b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 5237984c4197SStefano Zampini /* BLAS integers */ 5238e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 5239e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 5240c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 5241727cdba6SStefano Zampini /* reuse */ 52420e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 52430e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 5244984c4197SStefano Zampini /* change of basis */ 5245b3d85658SStefano Zampini PetscBool qr_needed; 52469162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 5247984c4197SStefano Zampini /* auxiliary stuff */ 524864efe560SStefano Zampini PetscInt *nnz,*is_indices; 52498a0068c3SStefano Zampini PetscInt ncc; 5250984c4197SStefano Zampini /* some quantities */ 525145a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 5252a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 5253984c4197SStefano Zampini 5254674ae819SStefano Zampini PetscFunctionBegin; 52558e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 52568e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 52578e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 525816909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 5259088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 5260088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 52610e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 52620e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 52630e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 52640e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 52650e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 5266088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5267cf5a6209SStefano Zampini 5268cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 52699162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 5270cf5a6209SStefano Zampini MatNullSpace nearnullsp; 5271cf5a6209SStefano Zampini const Vec *nearnullvecs; 5272cf5a6209SStefano Zampini Vec *localnearnullsp; 5273cf5a6209SStefano Zampini PetscScalar *array; 5274cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 5275cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 5276674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 5277b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 5278674ae819SStefano Zampini PetscScalar *work; 5279674ae819SStefano Zampini PetscReal *singular_vals; 5280674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5281674ae819SStefano Zampini PetscReal *rwork; 5282674ae819SStefano Zampini #endif 5283674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5284674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 5285674ae819SStefano Zampini #else 5286964fefecSStefano Zampini PetscBLASInt dummy_int=1; 5287964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 5288674ae819SStefano Zampini #endif 5289674ae819SStefano Zampini 5290674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 5291d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 5292e4d548c7SStefano Zampini /* print some info */ 52935c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 5294e4d548c7SStefano Zampini PetscInt nv; 5295e4d548c7SStefano Zampini 5296c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5297e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 5298e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5299e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5300e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 5301e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 5302e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 5303e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5304e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5305e4d548c7SStefano Zampini } 5306e4d548c7SStefano Zampini 5307d06fc5fdSStefano Zampini /* free unneeded index sets */ 5308d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 5309d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 5310674ae819SStefano Zampini } 5311d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 5312d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5313d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5314d06fc5fdSStefano Zampini } 5315d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5316d06fc5fdSStefano Zampini n_ISForEdges = 0; 5317d06fc5fdSStefano Zampini } 5318d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 5319d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5320d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5321d06fc5fdSStefano Zampini } 5322d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5323d06fc5fdSStefano Zampini n_ISForFaces = 0; 5324d06fc5fdSStefano Zampini } 532570022509SStefano Zampini 5326674ae819SStefano Zampini /* check if near null space is attached to global mat */ 5327674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 5328674ae819SStefano Zampini if (nearnullsp) { 5329674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 5330f4ddd8eeSStefano Zampini /* remove any stored info */ 5331f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 5332f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5333f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 5334f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 5335f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 5336473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5337f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 5338f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 5339f4ddd8eeSStefano Zampini } 5340984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 5341984c4197SStefano Zampini nnsp_size = 0; 5342674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 5343674ae819SStefano Zampini } 5344984c4197SStefano Zampini /* get max number of constraints on a single cc */ 5345984c4197SStefano Zampini max_constraints = nnsp_size; 5346984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 5347984c4197SStefano Zampini 5348674ae819SStefano Zampini /* 5349674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 53509162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 53519162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 53529162d606SStefano Zampini There can be multiple constraints per connected component 5353674ae819SStefano Zampini */ 5354674ae819SStefano Zampini n_vertices = 0; 5355674ae819SStefano Zampini if (ISForVertices) { 5356674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 5357674ae819SStefano Zampini } 53589162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 53599162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 53609162d606SStefano Zampini 53619162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 53629162d606SStefano Zampini total_counts *= max_constraints; 5363674ae819SStefano Zampini total_counts += n_vertices; 53644641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 53659162d606SStefano Zampini 5366674ae819SStefano Zampini total_counts = 0; 5367674ae819SStefano Zampini max_size_of_constraint = 0; 5368674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 53699162d606SStefano Zampini IS used_is; 5370674ae819SStefano Zampini if (i<n_ISForEdges) { 53719162d606SStefano Zampini used_is = ISForEdges[i]; 5372674ae819SStefano Zampini } else { 53739162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 5374674ae819SStefano Zampini } 53759162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 5376674ae819SStefano Zampini total_counts += j; 5377674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 5378674ae819SStefano Zampini } 53799162d606SStefano 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); 53809162d606SStefano Zampini 5381984c4197SStefano Zampini /* get local part of global near null space vectors */ 5382785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 5383984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5384984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 5385e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5386e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5387984c4197SStefano Zampini } 5388674ae819SStefano Zampini 5389242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 5390242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 5391a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 5392242a89d7SStefano Zampini 5393984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 5394a773dcb8SStefano Zampini if (!skip_lapack) { 5395674ae819SStefano Zampini PetscScalar temp_work; 5396911cabfeSStefano Zampini 5397674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5398984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 5399785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 5400785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 5401785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 5402674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5403785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 5404674ae819SStefano Zampini #endif 5405674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5406c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 5407c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 5408674ae819SStefano Zampini lwork = -1; 5409674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5410674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5411c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 5412674ae819SStefano Zampini #else 5413c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 5414674ae819SStefano Zampini #endif 5415674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5416984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 5417674ae819SStefano Zampini #else /* on missing GESVD */ 5418674ae819SStefano Zampini /* SVD */ 5419674ae819SStefano Zampini PetscInt max_n,min_n; 5420674ae819SStefano Zampini max_n = max_size_of_constraint; 5421984c4197SStefano Zampini min_n = max_constraints; 5422984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 5423674ae819SStefano Zampini min_n = max_size_of_constraint; 5424984c4197SStefano Zampini max_n = max_constraints; 5425674ae819SStefano Zampini } 5426785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 5427674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5428785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 5429674ae819SStefano Zampini #endif 5430674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5431674ae819SStefano Zampini lwork = -1; 5432e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 5433e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 5434b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 5435674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5436674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 54379162d606SStefano 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)); 5438674ae819SStefano Zampini #else 54399162d606SStefano 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)); 5440674ae819SStefano Zampini #endif 5441674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5442984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 5443984c4197SStefano Zampini #endif /* on missing GESVD */ 5444674ae819SStefano Zampini /* Allocate optimal workspace */ 5445674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 5446854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 5447674ae819SStefano Zampini } 5448674ae819SStefano Zampini /* Now we can loop on constraining sets */ 5449674ae819SStefano Zampini total_counts = 0; 54509162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 54519162d606SStefano Zampini constraints_data_ptr[0] = 0; 5452674ae819SStefano Zampini /* vertices */ 54539162d606SStefano Zampini if (n_vertices) { 5454674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 54559162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 5456674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 54579162d606SStefano Zampini constraints_n[total_counts] = 1; 54589162d606SStefano Zampini constraints_data[total_counts] = 1.0; 54599162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 54609162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 5461674ae819SStefano Zampini total_counts++; 5462674ae819SStefano Zampini } 5463674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5464674ae819SStefano Zampini n_vertices = total_counts; 5465674ae819SStefano Zampini } 5466984c4197SStefano Zampini 5467674ae819SStefano Zampini /* edges and faces */ 54689162d606SStefano Zampini total_counts_cc = total_counts; 5469911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 54709162d606SStefano Zampini IS used_is; 54719162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 54729162d606SStefano Zampini 5473911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 54749162d606SStefano Zampini used_is = ISForEdges[ncc]; 5475984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 5476674ae819SStefano Zampini } else { 54779162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 5478984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 5479674ae819SStefano Zampini } 5480674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 54819162d606SStefano Zampini 54829162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 54839162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5484984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 5485984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 5486674ae819SStefano Zampini if (nnsp_has_cnst) { 54875b08dc53SStefano Zampini PetscScalar quad_value; 54889162d606SStefano Zampini 54899162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 54909162d606SStefano Zampini idxs_copied = PETSC_TRUE; 54919162d606SStefano Zampini 5492a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 5493674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 5494a773dcb8SStefano Zampini } else { 5495a773dcb8SStefano Zampini quad_value = 1.0; 5496a773dcb8SStefano Zampini } 5497674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 54989162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 5499674ae819SStefano Zampini } 55009162d606SStefano Zampini temp_constraints++; 5501674ae819SStefano Zampini total_counts++; 5502674ae819SStefano Zampini } 5503674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 5504984c4197SStefano Zampini PetscReal real_value; 55059162d606SStefano Zampini PetscScalar *ptr_to_data; 55069162d606SStefano Zampini 5507984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 55089162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 5509674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 55109162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 5511674ae819SStefano Zampini } 5512984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 5513984c4197SStefano Zampini /* check if array is null on the connected component */ 5514e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 55159162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 55165b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 5517674ae819SStefano Zampini temp_constraints++; 5518674ae819SStefano Zampini total_counts++; 55199162d606SStefano Zampini if (!idxs_copied) { 55209162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 55219162d606SStefano Zampini idxs_copied = PETSC_TRUE; 5522674ae819SStefano Zampini } 5523674ae819SStefano Zampini } 55249162d606SStefano Zampini } 55259162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 552645a1bb75SStefano Zampini valid_constraints = temp_constraints; 5527eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 5528a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 55299162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 55309162d606SStefano Zampini 55319162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5532a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 55339162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 5534a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 55359162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 5536a773dcb8SStefano Zampini } else { /* perform SVD */ 5537984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 55389162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5539674ae819SStefano Zampini 5540674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5541984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 5542984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 5543984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 5544984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 5545984c4197SStefano Zampini from that computed using LAPACKgesvd 5546984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 5547984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 5548984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 5549674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 5550e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5551984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5552674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 5553674ae819SStefano Zampini for (k=0;k<j+1;k++) { 55549162d606SStefano 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)); 5555674ae819SStefano Zampini } 5556674ae819SStefano Zampini } 5557e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 5558e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5559e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 5560674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5561c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 5562674ae819SStefano Zampini #else 5563c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 5564674ae819SStefano Zampini #endif 5565674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5566984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 5567984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 5568674ae819SStefano Zampini j = 0; 5569984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 5570674ae819SStefano Zampini total_counts = total_counts-j; 557145a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 5572e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 5573c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5574c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5575c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 5576c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5577c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 5578c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5579674ae819SStefano Zampini if (j<temp_constraints) { 5580984c4197SStefano Zampini PetscInt ii; 5581984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 5582674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 55839162d606SStefano 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)); 5584674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5585984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 5586674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 55879162d606SStefano 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]; 5588674ae819SStefano Zampini } 5589674ae819SStefano Zampini } 5590674ae819SStefano Zampini } 5591674ae819SStefano Zampini #else /* on missing GESVD */ 5592e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5593e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5594b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5595674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5596674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 55979162d606SStefano 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)); 5598674ae819SStefano Zampini #else 55999162d606SStefano 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)); 5600674ae819SStefano Zampini #endif 5601984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 5602674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5603984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 5604e310c8b4SStefano Zampini k = temp_constraints; 5605e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 5606674ae819SStefano Zampini j = 0; 5607e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 560845a1bb75SStefano Zampini valid_constraints = k-j; 5609911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 5610984c4197SStefano Zampini #endif /* on missing GESVD */ 5611674ae819SStefano Zampini } 5612a773dcb8SStefano Zampini } 56139162d606SStefano Zampini /* update pointers information */ 56149162d606SStefano Zampini if (valid_constraints) { 56159162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 56169162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 56179162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 56189162d606SStefano Zampini /* set change_of_basis flag */ 561945a1bb75SStefano Zampini if (boolforchange) { 5620b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 56219162d606SStefano Zampini } 5622b3d85658SStefano Zampini total_counts_cc++; 562345a1bb75SStefano Zampini } 562445a1bb75SStefano Zampini } 5625984c4197SStefano Zampini /* free workspace */ 56268f1c130eSStefano Zampini if (!skip_lapack) { 5627984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 5628984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5629984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 5630984c4197SStefano Zampini #endif 5631984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 5632984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5633984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 5634984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 5635984c4197SStefano Zampini #endif 5636984c4197SStefano Zampini } 5637984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5638984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 5639984c4197SStefano Zampini } 5640984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 5641cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 5642cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5643cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5644cf5a6209SStefano Zampini } 5645cf5a6209SStefano Zampini if (n_ISForFaces) { 5646cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5647cf5a6209SStefano Zampini } 5648cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5649cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5650cf5a6209SStefano Zampini } 5651cf5a6209SStefano Zampini if (n_ISForEdges) { 5652cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5653cf5a6209SStefano Zampini } 5654cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 565508122e43SStefano Zampini } else { 565608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5657984c4197SStefano Zampini 565808122e43SStefano Zampini total_counts = 0; 565908122e43SStefano Zampini n_vertices = 0; 5660d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 5661d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 566208122e43SStefano Zampini } 566308122e43SStefano Zampini max_constraints = 0; 56649162d606SStefano Zampini total_counts_cc = 0; 566508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 566608122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 56679162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 566808122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 566908122e43SStefano Zampini } 56709162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 56719162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 56729162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 56739162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 567474d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 56759162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 56769162d606SStefano Zampini total_counts_cc = 0; 56779162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 56789162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 56799162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 568008122e43SStefano Zampini } 568108122e43SStefano Zampini } 56829162d606SStefano Zampini #if 0 56839162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 56849162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 56859162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 56869162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 56879162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 56889162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 56899162d606SStefano Zampini } 56909162d606SStefano Zampini printf("\n"); 56919162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 56929162d606SStefano Zampini } 56931b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 56948bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 56951b968477SStefano Zampini } 56961b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 56978bec7fa6SStefano 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]); 56981b968477SStefano Zampini } 569908122e43SStefano Zampini #endif 570008122e43SStefano Zampini 57018bec7fa6SStefano Zampini max_size_of_constraint = 0; 57029162d606SStefano 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]); 57039162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 570408122e43SStefano Zampini /* Change of basis */ 5705b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 570608122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 570708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 570808122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 5709b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 571008122e43SStefano Zampini } 571108122e43SStefano Zampini } 571208122e43SStefano Zampini } 571308122e43SStefano Zampini } 5714984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 57154f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 571608122e43SStefano Zampini 57179162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 57189162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 5719eee23b56SStefano 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); 5720674ae819SStefano Zampini 5721674ae819SStefano Zampini /* Create constraint matrix */ 5722674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 572316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 5724984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 5725984c4197SStefano Zampini 5726984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 5727a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 5728a717540cSStefano Zampini qr_needed = PETSC_FALSE; 572974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 5730984c4197SStefano Zampini total_primal_vertices=0; 5731b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 57329162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 57339162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 573472b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 57359162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 5736b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 573764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 57389162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 57399162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 5740a717540cSStefano Zampini } 5741b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 574291af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 5743a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 5744a717540cSStefano Zampini qr_needed = PETSC_TRUE; 5745a717540cSStefano Zampini } 5746fa434743SStefano Zampini } else { 5747b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 5748fa434743SStefano Zampini } 5749a717540cSStefano Zampini } 5750b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 5751b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 5752674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 575370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 57544f1b2e48SStefano 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); 57550e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 57560e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 5757984c4197SStefano Zampini 5758984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 575974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 5760785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 5761984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 576274d5cdf7SStefano Zampini 5763984c4197SStefano Zampini j = total_primal_vertices; 576474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 5765b3d85658SStefano Zampini cum = total_primal_vertices; 57669162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 57674641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 5768b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 5769b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 5770b3d85658SStefano Zampini cum++; 57719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 577274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 577374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 577474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 577574d5cdf7SStefano Zampini } 57769162d606SStefano Zampini j += constraints_n[i]; 5777674ae819SStefano Zampini } 5778674ae819SStefano Zampini } 5779674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 5780674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 5781088faed8SStefano Zampini 5782674ae819SStefano Zampini /* set values in constraint matrix */ 5783984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 57840e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5785674ae819SStefano Zampini } 5786984c4197SStefano Zampini total_counts = total_primal_vertices; 57879162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 57884641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 57899162d606SStefano Zampini PetscInt *cols; 57909162d606SStefano Zampini 57919162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 57929162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 57939162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 57949162d606SStefano Zampini PetscInt row = total_counts+k; 57959162d606SStefano Zampini PetscScalar *vals; 57969162d606SStefano Zampini 57979162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 57989162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 57999162d606SStefano Zampini } 58009162d606SStefano Zampini total_counts += constraints_n[i]; 5801674ae819SStefano Zampini } 5802674ae819SStefano Zampini } 5803674ae819SStefano Zampini /* assembling */ 5804674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5805674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5806088faed8SStefano Zampini 5807984c4197SStefano Zampini /* 58086a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5809984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 5810f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5811984c4197SStefano Zampini */ 5812674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 5813674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 5814026de310SStefano Zampini /* dual and primal dofs on a single cc */ 5815984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 5816984c4197SStefano Zampini /* working stuff for GEQRF */ 581781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 5818984c4197SStefano Zampini PetscBLASInt lqr_work; 5819984c4197SStefano Zampini /* working stuff for UNGQR */ 5820984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 5821984c4197SStefano Zampini PetscBLASInt lgqr_work; 5822984c4197SStefano Zampini /* working stuff for TRTRS */ 5823984c4197SStefano Zampini PetscScalar *trs_rhs; 58243f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 5825984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 5826984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 5827984c4197SStefano Zampini PetscScalar *start_vals; 5828984c4197SStefano Zampini /* working stuff for values insertion */ 58294641a718SStefano Zampini PetscBT is_primal; 583064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 5831906d46d4SStefano Zampini /* matrix sizes */ 5832906d46d4SStefano Zampini PetscInt global_size,local_size; 5833906d46d4SStefano Zampini /* temporary change of basis */ 5834906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 5835cf5a6209SStefano Zampini /* extra space for debugging */ 5836cf5a6209SStefano Zampini PetscScalar *dbg_work; 5837984c4197SStefano Zampini 5838906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 5839906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 584016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 5841bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 5842906d46d4SStefano Zampini /* nonzeros for local mat */ 5843bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 58441dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5845bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 58461dd7afcfSStefano Zampini } else { 58471dd7afcfSStefano Zampini const PetscInt *ii; 58481dd7afcfSStefano Zampini PetscInt n; 58491dd7afcfSStefano Zampini PetscBool flg_row; 58501dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 58511dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 58521dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 58531dd7afcfSStefano Zampini } 58549162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 5855a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 58569162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 5857a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 58589162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 5859a717540cSStefano Zampini } else { 58609162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 58619162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 5862a717540cSStefano Zampini } 5863a717540cSStefano Zampini } 5864a717540cSStefano Zampini } 5865906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 5866bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 58671dd7afcfSStefano Zampini /* Set interior change in the matrix */ 58681dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5869bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 5870906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 5871a717540cSStefano Zampini } 58721dd7afcfSStefano Zampini } else { 58731dd7afcfSStefano Zampini const PetscInt *ii,*jj; 58741dd7afcfSStefano Zampini PetscScalar *aa; 58751dd7afcfSStefano Zampini PetscInt n; 58761dd7afcfSStefano Zampini PetscBool flg_row; 58771dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 58781dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 58791dd7afcfSStefano Zampini for (i=0;i<n;i++) { 58801dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 58811dd7afcfSStefano Zampini } 58821dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 58831dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 58841dd7afcfSStefano Zampini } 5885a717540cSStefano Zampini 5886a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5887a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5888a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5889a717540cSStefano Zampini } 5890a717540cSStefano Zampini 5891a717540cSStefano Zampini 5892a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 5893a717540cSStefano Zampini /* 5894a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 5895a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 5896a717540cSStefano Zampini 5897a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 5898a717540cSStefano Zampini 5899a6b551f4SStefano 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) 5900a6b551f4SStefano Zampini 5901a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 5902a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 5903a717540cSStefano Zampini | ... | 5904a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 5905a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 5906a717540cSStefano Zampini 5907a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 5908a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 5909a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 5910a6b551f4SStefano Zampini 5911a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 5912a717540cSStefano Zampini */ 5913a717540cSStefano Zampini if (qr_needed) { 5914984c4197SStefano Zampini /* space to store Q */ 5915854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 59164e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 59174e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 5918984c4197SStefano Zampini /* first we issue queries for optimal work */ 59193f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 59203f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 59213f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5922984c4197SStefano Zampini lqr_work = -1; 59233f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 5924984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 5925984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 5926785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 5927984c4197SStefano Zampini lgqr_work = -1; 59283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 59293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 59303f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 59313f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 59323f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 59333f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 5934984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 5935984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 5936785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 5937984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 5938785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 5939a717540cSStefano Zampini /* allocating workspace for check */ 5940a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5941cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 5942a717540cSStefano Zampini } 5943a717540cSStefano Zampini } 5944984c4197SStefano Zampini /* array to store whether a node is primal or not */ 59454641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 5946473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 59470e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 5948eee23b56SStefano 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); 594939e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 595039e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 595139e2fb2aSStefano Zampini } 595239e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 5953984c4197SStefano Zampini 5954a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 59559162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 59569162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 59574641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 5958984c4197SStefano Zampini /* get constraint info */ 59599162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 5960984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 5961984c4197SStefano Zampini 5962984c4197SStefano Zampini if (pcbddc->dbg_flag) { 59639162d606SStefano 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); 5964674ae819SStefano Zampini } 5965984c4197SStefano Zampini 5966fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 5967a717540cSStefano Zampini 5968a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 5969a717540cSStefano Zampini if (pcbddc->dbg_flag) { 59709162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5971a717540cSStefano Zampini } 5972984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 59739162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5974984c4197SStefano Zampini 5975984c4197SStefano Zampini /* compute QR decomposition of constraints */ 59763f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 59773f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 59783f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5979674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59803f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 5981984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 5982674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5983984c4197SStefano Zampini 5984984c4197SStefano Zampini /* explictly compute R^-T */ 5985984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 5986984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 59873f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 59883f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 59893f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 59903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 5991984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59923f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 5993984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 5994984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5995984c4197SStefano Zampini 5996a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 59973f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 59983f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 59993f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 60003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6001984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 60023f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 6003984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 6004984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6005984c4197SStefano Zampini 6006984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 6007984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 6008984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 60093f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 60103f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 60113f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 60123f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 60133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 60143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 6015984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 60169162d606SStefano 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)); 6017984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 60189162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 6019984c4197SStefano Zampini 6020984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 60219162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 6022984c4197SStefano Zampini /* insert cols for primal dofs */ 6023984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 6024984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 60259162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6026906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6027984c4197SStefano Zampini } 6028984c4197SStefano Zampini /* insert cols for dual dofs */ 6029984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 60309162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 6031984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 60329162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6033906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6034984c4197SStefano Zampini j++; 6035674ae819SStefano Zampini } 6036674ae819SStefano Zampini } 6037984c4197SStefano Zampini 6038984c4197SStefano Zampini /* check change of basis */ 6039984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6040984c4197SStefano Zampini PetscInt ii,jj; 6041984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 6042c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 6043c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 6044c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 6045c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6046c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 6047c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 6048984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6049cf5a6209SStefano 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)); 6050984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6051984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6052984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6053cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 6054cf5a6209SStefano 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; 6055674ae819SStefano Zampini } 6056674ae819SStefano Zampini } 6057984c4197SStefano Zampini if (!valid_qr) { 605822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 6059984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6060984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6061cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 6062cf5a6209SStefano 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])); 6063674ae819SStefano Zampini } 6064cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 6065cf5a6209SStefano 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])); 6066984c4197SStefano Zampini } 6067984c4197SStefano Zampini } 6068984c4197SStefano Zampini } 6069674ae819SStefano Zampini } else { 607022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 6071674ae819SStefano Zampini } 6072674ae819SStefano Zampini } 6073a717540cSStefano Zampini } else { /* simple transformation block */ 6074a717540cSStefano Zampini PetscInt row,col; 6075a6b551f4SStefano Zampini PetscScalar val,norm; 6076a6b551f4SStefano Zampini 6077a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 60789162d606SStefano 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)); 6079a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 60809162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 60819162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6082bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 60839162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 6084906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 60859162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 6086a717540cSStefano Zampini } else { 6087a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 60889162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6089a717540cSStefano Zampini if (row != col) { 60909162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 6091a717540cSStefano Zampini } else { 60929162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 6093a717540cSStefano Zampini } 6094906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 6095a717540cSStefano Zampini } 6096a717540cSStefano Zampini } 6097a717540cSStefano Zampini } 609898a51de6SStefano Zampini if (pcbddc->dbg_flag) { 609922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 6100a717540cSStefano Zampini } 6101674ae819SStefano Zampini } 6102984c4197SStefano Zampini } else { 6103984c4197SStefano Zampini if (pcbddc->dbg_flag) { 61049162d606SStefano 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); 6105674ae819SStefano Zampini } 6106674ae819SStefano Zampini } 6107674ae819SStefano Zampini } 6108a717540cSStefano Zampini 6109a717540cSStefano Zampini /* free workspace */ 6110a717540cSStefano Zampini if (qr_needed) { 6111984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6112cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 6113984c4197SStefano Zampini } 6114984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 6115984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 6116984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 6117984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 6118984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 6119674ae819SStefano Zampini } 6120a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 6121906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6122906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6123906d46d4SStefano Zampini 6124906d46d4SStefano Zampini /* assembling of global change of variable */ 612588c03ad3SStefano Zampini if (!pcbddc->fake_change) { 6126bbb9e6c6SStefano Zampini Mat tmat; 612716f15bc4SStefano Zampini PetscInt bs; 612816f15bc4SStefano Zampini 6129906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 6130906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 6131bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 6132bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 6133bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6134bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 613516f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 613616f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 6137906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 6138bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 6139bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6140bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6141bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6142bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6143e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6144e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6145bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 6146bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 614788c03ad3SStefano Zampini 6148906d46d4SStefano Zampini /* check */ 6149906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 6150906d46d4SStefano Zampini PetscReal error; 6151906d46d4SStefano Zampini Vec x,x_change; 6152906d46d4SStefano Zampini 6153906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 6154906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 6155906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 6156906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 6157e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6158e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6159bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 6160e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6161e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6162906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 6163906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 6164906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 6165637e8532SStefano Zampini if (error > PETSC_SMALL) { 6166637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 6167637e8532SStefano Zampini } 6168906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 6169906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 6170906d46d4SStefano Zampini } 6171b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 6172b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 6173b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6174bf3a8328SStefano Zampini 61759a962809SStefano 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); 6176b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 6177ac632422SStefano Zampini Mat S_new,tmat; 6178bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 6179bbb9e6c6SStefano Zampini 6180bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 61817dae84e0SHong Zhang ierr = MatCreateSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 6182bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6183bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 6184bf3a8328SStefano Zampini IS is_V; 6185b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 6186b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 6187b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 6188b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 6189b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 6190bf3a8328SStefano Zampini } 6191bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 6192ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6193b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 6194ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6195bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6196bf3a8328SStefano Zampini const PetscScalar *array; 6197bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 6198bf3a8328SStefano Zampini PetscInt i,n_V; 6199bf3a8328SStefano Zampini 6200b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6201b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 6202b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6203b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6204b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 6205b087196eSStefano Zampini for (i=0;i<n_V;i++) { 6206b087196eSStefano Zampini PetscScalar val; 6207b087196eSStefano Zampini PetscInt idx; 6208b087196eSStefano Zampini 6209b087196eSStefano Zampini idx = idxs_V[i]; 6210b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 6211b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 6212b087196eSStefano Zampini } 6213b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6214b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6215bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 6216bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6217bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6218bf3a8328SStefano Zampini } 6219ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 6220ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6221ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 6222ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6223b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 6224ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6225bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6226b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6227bf3a8328SStefano Zampini } 6228ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 6229ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6230ac632422SStefano Zampini } 6231b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 623288c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6233b96c3477SStefano Zampini } 6234c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 6235b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 6236c9db6a07SStefano Zampini PetscInt i; 6237c9db6a07SStefano Zampini 6238c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6239c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 6240c9db6a07SStefano Zampini } 6241c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 6242c9db6a07SStefano Zampini } 6243b96c3477SStefano Zampini } 624416909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 624516909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 624616909a7fSStefano Zampini } else { 6247906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 624816909a7fSStefano Zampini } 62491dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 625027b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 625172b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 625272b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 625372b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 625472b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 625572b8c272SStefano Zampini } 62561dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 625727b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 6258b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 6259b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 6260906d46d4SStefano Zampini } else { 62611dd7afcfSStefano Zampini Mat benign_global = NULL; 626227b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 62631dd7afcfSStefano Zampini Mat tmat; 62641dd7afcfSStefano Zampini 62651dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 62661dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 62671dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 62681dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 62691dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 62701dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 62711dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 62721dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 62731dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 62741dd7afcfSStefano Zampini if (pcbddc->benign_change) { 62751dd7afcfSStefano Zampini Mat M; 62761dd7afcfSStefano Zampini 62771dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 62781dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 62791dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 62801dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 6281906d46d4SStefano Zampini } else { 62821dd7afcfSStefano Zampini Mat eye; 62831dd7afcfSStefano Zampini PetscScalar *array; 62841dd7afcfSStefano Zampini 62851dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 62861dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 62871dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 62881dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 6289906d46d4SStefano Zampini } 62901dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 62911dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 62921dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 62931dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 62941dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 62951dd7afcfSStefano Zampini } 62961dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 62971dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 62981dd7afcfSStefano Zampini } 62991dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 63001dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 63011dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 630227b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 63031dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 63041dd7afcfSStefano Zampini } 63051dd7afcfSStefano Zampini } 630616909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 630716909a7fSStefano Zampini IS is_global; 630816909a7fSStefano Zampini const PetscInt *gidxs; 630916909a7fSStefano Zampini 631016909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 631116909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 631216909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 63137dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 631416909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 631516909a7fSStefano Zampini } 63161dd7afcfSStefano Zampini } 63171dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 63181dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 6319b9b85e73SStefano Zampini } 6320a717540cSStefano Zampini 632172b8c272SStefano Zampini if (!pcbddc->fake_change) { 63224f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 63234f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 63244f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 63254f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 6326019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 6327019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 6328019a44ceSStefano Zampini pcbddc->local_primal_size++; 6329019a44ceSStefano Zampini } 6330019a44ceSStefano Zampini 6331019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 6332727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 6333727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 63349f47a83aSStefano 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); 6335c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 63360e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 63379f47a83aSStefano 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); 6338727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 6339727cdba6SStefano Zampini } 63400e6343abSStefano Zampini } 6341727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 6342b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 634372b8c272SStefano Zampini } 634472b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 6345727cdba6SStefano Zampini 6346a717540cSStefano Zampini /* flush dbg viewer */ 6347b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 6348b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6349b8ffe317SStefano Zampini } 6350a717540cSStefano Zampini 6351e310c8b4SStefano Zampini /* free workspace */ 6352a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 63534641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 635408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 63559162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 63569162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 635708122e43SStefano Zampini } else { 63589162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 63599162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 63609162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 636108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 636208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 63639162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 63649162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 636508122e43SStefano Zampini } 6366674ae819SStefano Zampini PetscFunctionReturn(0); 6367674ae819SStefano Zampini } 6368674ae819SStefano Zampini 6369674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 6370674ae819SStefano Zampini { 637171582508SStefano Zampini ISLocalToGlobalMapping map; 6372674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 6373674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 637466da6bd7Sstefano_zampini PetscInt i,N; 637566da6bd7Sstefano_zampini PetscBool rcsr = PETSC_FALSE; 637666da6bd7Sstefano_zampini PetscErrorCode ierr; 6377674ae819SStefano Zampini 6378674ae819SStefano Zampini PetscFunctionBegin; 63798af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 6380b03ebc13SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 63818e61c736SStefano Zampini /* Reset previously computed graph */ 63828e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 6383674ae819SStefano Zampini /* Init local Graph struct */ 63847fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 638571582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 6386be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 6387674ae819SStefano Zampini 63887a0e7b2cSstefano_zampini if (pcbddc->user_primal_vertices_local && !pcbddc->user_primal_vertices) { 63897a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 63907a0e7b2cSstefano_zampini } 6391575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 63929a962809SStefano 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); 63939577ea80SStefano Zampini 6394674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 639566da6bd7Sstefano_zampini if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) { 63964d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 63974d379d7bSStefano Zampini PetscInt nvtxs; 6398e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 6399674ae819SStefano Zampini 64002fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 64012fffb893SStefano Zampini if (flg_row) { 64024d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 6403b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 64042fffb893SStefano Zampini } 64052fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 640666da6bd7Sstefano_zampini rcsr = PETSC_TRUE; 6407674ae819SStefano Zampini } 64089b28b941SStefano Zampini if (pcbddc->dbg_flag) { 64099b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6410674ae819SStefano Zampini } 6411674ae819SStefano Zampini 6412674ae819SStefano Zampini /* Setup of Graph */ 64134b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 641414f95afaSStefano 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); 6415674ae819SStefano Zampini 64164f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 64174f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 64184f1b2e48SStefano Zampini PetscInt *local_subs; 64194f1b2e48SStefano Zampini 64204f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 64214f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 64224f1b2e48SStefano Zampini const PetscInt *idxs; 64234f1b2e48SStefano Zampini PetscInt nl,j; 64244f1b2e48SStefano Zampini 64254f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 64264f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 642771582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 64284f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 64294f1b2e48SStefano Zampini } 64304f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 64314f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 64324f1b2e48SStefano Zampini } 64338af8fcf9SStefano Zampini } 64344f1b2e48SStefano Zampini 6435cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 6436674ae819SStefano Zampini /* Graph's connected components analysis */ 6437674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 643871582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 64398af8fcf9SStefano Zampini } 644066da6bd7Sstefano_zampini if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0; 6441674ae819SStefano Zampini PetscFunctionReturn(0); 6442674ae819SStefano Zampini } 6443674ae819SStefano Zampini 64449a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 64459a7d3425SStefano Zampini { 64469a7d3425SStefano Zampini PetscInt i,j; 64479a7d3425SStefano Zampini PetscScalar *alphas; 64489a7d3425SStefano Zampini PetscErrorCode ierr; 64499a7d3425SStefano Zampini 64509a7d3425SStefano Zampini PetscFunctionBegin; 6451785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 64529a7d3425SStefano Zampini for (i=0;i<n;i++) { 64539a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 6454669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 6455669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 6456669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 64579a7d3425SStefano Zampini } 64589a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 64599a7d3425SStefano Zampini PetscFunctionReturn(0); 64609a7d3425SStefano Zampini } 64619a7d3425SStefano Zampini 6462bb360cb4SStefano Zampini PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 6463e7931f94SStefano Zampini { 646457de7509SStefano Zampini Mat A; 6465e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 6466e7931f94SStefano Zampini PetscMPIInt size,rank,color; 646752e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 646852e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 6469bb360cb4SStefano Zampini PetscInt im_active,active_procs,N,n,i,j,threshold = 2; 647057de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 647127b6a85dSStefano Zampini PetscInt xadj_count,*count; 647227b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 647327b6a85dSStefano Zampini PetscSubcomm psubcomm; 647427b6a85dSStefano Zampini MPI_Comm subcomm; 647552e5ac9dSStefano Zampini PetscErrorCode ierr; 6476a57a6d2fSStefano Zampini 6477e7931f94SStefano Zampini PetscFunctionBegin; 647857de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 647957de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 6480fbfcfee5SBarry Smith if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME); 648157de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 648257de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 648357de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 648457de7509SStefano Zampini 648557de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 648657de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 648757de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 648857de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 648957de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 6490bb360cb4SStefano Zampini im_active = !!n; 649157de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 649257de7509SStefano Zampini void_procs = size - active_procs; 649357de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 649457de7509SStefano Zampini if (void_procs) { 649557de7509SStefano Zampini PetscInt ncand; 649657de7509SStefano Zampini 649757de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 649857de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 649957de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 650057de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 650157de7509SStefano Zampini if (!procs_candidates[i]) { 650257de7509SStefano Zampini procs_candidates[ncand++] = i; 650357de7509SStefano Zampini } 650457de7509SStefano Zampini } 650557de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 650657de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 650757de7509SStefano Zampini } 650857de7509SStefano Zampini 6509bb360cb4SStefano Zampini /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix 651014f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 6511bb360cb4SStefano Zampini ierr = MatGetSize(mat,&N,NULL);CHKERRQ(ierr); 6512bb360cb4SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) { 651314f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 651414f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 651514f0bfb9SStefano Zampini else dest = rank; 651657de7509SStefano Zampini if (im_active) { 651757de7509SStefano Zampini issize = 1; 651857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 651914f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 652057de7509SStefano Zampini } else { 652114f0bfb9SStefano Zampini isidx = dest; 652257de7509SStefano Zampini } 652357de7509SStefano Zampini } else { 652457de7509SStefano Zampini issize = 0; 652557de7509SStefano Zampini isidx = -1; 652657de7509SStefano Zampini } 6527bb360cb4SStefano Zampini if (*n_subdomains != 1) *n_subdomains = active_procs; 652857de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 6529daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 653057de7509SStefano Zampini PetscFunctionReturn(0); 653157de7509SStefano Zampini } 6532c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 6533c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 653427b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 6535e7931f94SStefano Zampini 6536e7931f94SStefano Zampini /* Get info on mapping */ 65373bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6538e7931f94SStefano Zampini 6539e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 6540785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 6541e7931f94SStefano Zampini xadj[0] = 0; 6542e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 6543785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 6544785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 6545bb360cb4SStefano Zampini ierr = PetscCalloc1(n,&count);CHKERRQ(ierr); 654627b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 654727b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 654827b6a85dSStefano Zampini count[shared[i][j]] += 1; 6549e7931f94SStefano Zampini 655027b6a85dSStefano Zampini xadj_count = 0; 65512b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 655227b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 655327b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 6554d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 6555d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 6556d023bfaeSStefano Zampini xadj_count++; 655727b6a85dSStefano Zampini break; 655827b6a85dSStefano Zampini } 6559e7931f94SStefano Zampini } 6560e7931f94SStefano Zampini } 6561d023bfaeSStefano Zampini xadj[1] = xadj_count; 656227b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 65633bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6564e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 6565e7931f94SStefano Zampini 65663837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 6567e7931f94SStefano Zampini 656827b6a85dSStefano Zampini /* Restrict work on active processes only */ 656927b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 657027b6a85dSStefano Zampini if (void_procs) { 657127b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 657227b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 657327b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 657427b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 657527b6a85dSStefano Zampini } else { 657627b6a85dSStefano Zampini psubcomm = NULL; 657727b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 657827b6a85dSStefano Zampini } 657927b6a85dSStefano Zampini 658027b6a85dSStefano Zampini v_wgt = NULL; 658127b6a85dSStefano Zampini if (!color) { 6582e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 6583e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 6584e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6585c8587f34SStefano Zampini } else { 658652e5ac9dSStefano Zampini Mat subdomain_adj; 658752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 658852e5ac9dSStefano Zampini MatPartitioning partitioner; 658927b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 659052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 659157de7509SStefano Zampini PetscMPIInt size; 6592b0c7d250SStefano Zampini PetscBool aggregate; 6593b0c7d250SStefano Zampini 659427b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 659527b6a85dSStefano Zampini if (void_procs) { 659627b6a85dSStefano Zampini PetscInt prank = rank; 6597785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 659827b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 6599e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 6600e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 6601c8587f34SStefano Zampini } 6602e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 660327b6a85dSStefano Zampini } else { 660427b6a85dSStefano Zampini oldranks = NULL; 660527b6a85dSStefano Zampini } 6606b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 660727b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 6608b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 6609b0c7d250SStefano Zampini PetscMPIInt nrank; 6610b0c7d250SStefano Zampini PetscScalar *vals; 6611b0c7d250SStefano Zampini 661227b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 6613b0c7d250SStefano Zampini lrows = 0; 6614b0c7d250SStefano Zampini if (nrank<redprocs) { 6615b0c7d250SStefano Zampini lrows = size/redprocs; 6616b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 6617b0c7d250SStefano Zampini } 661827b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 6619b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 6620b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6621b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6622b0c7d250SStefano Zampini row = nrank; 6623b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 6624b0c7d250SStefano Zampini cols = adjncy; 6625b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 6626b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 6627b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 6628b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6629b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 663052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 663152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 663252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6633b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 663427b6a85dSStefano Zampini if (use_vwgt) { 663527b6a85dSStefano Zampini Vec v; 663627b6a85dSStefano Zampini const PetscScalar *array; 663727b6a85dSStefano Zampini PetscInt nl; 663827b6a85dSStefano Zampini 663927b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 6640bb360cb4SStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)n,INSERT_VALUES);CHKERRQ(ierr); 664127b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 664227b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 664327b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 664427b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 664527b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 664622db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 664727b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 664827b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 664927b6a85dSStefano Zampini } 6650b0c7d250SStefano Zampini } else { 665127b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 665227b6a85dSStefano Zampini if (use_vwgt) { 665327b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 6654bb360cb4SStefano Zampini v_wgt[0] = n; 665527b6a85dSStefano Zampini } 6656b0c7d250SStefano Zampini } 665722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 6658e7931f94SStefano Zampini 6659e7931f94SStefano Zampini /* Partition */ 666027b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 6661e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 666227b6a85dSStefano Zampini if (v_wgt) { 6663e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 6664c8587f34SStefano Zampini } 666557de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 666657de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 6667e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 6668e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 666922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 6670e7931f94SStefano Zampini 667152e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 66726583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 667352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 667452e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 667557de7509SStefano Zampini if (!aggregate) { 667657de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 667727b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 667827b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 667927b6a85dSStefano Zampini #endif 668057de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 668127b6a85dSStefano Zampini } else if (oldranks) { 6682b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 668327b6a85dSStefano Zampini } else { 668427b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 668557de7509SStefano Zampini } 668628143c3dSStefano Zampini } else { 6687b0c7d250SStefano Zampini PetscInt idxs[1]; 6688b0c7d250SStefano Zampini PetscMPIInt tag; 6689b0c7d250SStefano Zampini MPI_Request *reqs; 6690b0c7d250SStefano Zampini 6691b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 6692b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 6693b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 669427b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 669528143c3dSStefano Zampini } 669627b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 6697b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6698b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 669957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 670027b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 670127b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 670227b6a85dSStefano Zampini #endif 670357de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 670427b6a85dSStefano Zampini } else if (oldranks) { 6705b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 670627b6a85dSStefano Zampini } else { 670727b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 6708e7931f94SStefano Zampini } 670957de7509SStefano Zampini } 671052e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6711e7931f94SStefano Zampini /* clean up */ 6712e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 671352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 6714e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 6715e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 6716e7931f94SStefano Zampini } 671727b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 671857de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 6719e7931f94SStefano Zampini 6720e7931f94SStefano Zampini /* assemble parallel IS for sends */ 6721e7931f94SStefano Zampini i = 1; 672227b6a85dSStefano Zampini if (!color) i=0; 672357de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 6724e7931f94SStefano Zampini PetscFunctionReturn(0); 6725e7931f94SStefano Zampini } 6726e7931f94SStefano Zampini 6727e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 6728e7931f94SStefano Zampini 67291e0482f5SStefano 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[]) 6730e7931f94SStefano Zampini { 673170cf5478SStefano Zampini Mat local_mat; 6732e7931f94SStefano Zampini IS is_sends_internal; 67339d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 67341ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 67359d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 6736e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 6737e7931f94SStefano Zampini PetscInt* l2gmap_indices; 6738e7931f94SStefano Zampini const PetscInt* is_indices; 6739e7931f94SStefano Zampini MatType new_local_type; 6740e7931f94SStefano Zampini /* buffers */ 6741e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 674228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 67439d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 6744e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 67451ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 6746e7931f94SStefano Zampini /* MPI */ 674728143c3dSStefano Zampini MPI_Comm comm,comm_n; 674828143c3dSStefano Zampini PetscSubcomm subcomm; 6749e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 675028143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 675128143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 67521ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 67531ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 67541ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 6755e7931f94SStefano Zampini PetscErrorCode ierr; 6756e7931f94SStefano Zampini 6757e7931f94SStefano Zampini PetscFunctionBegin; 675857de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 6759e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 6760fbfcfee5SBarry Smith if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME); 676157de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 676257de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 676357de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 676457de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 676557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 67661ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 67671ae86dd6SStefano Zampini if (nvecs) { 67681ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 67691ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 67701ae86dd6SStefano Zampini } 677157de7509SStefano Zampini /* further checks */ 6772e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6773e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 6774e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 6775e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 6776e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 677757de7509SStefano Zampini if (reuse && *mat_n) { 677870cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 677957de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 678070cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 678128143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 678270cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 678370cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 678470cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 678570cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 678670cf5478SStefano Zampini } 6787e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 6788e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 678957de7509SStefano Zampini 6790e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 6791e7931f94SStefano Zampini if (!is_sends) { 679228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 6793bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 6794c8587f34SStefano Zampini } else { 6795e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 6796e7931f94SStefano Zampini is_sends_internal = is_sends; 6797c8587f34SStefano Zampini } 6798e7931f94SStefano Zampini 6799e7931f94SStefano Zampini /* get comm */ 6800a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 6801e7931f94SStefano Zampini 6802e7931f94SStefano Zampini /* compute number of sends */ 6803e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 6804e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 6805e7931f94SStefano Zampini 6806e7931f94SStefano Zampini /* compute number of receives */ 6807e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 6808785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 6809e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 6810e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6811e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 6812e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 6813e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 6814e7931f94SStefano Zampini 681528143c3dSStefano Zampini /* restrict comm if requested */ 681628143c3dSStefano Zampini subcomm = 0; 681728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 681828143c3dSStefano Zampini if (restrict_comm) { 6819779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 6820779c1cceSStefano Zampini 682128143c3dSStefano Zampini color = 0; 682253a05cb3SStefano Zampini if (restrict_full) { 682353a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 682453a05cb3SStefano Zampini } else { 682553a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 682653a05cb3SStefano Zampini } 6827b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 682828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 682928143c3dSStefano Zampini /* check if reuse has been requested */ 683057de7509SStefano Zampini if (reuse) { 683128143c3dSStefano Zampini if (*mat_n) { 683228143c3dSStefano Zampini PetscMPIInt subcommsize2; 683328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 683428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 683528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 683628143c3dSStefano Zampini } else { 683728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 683828143c3dSStefano Zampini } 683928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 6840779c1cceSStefano Zampini PetscMPIInt rank; 6841779c1cceSStefano Zampini 6842779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 684328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 684428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 684528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 6846306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 684728143c3dSStefano Zampini } 684828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 684928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 685028143c3dSStefano Zampini } else { 685128143c3dSStefano Zampini comm_n = comm; 685228143c3dSStefano Zampini } 685328143c3dSStefano Zampini 6854e7931f94SStefano Zampini /* prepare send/receive buffers */ 6855785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 6856e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 6857785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 6858e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 685928143c3dSStefano Zampini if (nis) { 6860854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 686128143c3dSStefano Zampini } 6862e7931f94SStefano Zampini 686328143c3dSStefano Zampini /* Get data from local matrices */ 68646c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 6865e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 6866e7931f94SStefano Zampini /* 6867e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 6868e7931f94SStefano Zampini send_buffer_idxs should contain: 6869e7931f94SStefano Zampini - MatType_PRIVATE type 6870e7931f94SStefano Zampini - PetscInt size_of_l2gmap 6871e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 6872e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 6873e7931f94SStefano Zampini */ 68746c4ed002SBarry Smith else { 6875e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 68763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 6877854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 6878e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 6879e7931f94SStefano Zampini send_buffer_idxs[1] = i; 68803bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6881e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 68823bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6883e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 6884e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6885e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 6886e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 6887c8587f34SStefano Zampini } 6888c8587f34SStefano Zampini } 6889e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 689028143c3dSStefano Zampini /* additional is (if any) */ 689128143c3dSStefano Zampini if (nis) { 689228143c3dSStefano Zampini PetscMPIInt psum; 689328143c3dSStefano Zampini PetscInt j; 689428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 689528143c3dSStefano Zampini PetscInt plen; 689628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 689728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 689828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 689928143c3dSStefano Zampini } 6900854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 690128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 690228143c3dSStefano Zampini PetscInt plen; 690328143c3dSStefano Zampini const PetscInt *is_array_idxs; 690428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 690528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 690628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 690728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 690828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 690928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 691028143c3dSStefano Zampini } 691128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 691228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 691328143c3dSStefano Zampini } 691428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 691528143c3dSStefano Zampini } 691628143c3dSStefano Zampini 6917e7931f94SStefano Zampini buf_size_idxs = 0; 6918e7931f94SStefano Zampini buf_size_vals = 0; 691928143c3dSStefano Zampini buf_size_idxs_is = 0; 69201ae86dd6SStefano Zampini buf_size_vecs = 0; 6921e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6922e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 6923e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 692428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 69251ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 6926e7931f94SStefano Zampini } 6927785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 6928785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 692995ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 69301ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 6931e7931f94SStefano Zampini 6932e7931f94SStefano Zampini /* get new tags for clean communications */ 6933e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 6934e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 693528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 69361ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 6937e7931f94SStefano Zampini 6938e7931f94SStefano Zampini /* allocate for requests */ 6939785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 6940785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 694195ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 69421ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 6943785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 6944785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 694595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 69461ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 6947e7931f94SStefano Zampini 6948e7931f94SStefano Zampini /* communications */ 6949e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6950e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 695128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 69521ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 6953e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6954e7931f94SStefano Zampini source_dest = onodes[i]; 6955e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 6956e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 6957e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6958e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 695928143c3dSStefano Zampini if (nis) { 696057de7509SStefano Zampini source_dest = onodes_is[i]; 696128143c3dSStefano 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); 696228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 696328143c3dSStefano Zampini } 69641ae86dd6SStefano Zampini if (nvecs) { 69651ae86dd6SStefano Zampini source_dest = onodes[i]; 69661ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 69671ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 69681ae86dd6SStefano Zampini } 6969e7931f94SStefano Zampini } 6970e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6971e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 6972e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 6973e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 697428143c3dSStefano Zampini if (nis) { 697528143c3dSStefano 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); 697628143c3dSStefano Zampini } 69771ae86dd6SStefano Zampini if (nvecs) { 69781ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 69791ae86dd6SStefano 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); 69801ae86dd6SStefano Zampini } 6981e7931f94SStefano Zampini } 6982e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6983e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 6984e7931f94SStefano Zampini 6985e7931f94SStefano Zampini /* assemble new l2g map */ 6986e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6987e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 69889d30be91SStefano Zampini new_local_rows = 0; 6989e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 69909d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6991e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6992e7931f94SStefano Zampini } 69939d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 6994e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 69959d30be91SStefano Zampini new_local_rows = 0; 6996e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 69979d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 69989d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6999e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7000e7931f94SStefano Zampini } 70019d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 70029d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 7003e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 7004e7931f94SStefano Zampini 7005e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 7006e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 7007e7931f94SStefano 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) */ 7008e7931f94SStefano Zampini if (n_recvs) { 700928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 7010e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 7011e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7012e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 7013e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 7014e7931f94SStefano Zampini break; 7015e7931f94SStefano Zampini } 7016e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7017e7931f94SStefano Zampini } 7018e7931f94SStefano Zampini switch (new_local_type_private) { 701928143c3dSStefano Zampini case MATDENSE_PRIVATE: 7020e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7021e7931f94SStefano Zampini bs = 1; 7022e7931f94SStefano Zampini break; 7023e7931f94SStefano Zampini case MATAIJ_PRIVATE: 7024e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7025e7931f94SStefano Zampini bs = 1; 7026e7931f94SStefano Zampini break; 7027e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 7028e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 7029e7931f94SStefano Zampini break; 7030e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 7031e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 7032e7931f94SStefano Zampini break; 7033e7931f94SStefano Zampini default: 7034fbfcfee5SBarry Smith SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,PETSC_FUNCTION_NAME); 7035e7931f94SStefano Zampini break; 7036e7931f94SStefano Zampini } 7037ed8ed4edSstefano_zampini } else { /* by default, new_local_type is seqaij */ 7038ed8ed4edSstefano_zampini new_local_type = MATSEQAIJ; 703928143c3dSStefano Zampini bs = 1; 7040e7931f94SStefano Zampini } 7041e7931f94SStefano Zampini 704270cf5478SStefano Zampini /* create MATIS object if needed */ 704357de7509SStefano Zampini if (!reuse) { 7044e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 7045e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 704670cf5478SStefano Zampini } else { 704770cf5478SStefano Zampini /* it also destroys the local matrices */ 704857de7509SStefano Zampini if (*mat_n) { 704970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 705057de7509SStefano Zampini } else { /* this is a fake object */ 705157de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 705257de7509SStefano Zampini } 705370cf5478SStefano Zampini } 705470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 7055e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 70569d30be91SStefano Zampini 70579d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 70589d30be91SStefano Zampini 70599d30be91SStefano Zampini /* Global to local map of received indices */ 70609d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 70619d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 70629d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 70639d30be91SStefano Zampini 70649d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 70659d30be91SStefano Zampini buf_size_idxs = 0; 70669d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 70679d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 70689d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 70699d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 70709d30be91SStefano Zampini } 70719d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 70729d30be91SStefano Zampini 70739d30be91SStefano Zampini /* set preallocation */ 70749d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 70759d30be91SStefano Zampini if (!newisdense) { 70769d30be91SStefano Zampini PetscInt *new_local_nnz=0; 70779d30be91SStefano Zampini 70789d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 70799d30be91SStefano Zampini if (n_recvs) { 70809d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 70819d30be91SStefano Zampini } 70829d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 70839d30be91SStefano Zampini PetscInt j; 70849d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 70859d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 70869d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 70879d30be91SStefano Zampini } 70889d30be91SStefano Zampini } else { 70899d30be91SStefano Zampini /* TODO */ 70909d30be91SStefano Zampini } 70919d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 70929d30be91SStefano Zampini } 70939d30be91SStefano Zampini if (new_local_nnz) { 70949d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 70959d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 70969d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 70979d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 70989d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 70999d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 71009d30be91SStefano Zampini } else { 71019d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 71029d30be91SStefano Zampini } 71039d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 71049d30be91SStefano Zampini } else { 71059d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 71069d30be91SStefano Zampini } 7107e7931f94SStefano Zampini 7108e7931f94SStefano Zampini /* set values */ 7109e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 71109d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 7111e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7112e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 7113e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 71149d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 7115e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7116e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7117e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 711828143c3dSStefano Zampini } else { 711928143c3dSStefano Zampini /* TODO */ 7120e7931f94SStefano Zampini } 7121e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7122e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 7123e7931f94SStefano Zampini } 7124e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7125e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 712670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 712770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 71289d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 7129e7931f94SStefano Zampini 7130dfd14d43SStefano Zampini #if 0 713128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 7132e7931f94SStefano Zampini Vec lvec,rvec; 7133e7931f94SStefano Zampini PetscReal infty_error; 7134e7931f94SStefano Zampini 71352a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 7136e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 7137e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 7138e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 713970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 7140e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 7141e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 7142e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 7143e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 7144e7931f94SStefano Zampini } 714528143c3dSStefano Zampini #endif 7146e7931f94SStefano Zampini 714728143c3dSStefano Zampini /* assemble new additional is (if any) */ 714828143c3dSStefano Zampini if (nis) { 714928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 715028143c3dSStefano Zampini 715128143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7152854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 715328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 715428143c3dSStefano Zampini psum = 0; 715528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 715628143c3dSStefano Zampini for (j=0;j<nis;j++) { 715728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 715828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 715928143c3dSStefano Zampini psum += plen; 716028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 716128143c3dSStefano Zampini } 716228143c3dSStefano Zampini } 7163854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 7164854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 716528143c3dSStefano Zampini for (i=1;i<nis;i++) { 716628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 716728143c3dSStefano Zampini } 716828143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 716928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 717028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 717128143c3dSStefano Zampini for (j=0;j<nis;j++) { 717228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 717328143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 717428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 717528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 717628143c3dSStefano Zampini } 717728143c3dSStefano Zampini } 717828143c3dSStefano Zampini for (i=0;i<nis;i++) { 717928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 718028143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 718128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 718228143c3dSStefano Zampini } 718328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 718428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 718528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 718628143c3dSStefano Zampini } 7187e7931f94SStefano Zampini /* free workspace */ 718828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 7189e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7190e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 7191e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7192e7931f94SStefano Zampini if (isdense) { 7193e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 7194e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 7195e7931f94SStefano Zampini } else { 7196e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 7197e7931f94SStefano Zampini } 719828143c3dSStefano Zampini if (nis) { 719928143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 720028143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 720128143c3dSStefano Zampini } 72021ae86dd6SStefano Zampini 72031ae86dd6SStefano Zampini if (nvecs) { 72041ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 72051ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 72061ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72071ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 72081ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 72091ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 72101ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 72111ae86dd6SStefano Zampini /* set values */ 72121ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 72131ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 72141ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72151ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 72161ae86dd6SStefano Zampini PetscInt j; 72171ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 72181ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 72191ae86dd6SStefano Zampini } 72201ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 72211ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 72221ae86dd6SStefano Zampini } 72231ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72241ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 72251ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 72261ae86dd6SStefano Zampini } 72271ae86dd6SStefano Zampini 72281ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 72291ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 7230e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 7231e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 72321ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 723328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 7234e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 7235e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 72361ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 723728143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 7238e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 7239e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 7240e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 7241e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 7242e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 724328143c3dSStefano Zampini if (nis) { 724428143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 724528143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 724628143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 724728143c3dSStefano Zampini } 724828143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 724928143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 725028143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 725128143c3dSStefano Zampini for (i=0;i<nis;i++) { 725228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 725328143c3dSStefano Zampini } 72541ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 72551ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 72561ae86dd6SStefano Zampini } 725753a05cb3SStefano Zampini *mat_n = NULL; 725828143c3dSStefano Zampini } 7259e7931f94SStefano Zampini PetscFunctionReturn(0); 7260e7931f94SStefano Zampini } 7261a57a6d2fSStefano Zampini 726212edc857SStefano Zampini /* temporary hack into ksp private data structure */ 7263af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 726412edc857SStefano Zampini 7265c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 7266c8587f34SStefano Zampini { 7267c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 7268c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 726920a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 72701ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 72711e0482f5SStefano Zampini Mat coarseG,t_coarse_mat_is; 72729881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 727320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 72746e683305SStefano Zampini IS coarse_is,*isarray; 72756e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 727630368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 7277f9eb5b7dSStefano Zampini PC pc_temp; 7278c8587f34SStefano Zampini PCType coarse_pc_type; 7279c8587f34SStefano Zampini KSPType coarse_ksp_type; 7280f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 72814f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 72821e0482f5SStefano Zampini PetscInt ncoarse,nedcfield; 728368457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 728422bc73bbSStefano Zampini PetscScalar *array; 728557de7509SStefano Zampini MatReuse coarse_mat_reuse; 728657de7509SStefano Zampini PetscBool restr, full_restr, have_void; 72877de4f681Sstefano_zampini PetscMPIInt commsize; 72889881197aSStefano Zampini PetscErrorCode ierr; 7289fdc09c96SStefano Zampini 7290c8587f34SStefano Zampini PetscFunctionBegin; 7291c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 729268457ee5SStefano 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 */ 7293fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 72945a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 72957de4f681Sstefano_zampini 72967de4f681Sstefano_zampini pcbddc->new_primal_space = PETSC_TRUE; 7297fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 7298f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 7299f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 7300f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 7301fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 730251bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 730351bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 7304727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 7305fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7306fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 7307fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7308f4ddd8eeSStefano Zampini } 7309fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 7310fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7311f4ddd8eeSStefano Zampini } 731270cf5478SStefano Zampini /* reset any subassembling information */ 731357de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 731470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 731557de7509SStefano Zampini } 73166e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 7317fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7318f4ddd8eeSStefano Zampini } 731957de7509SStefano Zampini /* assemble coarse matrix */ 732057de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 732157de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 732257de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 732357de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 732418a45a71SStefano Zampini } else { 732557de7509SStefano Zampini coarse_mat = NULL; 732657de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 73276e683305SStefano Zampini } 7328e7931f94SStefano Zampini 7329abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 7330abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 7331abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 7332abbbba34SStefano Zampini 7333abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 733422bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 733522bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 733622bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 733722bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 7338e176bc59SStefano 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); 73396e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 73406e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 73416e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7342abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 7343abbbba34SStefano Zampini 734457de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 734557de7509SStefano Zampini im_active = !!(pcis->n); 734657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 734757de7509SStefano Zampini 734814f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 734957de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 735057de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 73517de4f681Sstefano_zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&commsize);CHKERRQ(ierr); 735257de7509SStefano Zampini coarse_mat_is = NULL; 735357de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 735457de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 73551ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 735657de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 735757de7509SStefano Zampini if (multilevel_requested) { 735857de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 735957de7509SStefano Zampini restr = PETSC_FALSE; 736057de7509SStefano Zampini full_restr = PETSC_FALSE; 736157de7509SStefano Zampini } else { 736257de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 736357de7509SStefano Zampini restr = PETSC_TRUE; 736457de7509SStefano Zampini full_restr = PETSC_TRUE; 736557de7509SStefano Zampini } 73667de4f681Sstefano_zampini if (!pcbddc->coarse_size || commsize == 1) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 736757de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 736857de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 7369a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 7370bb360cb4SStefano Zampini if (multilevel_requested) { 7371bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7372bb360cb4SStefano Zampini } else { 7373bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7374bb360cb4SStefano Zampini } 7375a198735bSStefano Zampini } else { 73767de4f681Sstefano_zampini PetscMPIInt rank; 7377a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 73787de4f681Sstefano_zampini have_void = (active_procs == (PetscInt)commsize) ? PETSC_FALSE : PETSC_TRUE; 7379a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 7380a198735bSStefano Zampini } 738157de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 738257de7509SStefano Zampini PetscInt psum; 738357de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 738457de7509SStefano Zampini else psum = 0; 738557de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 73867de4f681Sstefano_zampini if (ncoarse < commsize) have_void = PETSC_TRUE; 738757de7509SStefano Zampini } 738857de7509SStefano Zampini /* determine if we can go multilevel */ 738957de7509SStefano Zampini if (multilevel_requested) { 739057de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 739157de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 739257de7509SStefano Zampini } 739357de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 739457de7509SStefano Zampini 7395e4d548c7SStefano Zampini /* dump subassembling pattern */ 7396e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 7397e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 7398e4d548c7SStefano Zampini } 7399e4d548c7SStefano Zampini 74006e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 74011e0482f5SStefano Zampini nedcfield = -1; 74021e0482f5SStefano Zampini if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal)) { /* protects from unneded computations */ 74036e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 74046e683305SStefano Zampini const PetscInt *idxs; 74056e683305SStefano Zampini ISLocalToGlobalMapping tmap; 74066e683305SStefano Zampini 74076e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 74080be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 74096e683305SStefano Zampini /* allocate space for temporary storage */ 7410854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 7411854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 74126e683305SStefano Zampini /* allocate for IS array */ 74136e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 74141e0482f5SStefano Zampini if (pcbddc->nedclocal) { 74151e0482f5SStefano Zampini if (pcbddc->nedfield > -1) { 74161e0482f5SStefano Zampini nedcfield = pcbddc->nedfield; 74171e0482f5SStefano Zampini } else { 74181e0482f5SStefano Zampini nedcfield = 0; 74191e0482f5SStefano Zampini if (nisdofs) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen (%d)",nisdofs); 74201e0482f5SStefano Zampini nisdofs = 1; 74211e0482f5SStefano Zampini } 74221e0482f5SStefano Zampini } 74236e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 742427b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 742530368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 7426854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 74276e683305SStefano Zampini /* dofs splitting */ 74286e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 74296e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 74301e0482f5SStefano Zampini if (nedcfield != i) { 74316e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 74326e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 74336e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 74346e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 74351e0482f5SStefano Zampini } else { 74361e0482f5SStefano Zampini ierr = ISGetLocalSize(pcbddc->nedclocal,&tsize);CHKERRQ(ierr); 74371e0482f5SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 74381e0482f5SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 7439eee23b56SStefano Zampini if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping coarse nedelec field! %d != %d\n",tsize,nout); 74401e0482f5SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 74411e0482f5SStefano Zampini } 74426e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 744330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 74446e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 74456e683305SStefano Zampini } 74466e683305SStefano Zampini /* neumann boundaries */ 74476e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 74486e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 74496e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 74506e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 74516e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 74526e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 74536e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 745430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 74556e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 74566e683305SStefano Zampini } 74576e683305SStefano Zampini /* free memory */ 74586e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 74596e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 74606e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 74616e683305SStefano Zampini } else { 74626e683305SStefano Zampini nis = 0; 74636e683305SStefano Zampini nisdofs = 0; 74646e683305SStefano Zampini nisneu = 0; 746530368db7SStefano Zampini nisvert = 0; 74666e683305SStefano Zampini isarray = NULL; 74676e683305SStefano Zampini } 74686e683305SStefano Zampini /* destroy no longer needed map */ 74696e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 74706e683305SStefano Zampini 747157de7509SStefano Zampini /* subassemble */ 747257de7509SStefano Zampini if (multilevel_allowed) { 74731ae86dd6SStefano Zampini Vec vp[1]; 74741ae86dd6SStefano Zampini PetscInt nvecs = 0; 747557de7509SStefano Zampini PetscBool reuse,reuser; 74761ae86dd6SStefano Zampini 747757de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 747857de7509SStefano Zampini else reuse = PETSC_FALSE; 747957de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 74801ae86dd6SStefano Zampini vp[0] = NULL; 74811ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 74821ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 74831ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 74841ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 74851ae86dd6SStefano Zampini nvecs = 1; 74861ae86dd6SStefano Zampini 74871ae86dd6SStefano Zampini if (pcbddc->divudotp) { 7488a198735bSStefano Zampini Mat B,loc_divudotp; 74891ae86dd6SStefano Zampini Vec v,p; 74901ae86dd6SStefano Zampini IS dummy; 74911ae86dd6SStefano Zampini PetscInt np; 74921ae86dd6SStefano Zampini 7493a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 7494a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 74951ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 74967dae84e0SHong Zhang ierr = MatCreateSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 74971ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 74981ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 74991ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 75001ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 75011ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 75021ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 75031ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 75041ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 75051ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 75061ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 75071ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 75081ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 750974e2c79eSStefano Zampini } 75101ae86dd6SStefano Zampini } 75111ae86dd6SStefano Zampini if (reuser) { 75121e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 751374e2c79eSStefano Zampini } else { 75141e0482f5SStefano 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); 75151ae86dd6SStefano Zampini } 75161ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 75171ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 75181ae86dd6SStefano Zampini PetscInt nl; 75191ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 75201ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 75211ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 75221ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 75231ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 75241ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 75251ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 75261ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 7527a198735bSStefano Zampini } else { 7528a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 75291ae86dd6SStefano Zampini } 75301ae86dd6SStefano Zampini } else { 75311e0482f5SStefano 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); 75326e683305SStefano Zampini } 753357de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 753457de7509SStefano Zampini PetscMPIInt size; 7535f913dca9SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size);CHKERRQ(ierr); 753657de7509SStefano Zampini if (!multilevel_allowed) { 753757de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 75386e683305SStefano Zampini } else { 753957de7509SStefano Zampini Mat A; 7540779c1cceSStefano Zampini 754157de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 754257de7509SStefano Zampini if (coarse_mat_is) { 754357de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 754457de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 754557de7509SStefano Zampini coarse_mat = coarse_mat_is; 754657de7509SStefano Zampini } 754757de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 754857de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 754957de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 7550779c1cceSStefano Zampini } 7551779c1cceSStefano Zampini } 755257de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 755357de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 75546e683305SStefano Zampini 75556e683305SStefano Zampini /* create local to global scatters for coarse problem */ 755668457ee5SStefano Zampini if (compute_vecs) { 75576e683305SStefano Zampini PetscInt lrows; 75586e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 755957de7509SStefano Zampini if (coarse_mat) { 756057de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 75616e683305SStefano Zampini } else { 75626e683305SStefano Zampini lrows = 0; 75636e683305SStefano Zampini } 75646e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 75656e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 75666e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 75676e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 75686e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 75696e683305SStefano Zampini } 75706e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 7571c8587f34SStefano Zampini 7572f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 7573f9eb5b7dSStefano Zampini if (multilevel_allowed) { 7574f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 7575f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 7576f9eb5b7dSStefano Zampini } else { 7577f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 7578f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 7579c8587f34SStefano Zampini } 7580c8587f34SStefano Zampini 75816e683305SStefano Zampini /* print some info if requested */ 75826e683305SStefano Zampini if (pcbddc->dbg_flag) { 75836e683305SStefano Zampini if (!multilevel_allowed) { 75846e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 75856e683305SStefano Zampini if (multilevel_requested) { 75866e683305SStefano 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); 75876e683305SStefano Zampini } else if (pcbddc->max_levels) { 75886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 75896e683305SStefano Zampini } 75906e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 75916e683305SStefano Zampini } 75926e683305SStefano Zampini } 75936e683305SStefano Zampini 75941e0482f5SStefano Zampini /* communicate coarse discrete gradient */ 75951e0482f5SStefano Zampini coarseG = NULL; 75961e0482f5SStefano Zampini if (pcbddc->nedcG && multilevel_allowed) { 75971e0482f5SStefano Zampini MPI_Comm ccomm; 75981e0482f5SStefano Zampini if (coarse_mat) { 75991e0482f5SStefano Zampini ccomm = PetscObjectComm((PetscObject)coarse_mat); 76001e0482f5SStefano Zampini } else { 76011e0482f5SStefano Zampini ccomm = MPI_COMM_NULL; 76021e0482f5SStefano Zampini } 76031e0482f5SStefano Zampini ierr = MatMPIAIJRestrict(pcbddc->nedcG,ccomm,&coarseG);CHKERRQ(ierr); 76041e0482f5SStefano Zampini } 76051e0482f5SStefano Zampini 7606f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 760757de7509SStefano Zampini if (coarse_mat) { 76086a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 76096e683305SStefano Zampini if (pcbddc->dbg_flag) { 761057de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 76116e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 76126e683305SStefano Zampini } 7613f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 7614312be037SStefano Zampini char prefix[256],str_level[16]; 7615e604994aSStefano Zampini size_t len; 76161e0482f5SStefano Zampini 761757de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 7618422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 7619c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 7620f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 762157de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7622c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 76236e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 7624c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 76251e0482f5SStefano Zampini /* TODO is this logic correct? should check for coarse_mat type */ 7626c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7627e604994aSStefano Zampini /* prefix */ 7628e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 7629e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 7630e604994aSStefano Zampini if (!pcbddc->current_level) { 7631e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 7632e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 7633c8587f34SStefano Zampini } else { 7634e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 7635312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 7636312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 763734d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 7638312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 7639e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 7640e604994aSStefano Zampini } 7641e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 76423e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 76433e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 76443e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 76453e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 7646f9eb5b7dSStefano Zampini /* allow user customization */ 7647f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 76483e3c6dadSStefano Zampini } 76493e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 765051bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 76513e3c6dadSStefano Zampini if (nisdofs) { 76523e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 76533e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 76543e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 76553e3c6dadSStefano Zampini } 76563e3c6dadSStefano Zampini } 76573e3c6dadSStefano Zampini if (nisneu) { 76583e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 76593e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 7660312be037SStefano Zampini } 766130368db7SStefano Zampini if (nisvert) { 766230368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 766330368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 766430368db7SStefano Zampini } 76651e0482f5SStefano Zampini if (coarseG) { 76661e0482f5SStefano Zampini ierr = PCBDDCSetDiscreteGradient(pc_temp,coarseG,1,nedcfield,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 76671e0482f5SStefano Zampini } 7668f9eb5b7dSStefano Zampini 7669f9eb5b7dSStefano Zampini /* get some info after set from options */ 7670f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 7671f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 76724f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 7673b76f3995Sstefano_zampini /* multilevel can only be requested via -pc_bddc_levels or PCBDDCSetLevels */ 7674b76f3995Sstefano_zampini if (isbddc && !multilevel_allowed) { 7675f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7676f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 7677f9eb5b7dSStefano Zampini } 7678b76f3995Sstefano_zampini /* multilevel cannot be done with coarse PCs different from BDDC or NN */ 7679b76f3995Sstefano_zampini if (multilevel_requested && !isbddc && !isnn) { 7680b76f3995Sstefano_zampini ierr = PCSetType(pc_temp,PCBDDC);CHKERRQ(ierr); 7681b76f3995Sstefano_zampini isbddc = PETSC_TRUE; 7682b76f3995Sstefano_zampini isnn = PETSC_FALSE; 7683b76f3995Sstefano_zampini } 768439f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 76854f3a063dSStefano Zampini if (isredundant) { 76864f3a063dSStefano Zampini KSP inner_ksp; 76874f3a063dSStefano Zampini PC inner_pc; 76889326c5c6Sstefano_zampini 76894f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 76904f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 76914f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 76924f3a063dSStefano Zampini } 7693f9eb5b7dSStefano Zampini 769457de7509SStefano Zampini /* parameters which miss an API */ 769557de7509SStefano Zampini if (isbddc) { 7696720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 7697720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 769857de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 769927b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 770027b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 7701a198735bSStefano Zampini Mat coarsedivudotp_is; 7702a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 7703a198735bSStefano Zampini IS row,col; 7704a198735bSStefano Zampini const PetscInt *gidxs; 7705a198735bSStefano Zampini PetscInt n,st,M,N; 7706a198735bSStefano Zampini 7707a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 7708a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 7709a198735bSStefano Zampini st = st-n; 7710a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 7711a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 7712a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 7713a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7714a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 7715a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7716a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 7717a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 7718a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 7719a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 7720a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 7721a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 7722a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 7723a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 7724a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 7725a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 7726a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 7727a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 7728a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 7729a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 77308ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 7731a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 7732720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 773359e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 7734720d30f9SStefano Zampini } 7735d4d8cf7bSStefano Zampini } 77369881197aSStefano Zampini 77373301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 77385a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 77393301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 77403301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 77413301b35fSStefano Zampini } 77423301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 77433301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 77443301b35fSStefano Zampini } 77453301b35fSStefano Zampini if (pc->pmat->spd_set) { 77463301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 77473301b35fSStefano Zampini } 774827b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 774927b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 775027b6a85dSStefano Zampini } 77516e683305SStefano Zampini /* set operators */ 77525f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 77536e683305SStefano Zampini if (pcbddc->dbg_flag) { 77546e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 77556e683305SStefano Zampini } 77566e683305SStefano Zampini } 77571e0482f5SStefano Zampini ierr = MatDestroy(&coarseG);CHKERRQ(ierr); 77586e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 7759b1ecc7b1SStefano Zampini #if 0 7760b9b85e73SStefano Zampini { 7761b9b85e73SStefano Zampini PetscViewer viewer; 7762b9b85e73SStefano Zampini char filename[256]; 7763b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 7764b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 77656a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 7766b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 7767f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 7768b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 7769b9b85e73SStefano Zampini } 7770b9b85e73SStefano Zampini #endif 7771f9eb5b7dSStefano Zampini 777298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 777398a51de6SStefano Zampini Vec crhs,csol; 777404708bb6SStefano Zampini 7775f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 7776f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 7777f347579bSStefano Zampini if (!csol) { 77782a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 7779f9eb5b7dSStefano Zampini } 7780f347579bSStefano Zampini if (!crhs) { 77812a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 7782f347579bSStefano Zampini } 7783b0f5fe93SStefano Zampini } 77841ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 7785b0f5fe93SStefano Zampini 7786b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 7787b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 7788b0f5fe93SStefano Zampini 7789b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 77904f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 77914f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 77924f1b2e48SStefano Zampini } 7793b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 7794b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 7795b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7796b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7797b0f5fe93SStefano Zampini if (coarse_mat) { 7798b0f5fe93SStefano Zampini Vec nullv; 7799b0f5fe93SStefano Zampini PetscScalar *array,*array2; 7800b0f5fe93SStefano Zampini PetscInt nl; 7801b0f5fe93SStefano Zampini 7802b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 7803b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 7804b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7805b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 7806b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 7807b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 7808b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7809b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 7810b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 7811b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 7812b0f5fe93SStefano Zampini } 7813b0f5fe93SStefano Zampini } 7814b0f5fe93SStefano Zampini 7815b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 7816b0f5fe93SStefano Zampini PetscBool ispreonly; 7817b0f5fe93SStefano Zampini 7818b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7819b0f5fe93SStefano Zampini PetscBool isnull; 7820b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 7821bef83e63SStefano Zampini if (isnull) { 7822b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 7823b0f5fe93SStefano Zampini } 7824bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 7825b0f5fe93SStefano Zampini } 7826b0f5fe93SStefano Zampini /* setup coarse ksp */ 7827b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 7828cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 7829cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 78306e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 7831c8587f34SStefano Zampini KSP check_ksp; 78322b510759SStefano Zampini KSPType check_ksp_type; 7833c8587f34SStefano Zampini PC check_pc; 78346e683305SStefano Zampini Vec check_vec,coarse_vec; 78356a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 78362b510759SStefano Zampini PetscInt its; 78376e683305SStefano Zampini PetscBool compute_eigs; 78386e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 78396e683305SStefano Zampini PetscInt neigs; 78408e185a42SStefano Zampini const char *prefix; 7841c8587f34SStefano Zampini 78422b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 78436e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 7844422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 784523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7846f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 7847e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 7848e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 7849e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 78502b510759SStefano Zampini if (ispreonly) { 78512b510759SStefano Zampini check_ksp_type = KSPPREONLY; 78526e683305SStefano Zampini compute_eigs = PETSC_FALSE; 78532b510759SStefano Zampini } else { 7854cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 78556e683305SStefano Zampini compute_eigs = PETSC_TRUE; 7856c8587f34SStefano Zampini } 7857c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 78586e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 78596e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 78606e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 7861a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 7862a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 7863a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 7864a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 7865c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 7866c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 7867c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 7868c8587f34SStefano Zampini /* create random vec */ 78692701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 7870c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 78716e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 7872c8587f34SStefano Zampini /* solve coarse problem */ 78736e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 7874cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 78756e683305SStefano Zampini if (compute_eigs) { 7876854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 7877854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 78786e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 78791ae86dd6SStefano Zampini if (neigs) { 78806e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 78816e683305SStefano Zampini lambda_min = eigs_r[0]; 78826e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 78832701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 78842701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 7885cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 7886cbcc2c2aSStefano Zampini } 7887c8587f34SStefano Zampini } 7888c8587f34SStefano Zampini } 78891ae86dd6SStefano Zampini } 7890cbcc2c2aSStefano Zampini 7891c8587f34SStefano Zampini /* check coarse problem residual error */ 78926e683305SStefano Zampini if (pcbddc->dbg_flag) { 78936e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 78946e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 78956e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 7896c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 78976e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 78986e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 7899779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 79006e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 79016e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 79026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 79036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 7904b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7905b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 7906b0f5fe93SStefano Zampini } 79076e683305SStefano Zampini if (compute_eigs) { 79086e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 7909b03ebc13SStefano Zampini KSPConvergedReason reason; 7910deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 7911c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 7912b03ebc13SStefano Zampini ierr = KSPGetConvergedReason(check_ksp,&reason);CHKERRQ(ierr); 79136e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 7914b03ebc13SStefano 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); 79156e683305SStefano Zampini for (i=0;i<neigs;i++) { 79166e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 7917c8587f34SStefano Zampini } 79186e683305SStefano Zampini } 79196e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 79206e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 79216e683305SStefano Zampini } 7922e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 79232701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 7924c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 79256e683305SStefano Zampini if (compute_eigs) { 79266e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 79276e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 7928c8587f34SStefano Zampini } 79296e683305SStefano Zampini } 79306e683305SStefano Zampini } 7931bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 7932cbcc2c2aSStefano Zampini /* print additional info */ 7933cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 79346e683305SStefano Zampini /* waits until all processes reaches this point */ 79356e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 7936cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 7937cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7938cbcc2c2aSStefano Zampini } 7939cbcc2c2aSStefano Zampini 79402b510759SStefano Zampini /* free memory */ 7941fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 7942c8587f34SStefano Zampini PetscFunctionReturn(0); 7943c8587f34SStefano Zampini } 7944674ae819SStefano Zampini 7945f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 7946f34684f1SStefano Zampini { 7947f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7948f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 7949f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 7950dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 7951dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 795273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 7953dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 7954f34684f1SStefano Zampini PetscErrorCode ierr; 7955f34684f1SStefano Zampini 7956f34684f1SStefano Zampini PetscFunctionBegin; 7957f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 79586c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 7959dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 79603bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 7961dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7962dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 79636583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 7964dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 7965dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 7966dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 79676c4ed002SBarry 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); 7968dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 7969dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7970dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 7971dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7972dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7973f34684f1SStefano Zampini 7974f34684f1SStefano Zampini /* check numbering */ 7975f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 7976019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 7977dc456d91SStefano Zampini PetscInt i; 7978b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 7979f34684f1SStefano Zampini 7980f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7981f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 7982f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 79831575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7984019a44ceSStefano Zampini /* counter */ 7985019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7986019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 7987019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7988019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7989019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7990019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7991f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 7992f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 7993727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 7994f34684f1SStefano Zampini } 7995f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 7996f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 7997f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7998e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7999e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8000e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8001e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8002f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8003019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8004f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8005019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 80062c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 800775c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 8008b9b85e73SStefano Zampini set_error = PETSC_TRUE; 80092c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 80102c66d082SStefano 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); 8011f34684f1SStefano Zampini } 8012f34684f1SStefano Zampini } 8013019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8014b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 8015f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8016f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8017f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 8018f34684f1SStefano Zampini } 8019f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8020f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8021e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8022e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8023f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 8024f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 8025b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 8026ca8b9ea9SStefano Zampini PetscInt *gidxs; 8027ca8b9ea9SStefano Zampini 8028ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 80293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 8030f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 8031f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8032f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 8033f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 80344bc2dc4bSStefano 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); 8035f34684f1SStefano Zampini } 8036f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8037ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 8038f34684f1SStefano Zampini } 8039f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 80401575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8041302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 8042f34684f1SStefano Zampini } 80438bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 8044f34684f1SStefano Zampini /* get back data */ 8045f34684f1SStefano Zampini *coarse_size_n = coarse_size; 8046f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 8047674ae819SStefano Zampini PetscFunctionReturn(0); 8048674ae819SStefano Zampini } 8049674ae819SStefano Zampini 8050a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 8051e456f2a8SStefano Zampini { 8052e456f2a8SStefano Zampini IS localis_t; 8053a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 8054e456f2a8SStefano Zampini PetscScalar *vals; 8055e456f2a8SStefano Zampini PetscErrorCode ierr; 8056e456f2a8SStefano Zampini 8057e456f2a8SStefano Zampini PetscFunctionBegin; 8058a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 8059e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 8060854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 8061e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 8062e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 8063a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 8064a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 80651035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 8066a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 80671035eff8SStefano Zampini } 8068a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 8069e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 8070e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 8071a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 8072a7dc3881SStefano Zampini /* now compute set in local ordering */ 8073a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8074a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8075a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8076a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 8077a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8078ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8079e456f2a8SStefano Zampini lsize++; 8080e456f2a8SStefano Zampini } 8081e456f2a8SStefano Zampini } 8082854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 8083a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8084ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8085e456f2a8SStefano Zampini idxs[lsize++] = i; 8086e456f2a8SStefano Zampini } 8087e456f2a8SStefano Zampini } 8088a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8089a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 8090e456f2a8SStefano Zampini *localis = localis_t; 8091e456f2a8SStefano Zampini PetscFunctionReturn(0); 8092e456f2a8SStefano Zampini } 8093906d46d4SStefano Zampini 809408122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 8095b96c3477SStefano Zampini { 8096a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8097b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8098b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 8099a64f4aa4SStefano Zampini Mat S_j; 8100b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 8101b96c3477SStefano Zampini PetscBool free_used_adj; 8102b96c3477SStefano Zampini PetscErrorCode ierr; 8103b96c3477SStefano Zampini 8104b96c3477SStefano Zampini PetscFunctionBegin; 8105b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 8106b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 810708122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 8108b96c3477SStefano Zampini used_xadj = NULL; 8109b96c3477SStefano Zampini used_adjncy = NULL; 8110b96c3477SStefano Zampini } else { 811108122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 811208122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 811308122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 811408122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 8115b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 8116b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 8117b96c3477SStefano Zampini } else { 81182fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 8119b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 8120b96c3477SStefano Zampini PetscInt nvtxs; 8121b96c3477SStefano Zampini 81222fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 81232fffb893SStefano Zampini if (flg_row) { 8124b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 8125b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 8126b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 8127b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 81282fffb893SStefano Zampini } else { 81292fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 81302fffb893SStefano Zampini used_xadj = NULL; 81312fffb893SStefano Zampini used_adjncy = NULL; 81322fffb893SStefano Zampini } 81332fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 8134b96c3477SStefano Zampini } 8135b96c3477SStefano Zampini } 8136d5574798SStefano Zampini 8137d5574798SStefano Zampini /* setup sub_schurs data */ 8138a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8139df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 8140df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 8141a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 814291af6908SStefano 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); 8143a64f4aa4SStefano Zampini } else { 81446816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 81454d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 8146a3df083aSStefano Zampini PetscInt benign_n; 814772b8c272SStefano Zampini Mat change = NULL; 81489d54b7f4SStefano Zampini Vec scaling = NULL; 814972b8c272SStefano Zampini IS change_primal = NULL; 8150a3df083aSStefano Zampini 81515feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 81525feab87aSStefano Zampini PetscInt n_vertices; 81535feab87aSStefano Zampini 81545feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 81552034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 81565feab87aSStefano Zampini } 815704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 815804708bb6SStefano Zampini if (!isseqaij) { 815904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 816004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 816104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 816204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 816304708bb6SStefano Zampini } else { 8164511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 816504708bb6SStefano Zampini } 816604708bb6SStefano Zampini } 8167a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 8168a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 8169ca92afb2SStefano Zampini } else { 8170a3df083aSStefano Zampini benign_n = 0; 8171ca92afb2SStefano Zampini } 8172b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 8173b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 8174b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 817572b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 817622db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 8177b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 817822db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 8179b7ab4a40SStefano Zampini } 8180b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 8181b7ab4a40SStefano 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 */ 8182b7ab4a40SStefano Zampini if (need_change) { 818388c03ad3SStefano Zampini PC_IS *pcisf; 818488c03ad3SStefano Zampini PC_BDDC *pcbddcf; 818588c03ad3SStefano Zampini PC pcf; 818688c03ad3SStefano Zampini 8187e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 818888c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 818988c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 819088c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 8191b9be95fcSstefano_zampini 819288c03ad3SStefano Zampini /* hacks */ 819388c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 819472b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 819572b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 819672b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 819772b8c272SStefano Zampini pcisf->n = pcis->n; 819872b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 819988c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 820088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 820188c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 820288c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 820388c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 820488c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 820572b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 820688c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 8207b9be95fcSstefano_zampini 8208b9be95fcSstefano_zampini /* setup constraints so that we can get information on primal vertices and change of basis (in local numbering) */ 820988c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 821072b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 821172b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 821272b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 821372b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 8214b9be95fcSstefano_zampini 821588c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 821672b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 821788c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 821888c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 821988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 822088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 822188c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 822288c03ad3SStefano Zampini pcf->ops->destroy = NULL; 8223b9be95fcSstefano_zampini pcf->ops->reset = NULL; 822488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 822588c03ad3SStefano Zampini } 82269d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 822791af6908SStefano 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); 822872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 822972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 8230ca92afb2SStefano Zampini } 8231d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8232b96c3477SStefano Zampini 8233b96c3477SStefano Zampini /* free adjacency */ 8234b96c3477SStefano Zampini if (free_used_adj) { 8235b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 8236b96c3477SStefano Zampini } 8237b96c3477SStefano Zampini PetscFunctionReturn(0); 8238b96c3477SStefano Zampini } 8239b96c3477SStefano Zampini 824008122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 8241b96c3477SStefano Zampini { 8242b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8243b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8244b96c3477SStefano Zampini PCBDDCGraph graph; 8245b96c3477SStefano Zampini PetscErrorCode ierr; 8246b96c3477SStefano Zampini 8247b96c3477SStefano Zampini PetscFunctionBegin; 8248b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 824908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 82503301b35fSStefano Zampini IS verticesIS,verticescomm; 82513301b35fSStefano Zampini PetscInt vsize,*idxs; 8252b96c3477SStefano Zampini 8253b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 82543301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 82553301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 82563301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 82573301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 8258c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 8259b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 8260be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 8261441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 82623301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 8263b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 8264b96c3477SStefano Zampini } else { 8265b96c3477SStefano Zampini graph = pcbddc->mat_graph; 8266b96c3477SStefano Zampini } 8267e4d548c7SStefano Zampini /* print some info */ 82685c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 8269e4d548c7SStefano Zampini IS vertices; 8270e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 8271c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 8272e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8273e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 8274e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8275e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 8276e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 8277e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 8278e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 8279e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8280e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8281c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8282e4d548c7SStefano Zampini } 8283b96c3477SStefano Zampini 8284b96c3477SStefano Zampini /* sub_schurs init */ 8285b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 8286b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 8287b334f244SStefano Zampini } 82888b6046baSStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 8289e62b6521Sstefano_zampini pcbddc->sub_schurs->prefix = ((PetscObject)pc)->prefix; 8290a64f4aa4SStefano Zampini 8291b96c3477SStefano Zampini /* free graph struct */ 829208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 8293b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 8294b96c3477SStefano Zampini } 8295b96c3477SStefano Zampini PetscFunctionReturn(0); 8296b96c3477SStefano Zampini } 8297fa34dd3eSStefano Zampini 8298fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 8299fa34dd3eSStefano Zampini { 8300fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8301fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8302fa34dd3eSStefano Zampini PetscErrorCode ierr; 8303fa34dd3eSStefano Zampini 8304fa34dd3eSStefano Zampini PetscFunctionBegin; 8305fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 8306fa34dd3eSStefano Zampini IS zerodiag = NULL; 83074f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 8308fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 83094f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 831075c01103SStefano Zampini PetscReal norm; 8311fa34dd3eSStefano Zampini PetscInt i; 8312fa34dd3eSStefano Zampini 8313fa34dd3eSStefano Zampini /* B0 and B0_B */ 8314fa34dd3eSStefano Zampini if (zerodiag) { 8315fa34dd3eSStefano Zampini IS dummy; 8316fa34dd3eSStefano Zampini 83174f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 83187dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 8319fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 8320fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 8321fa34dd3eSStefano Zampini } 8322fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 8323fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 8324fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 8325fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8326fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8327fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8328fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8329fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 8330fa34dd3eSStefano Zampini /* S_j */ 8331fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8332fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 8333fa34dd3eSStefano Zampini 8334fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 8335fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 8336fa34dd3eSStefano Zampini /* continuous in primal space */ 8337fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 8338fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8339fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8340fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 83414f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 83424f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 8343fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8344fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8345fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8346fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8347fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8348fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8349fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 8350fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 8351fa34dd3eSStefano Zampini 8352fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 8353fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 8354fa34dd3eSStefano Zampini /* local with Schur */ 8355fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 8356fa34dd3eSStefano Zampini if (zerodiag) { 8357fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 83584f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 8359fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8360fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 8361fa34dd3eSStefano Zampini } 8362fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 8363fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8364fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8365fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8366fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8367fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 8368fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8369fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8370fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 8371fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8372fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8373fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8374fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8375fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8376fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 8377fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 8378fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8379fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8380fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8381fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8382fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8383fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8384fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 8385fa34dd3eSStefano Zampini if (zerodiag) { 8386fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 8387fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 83884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 8389fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8390fa34dd3eSStefano Zampini } 8391fa34dd3eSStefano Zampini /* BDDC */ 8392fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 8393fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 8394fa34dd3eSStefano Zampini 8395fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 8396fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 8397fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 8398fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 83994f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 84004f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 8401fa34dd3eSStefano Zampini } 84024f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 8403fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 8404fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 8405fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 8406fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8407fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 8408fa34dd3eSStefano Zampini } 8409fa34dd3eSStefano Zampini PetscFunctionReturn(0); 8410fa34dd3eSStefano Zampini } 84111e0482f5SStefano Zampini 84121e0482f5SStefano Zampini #include <../src/mat/impls/aij/mpi/mpiaij.h> 84131e0482f5SStefano Zampini PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B) 84141e0482f5SStefano Zampini { 84151e0482f5SStefano Zampini Mat At; 84161e0482f5SStefano Zampini IS rows; 84171e0482f5SStefano Zampini PetscInt rst,ren; 84181e0482f5SStefano Zampini PetscErrorCode ierr; 84191e0482f5SStefano Zampini PetscLayout rmap; 84201e0482f5SStefano Zampini 84211e0482f5SStefano Zampini PetscFunctionBegin; 84221e0482f5SStefano Zampini rst = ren = 0; 84231e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 84241e0482f5SStefano Zampini ierr = PetscLayoutCreate(ccomm,&rmap);CHKERRQ(ierr); 84251e0482f5SStefano Zampini ierr = PetscLayoutSetSize(rmap,A->rmap->N);CHKERRQ(ierr); 84261e0482f5SStefano Zampini ierr = PetscLayoutSetBlockSize(rmap,1);CHKERRQ(ierr); 84271e0482f5SStefano Zampini ierr = PetscLayoutSetUp(rmap);CHKERRQ(ierr); 84281e0482f5SStefano Zampini ierr = PetscLayoutGetRange(rmap,&rst,&ren);CHKERRQ(ierr); 84291e0482f5SStefano Zampini } 8430e07686f2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)A),ren-rst,rst,1,&rows);CHKERRQ(ierr); 84317dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,rows,NULL,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 84321e0482f5SStefano Zampini ierr = ISDestroy(&rows);CHKERRQ(ierr); 84331e0482f5SStefano Zampini 84341e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 84351e0482f5SStefano Zampini Mat_MPIAIJ *a,*b; 84361e0482f5SStefano Zampini IS from,to; 84371e0482f5SStefano Zampini Vec gvec; 84381e0482f5SStefano Zampini PetscInt lsize; 84391e0482f5SStefano Zampini 84401e0482f5SStefano Zampini ierr = MatCreate(ccomm,B);CHKERRQ(ierr); 84411e0482f5SStefano Zampini ierr = MatSetSizes(*B,ren-rst,PETSC_DECIDE,PETSC_DECIDE,At->cmap->N);CHKERRQ(ierr); 84421e0482f5SStefano Zampini ierr = MatSetType(*B,MATAIJ);CHKERRQ(ierr); 84431e0482f5SStefano Zampini ierr = PetscLayoutDestroy(&((*B)->rmap));CHKERRQ(ierr); 84441e0482f5SStefano Zampini ierr = PetscLayoutSetUp((*B)->cmap);CHKERRQ(ierr); 84451e0482f5SStefano Zampini a = (Mat_MPIAIJ*)At->data; 84461e0482f5SStefano Zampini b = (Mat_MPIAIJ*)(*B)->data; 84471e0482f5SStefano Zampini ierr = MPI_Comm_size(ccomm,&b->size);CHKERRQ(ierr); 84481e0482f5SStefano Zampini ierr = MPI_Comm_rank(ccomm,&b->rank);CHKERRQ(ierr); 84491e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->A);CHKERRQ(ierr); 84501e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->B);CHKERRQ(ierr); 84511e0482f5SStefano Zampini b->A = a->A; 84521e0482f5SStefano Zampini b->B = a->B; 84531e0482f5SStefano Zampini 84541e0482f5SStefano Zampini b->donotstash = a->donotstash; 84551e0482f5SStefano Zampini b->roworiented = a->roworiented; 84561e0482f5SStefano Zampini b->rowindices = 0; 84571e0482f5SStefano Zampini b->rowvalues = 0; 84581e0482f5SStefano Zampini b->getrowactive = PETSC_FALSE; 84591e0482f5SStefano Zampini 84601e0482f5SStefano Zampini (*B)->rmap = rmap; 84611e0482f5SStefano Zampini (*B)->factortype = A->factortype; 84621e0482f5SStefano Zampini (*B)->assembled = PETSC_TRUE; 84631e0482f5SStefano Zampini (*B)->insertmode = NOT_SET_VALUES; 84641e0482f5SStefano Zampini (*B)->preallocated = PETSC_TRUE; 84651e0482f5SStefano Zampini 84661e0482f5SStefano Zampini if (a->colmap) { 84671e0482f5SStefano Zampini #if defined(PETSC_USE_CTABLE) 84681e0482f5SStefano Zampini ierr = PetscTableCreateCopy(a->colmap,&b->colmap);CHKERRQ(ierr); 84691e0482f5SStefano Zampini #else 84701e0482f5SStefano Zampini ierr = PetscMalloc1(At->cmap->N,&b->colmap);CHKERRQ(ierr); 84711e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)*B,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 84721e0482f5SStefano Zampini ierr = PetscMemcpy(b->colmap,a->colmap,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 84731e0482f5SStefano Zampini #endif 84741e0482f5SStefano Zampini } else b->colmap = 0; 84751e0482f5SStefano Zampini if (a->garray) { 84761e0482f5SStefano Zampini PetscInt len; 84771e0482f5SStefano Zampini len = a->B->cmap->n; 84781e0482f5SStefano Zampini ierr = PetscMalloc1(len+1,&b->garray);CHKERRQ(ierr); 84791e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)(*B),len*sizeof(PetscInt));CHKERRQ(ierr); 84801e0482f5SStefano Zampini if (len) { ierr = PetscMemcpy(b->garray,a->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 84811e0482f5SStefano Zampini } else b->garray = 0; 84821e0482f5SStefano Zampini 84831e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->lvec);CHKERRQ(ierr); 84841e0482f5SStefano Zampini b->lvec = a->lvec; 84851e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->lvec);CHKERRQ(ierr); 84861e0482f5SStefano Zampini 84871e0482f5SStefano Zampini /* cannot use VecScatterCopy */ 84881e0482f5SStefano Zampini ierr = VecGetLocalSize(b->lvec,&lsize);CHKERRQ(ierr); 84891e0482f5SStefano Zampini ierr = ISCreateGeneral(ccomm,lsize,b->garray,PETSC_USE_POINTER,&from);CHKERRQ(ierr); 84901e0482f5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,lsize,0,1,&to);CHKERRQ(ierr); 84911e0482f5SStefano Zampini ierr = MatCreateVecs(*B,&gvec,NULL);CHKERRQ(ierr); 84921e0482f5SStefano Zampini ierr = VecScatterCreate(gvec,from,b->lvec,to,&b->Mvctx);CHKERRQ(ierr); 84931e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->Mvctx);CHKERRQ(ierr); 84941e0482f5SStefano Zampini ierr = ISDestroy(&from);CHKERRQ(ierr); 84951e0482f5SStefano Zampini ierr = ISDestroy(&to);CHKERRQ(ierr); 84961e0482f5SStefano Zampini ierr = VecDestroy(&gvec);CHKERRQ(ierr); 84971e0482f5SStefano Zampini } 84981e0482f5SStefano Zampini ierr = MatDestroy(&At);CHKERRQ(ierr); 84991e0482f5SStefano Zampini PetscFunctionReturn(0); 85001e0482f5SStefano Zampini } 8501