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; 1636*ef24a9ddSStefano Zampini PetscBool monolithic = PETSC_FALSE; 16371f4df5f7SStefano Zampini 16381f4df5f7SStefano Zampini PetscFunctionBegin; 1639*ef24a9ddSStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC topology options","PC");CHKERRQ(ierr); 1640*ef24a9ddSStefano Zampini ierr = PetscOptionsBool("-pc_bddc_monolithic","Don't split dofs by block size",NULL,monolithic,&monolithic,NULL);CHKERRQ(ierr); 1641*ef24a9ddSStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 16421f4df5f7SStefano Zampini ierr = MatCreateVecs(pc->pmat,&global,NULL);CHKERRQ(ierr); 16431f4df5f7SStefano Zampini /* need to convert from global to local topology information and remove references to information in global ordering */ 16441f4df5f7SStefano Zampini ierr = MatCreateVecs(matis->A,&local,NULL);CHKERRQ(ierr); 16451f4df5f7SStefano Zampini if (pcbddc->user_provided_isfordofs) { 16461f4df5f7SStefano Zampini if (pcbddc->n_ISForDofs) { 16471f4df5f7SStefano Zampini PetscInt i; 16481f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16491f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 16501f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16511f4df5f7SStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 16521f4df5f7SStefano Zampini } 16531f4df5f7SStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 16541f4df5f7SStefano Zampini pcbddc->n_ISForDofs = 0; 16551f4df5f7SStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 16561f4df5f7SStefano Zampini } 16571f4df5f7SStefano Zampini } else { 1658986cdee1SStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering if bs > 1 */ 16591f4df5f7SStefano Zampini PetscInt i, n = matis->A->rmap->n; 1660986cdee1SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&i);CHKERRQ(ierr); 1661*ef24a9ddSStefano Zampini if (i > 1 && !monolithic) { 1662986cdee1SStefano Zampini pcbddc->n_ISForDofsLocal = i; 16631f4df5f7SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 16641f4df5f7SStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16651f4df5f7SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16661f4df5f7SStefano Zampini } 16671f4df5f7SStefano Zampini } 16687a0e7b2cSstefano_zampini } else { 16697a0e7b2cSstefano_zampini PetscInt i; 16707a0e7b2cSstefano_zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 16717a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 16727a0e7b2cSstefano_zampini } 16731f4df5f7SStefano Zampini } 1674986cdee1SStefano Zampini } 16751f4df5f7SStefano Zampini 16761f4df5f7SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { 16771f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16787a0e7b2cSstefano_zampini } else if (pcbddc->DirichletBoundariesLocal) { 16797a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 16801f4df5f7SStefano Zampini } 16811f4df5f7SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { 16821f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16837a0e7b2cSstefano_zampini } else if (pcbddc->NeumannBoundariesLocal) { 16847a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 16851f4df5f7SStefano Zampini } 16861f4df5f7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { 16871f4df5f7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 16881f4df5f7SStefano Zampini } 16891f4df5f7SStefano Zampini ierr = VecDestroy(&global);CHKERRQ(ierr); 16901f4df5f7SStefano Zampini ierr = VecDestroy(&local);CHKERRQ(ierr); 16917a0e7b2cSstefano_zampini 16927a0e7b2cSstefano_zampini PetscFunctionReturn(0); 16937a0e7b2cSstefano_zampini } 16947a0e7b2cSstefano_zampini 16957a0e7b2cSstefano_zampini PetscErrorCode PCBDDCConsistencyCheckIS(PC pc, MPI_Op mop, IS *is) 16967a0e7b2cSstefano_zampini { 16977a0e7b2cSstefano_zampini Mat_IS *matis = (Mat_IS*)(pc->pmat->data); 16987a0e7b2cSstefano_zampini PetscErrorCode ierr; 16997a0e7b2cSstefano_zampini IS nis; 17007a0e7b2cSstefano_zampini const PetscInt *idxs; 17017a0e7b2cSstefano_zampini PetscInt i,nd,n = matis->A->rmap->n,*nidxs,nnd; 17027a0e7b2cSstefano_zampini PetscBool *ld; 17037a0e7b2cSstefano_zampini 17047a0e7b2cSstefano_zampini PetscFunctionBegin; 17057a0e7b2cSstefano_zampini if (mop != MPI_LAND && mop != MPI_LOR) SETERRQ(PetscObjectComm((PetscObject)(pc)),PETSC_ERR_SUP,"Supported are MPI_LAND and MPI_LOR"); 17067a0e7b2cSstefano_zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 17077a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 17087a0e7b2cSstefano_zampini /* init rootdata with true */ 17097a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_rootdata; 17107a0e7b2cSstefano_zampini for (i=0;i<pc->pmat->rmap->n;i++) ld[i] = PETSC_TRUE; 17117a0e7b2cSstefano_zampini } else { 17127a0e7b2cSstefano_zampini ierr = PetscMemzero(matis->sf_rootdata,pc->pmat->rmap->n*sizeof(PetscBool));CHKERRQ(ierr); 17137a0e7b2cSstefano_zampini } 17147a0e7b2cSstefano_zampini ierr = PetscMemzero(matis->sf_leafdata,n*sizeof(PetscBool));CHKERRQ(ierr); 17157a0e7b2cSstefano_zampini ierr = ISGetLocalSize(*is,&nd);CHKERRQ(ierr); 17167a0e7b2cSstefano_zampini ierr = ISGetIndices(*is,&idxs);CHKERRQ(ierr); 17177a0e7b2cSstefano_zampini ld = (PetscBool*) matis->sf_leafdata; 17187a0e7b2cSstefano_zampini for (i=0;i<nd;i++) 17197a0e7b2cSstefano_zampini if (-1 < idxs[i] && idxs[i] < n) 17207a0e7b2cSstefano_zampini ld[idxs[i]] = PETSC_TRUE; 17217a0e7b2cSstefano_zampini ierr = ISRestoreIndices(*is,&idxs);CHKERRQ(ierr); 17227a0e7b2cSstefano_zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 17237a0e7b2cSstefano_zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);CHKERRQ(ierr); 17247a0e7b2cSstefano_zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 17257a0e7b2cSstefano_zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 17267a0e7b2cSstefano_zampini if (mop == MPI_LAND) { 17277a0e7b2cSstefano_zampini ierr = PetscMalloc1(nd,&nidxs);CHKERRQ(ierr); 17287a0e7b2cSstefano_zampini } else { 17297a0e7b2cSstefano_zampini ierr = PetscMalloc1(n,&nidxs);CHKERRQ(ierr); 17307a0e7b2cSstefano_zampini } 17317a0e7b2cSstefano_zampini for (i=0,nnd=0;i<n;i++) 17327a0e7b2cSstefano_zampini if (ld[i]) 17337a0e7b2cSstefano_zampini nidxs[nnd++] = i; 17347a0e7b2cSstefano_zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(*is)),nnd,nidxs,PETSC_OWN_POINTER,&nis);CHKERRQ(ierr); 17357a0e7b2cSstefano_zampini ierr = ISDestroy(is);CHKERRQ(ierr); 17367a0e7b2cSstefano_zampini *is = nis; 17371f4df5f7SStefano Zampini PetscFunctionReturn(0); 17381f4df5f7SStefano Zampini } 17391f4df5f7SStefano Zampini 17403e589ea0SStefano Zampini PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z) 17413e589ea0SStefano Zampini { 17423e589ea0SStefano Zampini PC_IS *pcis = (PC_IS*)(pc->data); 17433e589ea0SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)(pc->data); 17443e589ea0SStefano Zampini PetscErrorCode ierr; 17453e589ea0SStefano Zampini 17463e589ea0SStefano Zampini PetscFunctionBegin; 17473e589ea0SStefano Zampini if (!pcbddc->benign_have_null) { 17483e589ea0SStefano Zampini PetscFunctionReturn(0); 17493e589ea0SStefano Zampini } 17503e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 17513e589ea0SStefano Zampini Vec swap; 17523e589ea0SStefano Zampini 17533e589ea0SStefano Zampini ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);CHKERRQ(ierr); 17543e589ea0SStefano Zampini swap = pcbddc->work_change; 17553e589ea0SStefano Zampini pcbddc->work_change = r; 17563e589ea0SStefano Zampini r = swap; 17573e589ea0SStefano Zampini } 17583e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 17593e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 17603e589ea0SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 17613e589ea0SStefano Zampini ierr = VecSet(z,0.);CHKERRQ(ierr); 17623e589ea0SStefano Zampini ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 17633e589ea0SStefano Zampini ierr = VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 17643e589ea0SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 1765f913dca9SStefano Zampini pcbddc->work_change = r; 17663e589ea0SStefano Zampini ierr = VecCopy(z,pcbddc->work_change);CHKERRQ(ierr); 17673e589ea0SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);CHKERRQ(ierr); 17683e589ea0SStefano Zampini } 17693e589ea0SStefano Zampini PetscFunctionReturn(0); 17703e589ea0SStefano Zampini } 17713e589ea0SStefano Zampini 1772a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 1773a3df083aSStefano Zampini { 1774a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1775a3df083aSStefano Zampini PetscErrorCode ierr; 1776a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 1777a3df083aSStefano Zampini 1778a3df083aSStefano Zampini PetscFunctionBegin; 1779a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 1780a3df083aSStefano Zampini if (transpose) { 1781a3df083aSStefano Zampini apply_right = ctx->apply_left; 1782a3df083aSStefano Zampini apply_left = ctx->apply_right; 1783a3df083aSStefano Zampini } else { 1784a3df083aSStefano Zampini apply_right = ctx->apply_right; 1785a3df083aSStefano Zampini apply_left = ctx->apply_left; 1786a3df083aSStefano Zampini } 1787a3df083aSStefano Zampini reset_x = PETSC_FALSE; 1788a3df083aSStefano Zampini if (apply_right) { 1789a3df083aSStefano Zampini const PetscScalar *ax; 1790a3df083aSStefano Zampini PetscInt nl,i; 1791a3df083aSStefano Zampini 1792a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 1793a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 1794a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 1795a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 1796a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1797a3df083aSStefano Zampini PetscScalar sum,val; 1798a3df083aSStefano Zampini const PetscInt *idxs; 1799a3df083aSStefano Zampini PetscInt nz,j; 1800a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1801a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1802a3df083aSStefano Zampini sum = 0.; 1803a3df083aSStefano Zampini if (ctx->apply_p0) { 1804a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 1805a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1806a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1807a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 1808a3df083aSStefano Zampini } 1809a3df083aSStefano Zampini } else { 1810a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1811a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 1812a3df083aSStefano Zampini } 1813a3df083aSStefano Zampini } 1814a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 1815a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1816a3df083aSStefano Zampini } 1817a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 1818a3df083aSStefano Zampini reset_x = PETSC_TRUE; 1819a3df083aSStefano Zampini } 1820a3df083aSStefano Zampini if (transpose) { 1821a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 1822a3df083aSStefano Zampini } else { 1823a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 1824a3df083aSStefano Zampini } 1825a3df083aSStefano Zampini if (reset_x) { 1826a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 1827a3df083aSStefano Zampini } 1828a3df083aSStefano Zampini if (apply_left) { 1829a3df083aSStefano Zampini PetscScalar *ay; 1830a3df083aSStefano Zampini PetscInt i; 1831a3df083aSStefano Zampini 1832a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 1833a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1834a3df083aSStefano Zampini PetscScalar sum,val; 1835a3df083aSStefano Zampini const PetscInt *idxs; 1836a3df083aSStefano Zampini PetscInt nz,j; 1837a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1838a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1839a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 1840a3df083aSStefano Zampini if (ctx->apply_p0) { 1841a3df083aSStefano Zampini sum = 0.; 1842a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1843a3df083aSStefano Zampini sum += ay[idxs[j]]; 1844a3df083aSStefano Zampini ay[idxs[j]] += val; 1845a3df083aSStefano Zampini } 1846a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 1847a3df083aSStefano Zampini } else { 1848a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 1849a3df083aSStefano Zampini ay[idxs[j]] += val; 1850a3df083aSStefano Zampini } 1851a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 1852a3df083aSStefano Zampini } 1853a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 1854a3df083aSStefano Zampini } 1855a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 1856a3df083aSStefano Zampini } 1857a3df083aSStefano Zampini PetscFunctionReturn(0); 1858a3df083aSStefano Zampini } 1859a3df083aSStefano Zampini 1860a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 1861a3df083aSStefano Zampini { 1862a3df083aSStefano Zampini PetscErrorCode ierr; 1863a3df083aSStefano Zampini 1864a3df083aSStefano Zampini PetscFunctionBegin; 1865a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 1866a3df083aSStefano Zampini PetscFunctionReturn(0); 1867a3df083aSStefano Zampini } 1868a3df083aSStefano Zampini 1869a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 1870a3df083aSStefano Zampini { 1871a3df083aSStefano Zampini PetscErrorCode ierr; 1872a3df083aSStefano Zampini 1873a3df083aSStefano Zampini PetscFunctionBegin; 1874a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 1875a3df083aSStefano Zampini PetscFunctionReturn(0); 1876a3df083aSStefano Zampini } 1877a3df083aSStefano Zampini 1878a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 1879a3df083aSStefano Zampini { 1880a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1881a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1882a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 1883a3df083aSStefano Zampini PetscErrorCode ierr; 1884a3df083aSStefano Zampini 1885a3df083aSStefano Zampini PetscFunctionBegin; 1886a3df083aSStefano Zampini if (!restore) { 18871dd7afcfSStefano Zampini Mat A_IB,A_BI; 1888a3df083aSStefano Zampini PetscScalar *work; 1889b334f244SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL; 1890a3df083aSStefano Zampini 18919a962809SStefano Zampini if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 18929a962809SStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(0); 1893a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 1894a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 1895a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 1896a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 1897a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 1898a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 1899a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 1900a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 1901a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 1902a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 1903a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 1904a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 1905059032f7SStefano Zampini if (reuse) { 1906a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 19071dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 1908059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 1909059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 1910059032f7SStefano Zampini PetscInt i; 1911059032f7SStefano Zampini 1912059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 1913059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1914059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1915059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1916059032f7SStefano Zampini } 1917059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 19181dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 1919059032f7SStefano Zampini } 1920a3df083aSStefano Zampini ctx->A = pcis->A_IB; 1921a3df083aSStefano Zampini ctx->work = work; 1922a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 1923a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1924a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1925a3df083aSStefano Zampini pcis->A_IB = A_IB; 1926a3df083aSStefano Zampini 1927a3df083aSStefano Zampini /* A_BI as A_IB^T */ 1928a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 1929a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 1930a3df083aSStefano Zampini pcis->A_BI = A_BI; 1931a3df083aSStefano Zampini } else { 19321dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 19331dd7afcfSStefano Zampini PetscFunctionReturn(0); 19341dd7afcfSStefano Zampini } 1935a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 1936a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 1937a3df083aSStefano Zampini pcis->A_IB = ctx->A; 19381dd7afcfSStefano Zampini ctx->A = NULL; 19391dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 19401dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 19411dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 19421dd7afcfSStefano Zampini if (ctx->free) { 1943059032f7SStefano Zampini PetscInt i; 19441dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 1945059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1946059032f7SStefano Zampini } 1947059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 1948059032f7SStefano Zampini } 1949a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 1950a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 1951a3df083aSStefano Zampini } 1952a3df083aSStefano Zampini PetscFunctionReturn(0); 1953a3df083aSStefano Zampini } 1954a3df083aSStefano Zampini 1955a3df083aSStefano Zampini /* used just in bddc debug mode */ 1956a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 1957a3df083aSStefano Zampini { 1958a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1959a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 1960a3df083aSStefano Zampini Mat An; 1961a3df083aSStefano Zampini PetscErrorCode ierr; 1962a3df083aSStefano Zampini 1963a3df083aSStefano Zampini PetscFunctionBegin; 1964a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 1965a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 1966a3df083aSStefano Zampini if (is1) { 19677dae84e0SHong Zhang ierr = MatCreateSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 1968a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 1969a3df083aSStefano Zampini } else { 1970a3df083aSStefano Zampini *B = An; 1971a3df083aSStefano Zampini } 1972a3df083aSStefano Zampini PetscFunctionReturn(0); 1973a3df083aSStefano Zampini } 1974a3df083aSStefano Zampini 19751cf9b237SStefano Zampini /* TODO: add reuse flag */ 19761cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 19771cf9b237SStefano Zampini { 19781cf9b237SStefano Zampini Mat Bt; 19791cf9b237SStefano Zampini PetscScalar *a,*bdata; 19801cf9b237SStefano Zampini const PetscInt *ii,*ij; 19811cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 19821cf9b237SStefano Zampini PetscBool flg_row; 19831cf9b237SStefano Zampini PetscErrorCode ierr; 19841cf9b237SStefano Zampini 19851cf9b237SStefano Zampini PetscFunctionBegin; 19861cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 19871cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 19881cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 19891cf9b237SStefano Zampini nnz = n; 19901cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 19911cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 19921cf9b237SStefano Zampini } 19931cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 19941cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 19951cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 19961cf9b237SStefano Zampini nnz = 0; 19971cf9b237SStefano Zampini bii[0] = 0; 19981cf9b237SStefano Zampini for (i=0;i<n;i++) { 19991cf9b237SStefano Zampini PetscInt j; 20001cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 20011cf9b237SStefano Zampini PetscScalar entry = a[j]; 20021cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 20031cf9b237SStefano Zampini bij[nnz] = ij[j]; 20041cf9b237SStefano Zampini bdata[nnz] = entry; 20051cf9b237SStefano Zampini nnz++; 20061cf9b237SStefano Zampini } 20071cf9b237SStefano Zampini } 20081cf9b237SStefano Zampini bii[i+1] = nnz; 20091cf9b237SStefano Zampini } 20101cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 20111cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 20121cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 20131cf9b237SStefano Zampini { 20141cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 20151cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 20161cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 20171cf9b237SStefano Zampini } 20181cf9b237SStefano Zampini *B = Bt; 20191cf9b237SStefano Zampini PetscFunctionReturn(0); 20201cf9b237SStefano Zampini } 20211cf9b237SStefano Zampini 20224f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 20234f1b2e48SStefano Zampini { 20244f1b2e48SStefano Zampini Mat B; 20254f1b2e48SStefano Zampini IS is_dummy,*cc_n; 20264f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 20274f1b2e48SStefano Zampini PCBDDCGraph graph; 20284f1b2e48SStefano Zampini PetscInt i,n; 20294f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 20304f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 20314f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 20324f1b2e48SStefano Zampini PetscErrorCode ierr; 20334f1b2e48SStefano Zampini 20344f1b2e48SStefano Zampini PetscFunctionBegin; 203563c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 203663c961adSStefano Zampini *ncc = 0; 203763c961adSStefano Zampini *cc = NULL; 203863c961adSStefano Zampini PetscFunctionReturn(0); 203963c961adSStefano Zampini } 20404f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 20414f1b2e48SStefano Zampini if (!isseqaij && filter) { 20421cf9b237SStefano Zampini PetscBool isseqdense; 20431cf9b237SStefano Zampini 20441cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 20451cf9b237SStefano Zampini if (!isseqdense) { 20464f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 20471cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 20481cf9b237SStefano Zampini PetscScalar *array; 20491cf9b237SStefano Zampini PetscReal chop=1.e-6; 20501cf9b237SStefano Zampini 20511cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 20521cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 20531cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 20541cf9b237SStefano Zampini for (i=0;i<n;i++) { 20551cf9b237SStefano Zampini PetscInt j; 20561cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 20571cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 20581cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 20591cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 20601cf9b237SStefano Zampini } 20611cf9b237SStefano Zampini } 20621cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 20639d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 20641cf9b237SStefano Zampini } 20654f1b2e48SStefano Zampini } else { 20664f1b2e48SStefano Zampini B = A; 20674f1b2e48SStefano Zampini } 20684f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 20694f1b2e48SStefano Zampini 20704f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 20714f1b2e48SStefano Zampini if (filter) { 20724f1b2e48SStefano Zampini PetscScalar *data; 20734f1b2e48SStefano Zampini PetscInt j,cum; 20744f1b2e48SStefano Zampini 20754f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 20764f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 20774f1b2e48SStefano Zampini cum = 0; 20784f1b2e48SStefano Zampini for (i=0;i<n;i++) { 20794f1b2e48SStefano Zampini PetscInt t; 20804f1b2e48SStefano Zampini 20814f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 20824f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 20834f1b2e48SStefano Zampini continue; 20844f1b2e48SStefano Zampini } 20854f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 20864f1b2e48SStefano Zampini } 20874f1b2e48SStefano Zampini t = xadj_filtered[i]; 20884f1b2e48SStefano Zampini xadj_filtered[i] = cum; 20894f1b2e48SStefano Zampini cum += t; 20904f1b2e48SStefano Zampini } 20914f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 20924f1b2e48SStefano Zampini } else { 20934f1b2e48SStefano Zampini xadj_filtered = NULL; 20944f1b2e48SStefano Zampini adjncy_filtered = NULL; 20954f1b2e48SStefano Zampini } 20964f1b2e48SStefano Zampini 20974f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 20984f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 20994f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 21004f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 21014f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 2102be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);CHKERRQ(ierr); 21034f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 21044f1b2e48SStefano Zampini if (xadj_filtered) { 21054f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 21064f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 21074f1b2e48SStefano Zampini } else { 21084f1b2e48SStefano Zampini graph->xadj = xadj; 21094f1b2e48SStefano Zampini graph->adjncy = adjncy; 21104f1b2e48SStefano Zampini } 21114f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 21124f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 21134f1b2e48SStefano Zampini /* partial clean up */ 21144f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 21154f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 21161cf9b237SStefano Zampini if (A != B) { 21174f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 21184f1b2e48SStefano Zampini } 21194f1b2e48SStefano Zampini 21204f1b2e48SStefano Zampini /* get back data */ 21211cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 21221cf9b237SStefano Zampini if (cc) { 21234f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 21244f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 21254f1b2e48SStefano 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); 21264f1b2e48SStefano Zampini } 21274f1b2e48SStefano Zampini *cc = cc_n; 21281cf9b237SStefano Zampini } 21294f1b2e48SStefano Zampini /* clean up graph */ 21304f1b2e48SStefano Zampini graph->xadj = 0; 21314f1b2e48SStefano Zampini graph->adjncy = 0; 21324f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 21334f1b2e48SStefano Zampini PetscFunctionReturn(0); 21344f1b2e48SStefano Zampini } 21354f1b2e48SStefano Zampini 21365408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 21375408967cSStefano Zampini { 21385408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 21395408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2140dee84bffSStefano Zampini IS dirIS = NULL; 21414f1b2e48SStefano Zampini PetscInt i; 21425408967cSStefano Zampini PetscErrorCode ierr; 21435408967cSStefano Zampini 21445408967cSStefano Zampini PetscFunctionBegin; 2145dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 21465408967cSStefano Zampini if (zerodiag) { 21475408967cSStefano Zampini Mat A; 21485408967cSStefano Zampini Vec vec3_N; 21495408967cSStefano Zampini PetscScalar *vals; 21505408967cSStefano Zampini const PetscInt *idxs; 2151d12d3064SStefano Zampini PetscInt nz,*count; 21525408967cSStefano Zampini 21535408967cSStefano Zampini /* p0 */ 21545408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 21555408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 21565408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 21575408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 21584f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 21595408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21605408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 21615408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 21625408967cSStefano Zampini /* v_I */ 21635408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 21645408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 21655408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21665408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 21675408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 21685408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 21695408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21705408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 21715408967cSStefano Zampini if (dirIS) { 21725408967cSStefano Zampini PetscInt n; 21735408967cSStefano Zampini 21745408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 21755408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 21765408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 21775408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 21785408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 21795408967cSStefano Zampini } 21805408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 21815408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 21825408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 21835408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 2184669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 21855408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 21865408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 21879a962809SStefano 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])); 21885408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 21895408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 2190d12d3064SStefano Zampini 2191d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 2192d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 2193d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2194d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 2195d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 2196d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 21979a962809SStefano 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]); 2198d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 2199d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 22005408967cSStefano Zampini } 2201dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 22025408967cSStefano Zampini 22035408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 22045408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 22054f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 22065408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 22074f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 22085408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 2209f2a566d8SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2210f2a566d8SStefano Zampini PetscInt val = PetscRealPart(pcbddc->benign_p0[i]); 2211f2a566d8SStefano 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); 2212f2a566d8SStefano Zampini } 22135408967cSStefano Zampini PetscFunctionReturn(0); 22145408967cSStefano Zampini } 22155408967cSStefano Zampini 2216339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 2217339f8db1SStefano Zampini { 2218339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 22194edc6404Sstefano_zampini IS pressures,zerodiag,zerodiag_save,*zerodiag_subs; 2220b0f5fe93SStefano Zampini PetscInt nz,n; 22214edc6404Sstefano_zampini PetscInt *interior_dofs,n_interior_dofs,nneu; 22224edc6404Sstefano_zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag,checkb; 2223339f8db1SStefano Zampini PetscErrorCode ierr; 2224339f8db1SStefano Zampini 2225339f8db1SStefano Zampini PetscFunctionBegin; 22269f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 22279f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 2228a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 2229a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 2230a3df083aSStefano Zampini } 2231a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 2232a3df083aSStefano Zampini pcbddc->benign_n = 0; 223328b8efb1Sstefano_zampini 223428b8efb1Sstefano_zampini /* if a local info on dofs is present, uses the last field for "pressures" (or fid by command line) 22354f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 22364f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 22374f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 22381ae86dd6SStefano Zampini since the local Schur complements are already SPD 22394f1b2e48SStefano Zampini */ 22404f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 22414f1b2e48SStefano Zampini have_null = PETSC_TRUE; 224240fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 22434f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 22444f1b2e48SStefano Zampini 224528b8efb1Sstefano_zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC benign options","PC");CHKERRQ(ierr); 224628b8efb1Sstefano_zampini ierr = PetscOptionsInt("-pc_bddc_pressure_field","Field id for pressures",NULL,p,&p,NULL);CHKERRQ(ierr); 224728b8efb1Sstefano_zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 224828b8efb1Sstefano_zampini if (p < 0 || p > pcbddc->n_ISForDofsLocal-1) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Invalid field id for pressures %D",p); 22494f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 22504f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 22514f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 22524f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 2253ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 225440fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 225540fa8d13SStefano Zampini if (!sorted) { 225640fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 225740fa8d13SStefano Zampini } 225840fa8d13SStefano Zampini } else { 225940fa8d13SStefano Zampini pressures = NULL; 226040fa8d13SStefano Zampini } 226197d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 226297d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 226327b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 226497d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 2265339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 2266339f8db1SStefano Zampini if (!sorted) { 2267339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 2268339f8db1SStefano Zampini } 22694edc6404Sstefano_zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 22704edc6404Sstefano_zampini zerodiag_save = zerodiag; 2271339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 22724f1b2e48SStefano Zampini if (!nz) { 22734f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 22744f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 227540fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 227640fa8d13SStefano Zampini } 22774f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 22784f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 22794f1b2e48SStefano Zampini zerodiag_subs = NULL; 22804f1b2e48SStefano Zampini pcbddc->benign_n = 0; 22811f4df5f7SStefano Zampini n_interior_dofs = 0; 22821f4df5f7SStefano Zampini interior_dofs = NULL; 22834edc6404Sstefano_zampini nneu = 0; 22844edc6404Sstefano_zampini if (pcbddc->NeumannBoundariesLocal) { 22854edc6404Sstefano_zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nneu);CHKERRQ(ierr); 22864edc6404Sstefano_zampini } 22873369cb78Sstefano_zampini checkb = (PetscBool)(!pcbddc->NeumannBoundariesLocal || pcbddc->current_level); 22884edc6404Sstefano_zampini if (checkb) { /* need to compute interior nodes */ 22891f4df5f7SStefano Zampini PetscInt n,i,j; 22901f4df5f7SStefano Zampini PetscInt n_neigh,*neigh,*n_shared,**shared; 22911f4df5f7SStefano Zampini PetscInt *iwork; 22921f4df5f7SStefano Zampini 22931f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);CHKERRQ(ierr); 22941f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 22951f4df5f7SStefano Zampini ierr = PetscCalloc1(n,&iwork);CHKERRQ(ierr); 22961f4df5f7SStefano Zampini ierr = PetscMalloc1(n,&interior_dofs);CHKERRQ(ierr); 229790648384SStefano Zampini for (i=1;i<n_neigh;i++) 22981f4df5f7SStefano Zampini for (j=0;j<n_shared[i];j++) 22991f4df5f7SStefano Zampini iwork[shared[i][j]] += 1; 23001f4df5f7SStefano Zampini for (i=0;i<n;i++) 23011f4df5f7SStefano Zampini if (!iwork[i]) 23021f4df5f7SStefano Zampini interior_dofs[n_interior_dofs++] = i; 23031f4df5f7SStefano Zampini ierr = PetscFree(iwork);CHKERRQ(ierr); 23041f4df5f7SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);CHKERRQ(ierr); 23051f4df5f7SStefano Zampini } 23064f1b2e48SStefano Zampini if (has_null_pressures) { 23074f1b2e48SStefano Zampini IS *subs; 23084edc6404Sstefano_zampini PetscInt nsubs,i,j,nl; 23091f4df5f7SStefano Zampini const PetscInt *idxs; 23101f4df5f7SStefano Zampini PetscScalar *array; 23111f4df5f7SStefano Zampini Vec *work; 23121f4df5f7SStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 23134f1b2e48SStefano Zampini 23144f1b2e48SStefano Zampini subs = pcbddc->local_subs; 23154f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 23161f4df5f7SStefano 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) */ 23174edc6404Sstefano_zampini if (checkb) { 23181f4df5f7SStefano Zampini ierr = VecDuplicateVecs(matis->y,2,&work);CHKERRQ(ierr); 23191f4df5f7SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nl);CHKERRQ(ierr); 23201f4df5f7SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 23211f4df5f7SStefano Zampini /* work[0] = 1_p */ 23221f4df5f7SStefano Zampini ierr = VecSet(work[0],0.);CHKERRQ(ierr); 23231f4df5f7SStefano Zampini ierr = VecGetArray(work[0],&array);CHKERRQ(ierr); 23241f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 23251f4df5f7SStefano Zampini ierr = VecRestoreArray(work[0],&array);CHKERRQ(ierr); 23261f4df5f7SStefano Zampini /* work[0] = 1_v */ 23271f4df5f7SStefano Zampini ierr = VecSet(work[1],1.);CHKERRQ(ierr); 23281f4df5f7SStefano Zampini ierr = VecGetArray(work[1],&array);CHKERRQ(ierr); 23291f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 0.; 23301f4df5f7SStefano Zampini ierr = VecRestoreArray(work[1],&array);CHKERRQ(ierr); 23311f4df5f7SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 23321f4df5f7SStefano Zampini } 23334f1b2e48SStefano Zampini if (nsubs > 1) { 23344f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 23354f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 23364f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 23374f1b2e48SStefano Zampini IS t_zerodiag_subs; 23384f1b2e48SStefano Zampini PetscInt nl; 23394f1b2e48SStefano Zampini 23404f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 23414f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 23424f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 23434f1b2e48SStefano Zampini if (nl) { 23444f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 23454f1b2e48SStefano Zampini 23464edc6404Sstefano_zampini if (checkb) { 23471f4df5f7SStefano Zampini ierr = VecSet(matis->x,0);CHKERRQ(ierr); 23481f4df5f7SStefano Zampini ierr = ISGetLocalSize(subs[i],&nl);CHKERRQ(ierr); 23491f4df5f7SStefano Zampini ierr = ISGetIndices(subs[i],&idxs);CHKERRQ(ierr); 23501f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 23511f4df5f7SStefano Zampini for (j=0;j<nl;j++) array[idxs[j]] = 1.; 23521f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 23531f4df5f7SStefano Zampini ierr = ISRestoreIndices(subs[i],&idxs);CHKERRQ(ierr); 23541f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[0],matis->x);CHKERRQ(ierr); 23551f4df5f7SStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 23561f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->y,work[1],matis->y);CHKERRQ(ierr); 23571f4df5f7SStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 23581f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 23591f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 23601f4df5f7SStefano Zampini valid = PETSC_FALSE; 23611f4df5f7SStefano Zampini break; 23621f4df5f7SStefano Zampini } 23631f4df5f7SStefano Zampini } 23641f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 23651f4df5f7SStefano Zampini } 23666632bad2Sstefano_zampini if (valid && nneu) { 23676632bad2Sstefano_zampini const PetscInt *idxs; 23681f4df5f7SStefano Zampini PetscInt nzb; 23691f4df5f7SStefano Zampini 23706632bad2Sstefano_zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 23716632bad2Sstefano_zampini ierr = ISGlobalToLocalMappingApply(l2g,IS_GTOLM_DROP,nneu,idxs,&nzb,NULL);CHKERRQ(ierr); 23726632bad2Sstefano_zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 23731f4df5f7SStefano Zampini if (nzb) valid = PETSC_FALSE; 23741f4df5f7SStefano Zampini } 23751f4df5f7SStefano Zampini if (valid && pressures) { 23764f1b2e48SStefano Zampini IS t_pressure_subs; 23774f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 23784f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 23794f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 23804f1b2e48SStefano Zampini } 23814f1b2e48SStefano Zampini if (valid) { 23824f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 23834f1b2e48SStefano Zampini pcbddc->benign_n++; 23844f1b2e48SStefano Zampini } else { 23854f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 23864f1b2e48SStefano Zampini } 23874f1b2e48SStefano Zampini } 23884f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 23894f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 23904f1b2e48SStefano Zampini } 23914f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 23924f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 23931f4df5f7SStefano Zampini 23946632bad2Sstefano_zampini if (nneu) valid = PETSC_FALSE; 23951f4df5f7SStefano Zampini if (valid && pressures) { 23964f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 23974f1b2e48SStefano Zampini } 23984edc6404Sstefano_zampini if (valid && checkb) { 23991f4df5f7SStefano Zampini ierr = MatMult(matis->A,work[0],matis->x);CHKERRQ(ierr); 24001f4df5f7SStefano Zampini ierr = VecPointwiseMult(matis->x,work[1],matis->x);CHKERRQ(ierr); 24011f4df5f7SStefano Zampini ierr = VecGetArray(matis->x,&array);CHKERRQ(ierr); 24021f4df5f7SStefano Zampini for (j=0;j<n_interior_dofs;j++) { 24031f4df5f7SStefano Zampini if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) { 24041f4df5f7SStefano Zampini valid = PETSC_FALSE; 24051f4df5f7SStefano Zampini break; 24061f4df5f7SStefano Zampini } 24071f4df5f7SStefano Zampini } 24081f4df5f7SStefano Zampini ierr = VecRestoreArray(matis->x,&array);CHKERRQ(ierr); 24091f4df5f7SStefano Zampini } 24104f1b2e48SStefano Zampini if (valid) { 24114f1b2e48SStefano Zampini pcbddc->benign_n = 1; 2412ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 24134f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 24144f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 24154f1b2e48SStefano Zampini } 24164f1b2e48SStefano Zampini } 24174edc6404Sstefano_zampini if (checkb) { 24181f4df5f7SStefano Zampini ierr = VecDestroyVecs(2,&work);CHKERRQ(ierr); 24194f1b2e48SStefano Zampini } 24201f4df5f7SStefano Zampini } 24211f4df5f7SStefano Zampini ierr = PetscFree(interior_dofs);CHKERRQ(ierr); 24224f1b2e48SStefano Zampini 24234f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2424b9b0e38cSStefano Zampini PetscInt n; 2425b9b0e38cSStefano Zampini 24264f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 24274f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 2428b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 2429b9b0e38cSStefano Zampini if (n) { 24304f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 24314f1b2e48SStefano Zampini have_null = PETSC_FALSE; 24324f1b2e48SStefano Zampini } 2433b9b0e38cSStefano Zampini } 24344f1b2e48SStefano Zampini 24354f1b2e48SStefano Zampini /* final check for null pressures */ 24364f1b2e48SStefano Zampini if (zerodiag && pressures) { 24374f1b2e48SStefano Zampini PetscInt nz,np; 24384f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 24394f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 24404f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 24414f1b2e48SStefano Zampini } 24424f1b2e48SStefano Zampini 24434f1b2e48SStefano Zampini if (recompute_zerodiag) { 24444f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 24454f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 24464f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 24474f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 24484f1b2e48SStefano Zampini } else { 24494f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 24504f1b2e48SStefano Zampini 24514f1b2e48SStefano Zampini nzn = 0; 24524f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 24534f1b2e48SStefano Zampini PetscInt ns; 24544f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 24554f1b2e48SStefano Zampini nzn += ns; 24564f1b2e48SStefano Zampini } 24574f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 24584f1b2e48SStefano Zampini nzn = 0; 24594f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 24604f1b2e48SStefano Zampini PetscInt ns,*idxs; 24614f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 24624f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 24634f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 24644f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 24654f1b2e48SStefano Zampini nzn += ns; 24664f1b2e48SStefano Zampini } 24674f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 24684f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 24694f1b2e48SStefano Zampini } 24704f1b2e48SStefano Zampini have_null = PETSC_FALSE; 24714f1b2e48SStefano Zampini } 24724f1b2e48SStefano Zampini 2473669cc0f4SStefano Zampini /* Prepare matrix to compute no-net-flux */ 2474a198735bSStefano Zampini if (pcbddc->compute_nonetflux && !pcbddc->divudotp) { 2475a198735bSStefano Zampini Mat A,loc_divudotp; 2476a198735bSStefano Zampini ISLocalToGlobalMapping rl2g,cl2g,l2gmap; 2477a198735bSStefano Zampini IS row,col,isused = NULL; 2478a198735bSStefano Zampini PetscInt M,N,n,st,n_isused; 2479a198735bSStefano Zampini 24801f4df5f7SStefano Zampini if (pressures) { 24811f4df5f7SStefano Zampini isused = pressures; 24821f4df5f7SStefano Zampini } else { 24834edc6404Sstefano_zampini isused = zerodiag_save; 24841f4df5f7SStefano Zampini } 2485a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);CHKERRQ(ierr); 2486669cc0f4SStefano Zampini ierr = MatISGetLocalMat(pc->pmat,&A);CHKERRQ(ierr); 24871ae86dd6SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 24881ae86dd6SStefano 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"); 2489a198735bSStefano Zampini n_isused = 0; 2490a198735bSStefano Zampini if (isused) { 2491a198735bSStefano Zampini ierr = ISGetLocalSize(isused,&n_isused);CHKERRQ(ierr); 2492a198735bSStefano Zampini } 2493a198735bSStefano Zampini ierr = MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2494a198735bSStefano Zampini st = st-n_isused; 24951ae86dd6SStefano Zampini if (n) { 2496a198735bSStefano Zampini const PetscInt *gidxs; 2497a198735bSStefano Zampini 24987dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);CHKERRQ(ierr); 2499a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 2500a198735bSStefano Zampini /* TODO: extend ISCreateStride with st = PETSC_DECIDE */ 2501a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2502a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2503a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 25041ae86dd6SStefano Zampini } else { 2505a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);CHKERRQ(ierr); 2506a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);CHKERRQ(ierr); 2507a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 2508a198735bSStefano Zampini } 2509a198735bSStefano Zampini ierr = MatGetSize(pc->pmat,NULL,&N);CHKERRQ(ierr); 2510a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 2511a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 2512a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 2513a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 2514a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 2515a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);CHKERRQ(ierr); 2516a198735bSStefano Zampini ierr = MatSetType(pcbddc->divudotp,MATIS);CHKERRQ(ierr); 2517a198735bSStefano Zampini ierr = MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 2518a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);CHKERRQ(ierr); 2519a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 2520a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 2521a198735bSStefano Zampini ierr = MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);CHKERRQ(ierr); 2522a198735bSStefano Zampini ierr = MatDestroy(&loc_divudotp);CHKERRQ(ierr); 25231ae86dd6SStefano Zampini ierr = MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 25241ae86dd6SStefano Zampini ierr = MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 25251ae86dd6SStefano Zampini } 25264edc6404Sstefano_zampini ierr = ISDestroy(&zerodiag_save);CHKERRQ(ierr); 2527b3afcdbeSStefano Zampini 2528b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 25294f1b2e48SStefano Zampini if (has_null_pressures) { 25304f1b2e48SStefano Zampini IS zerodiagc; 25314f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 25324f1b2e48SStefano Zampini PetscInt i,s,*nnz; 25334f1b2e48SStefano Zampini 25344f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 2535339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 2536339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 2537339f8db1SStefano Zampini /* local change of basis for pressures */ 2538339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 253997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 2540339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 2541339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2542339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 25434f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 25444f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 25454f1b2e48SStefano Zampini PetscInt nzs,j; 25464f1b2e48SStefano Zampini 25474f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 25484f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 25494f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 25504f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 25514f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 25524f1b2e48SStefano Zampini } 2553339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 2554339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2555339f8db1SStefano Zampini /* set identity on velocities */ 2556339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 2557339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 2558339f8db1SStefano Zampini } 25594f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 25604f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 25619f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 25624f1b2e48SStefano 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); 2563339f8db1SStefano Zampini /* set change on pressures */ 25644f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 25654f1b2e48SStefano Zampini PetscScalar *array; 25664f1b2e48SStefano Zampini PetscInt nzs; 25674f1b2e48SStefano Zampini 25684f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 25694f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 25704f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 2571339f8db1SStefano Zampini PetscScalar vals[2]; 2572339f8db1SStefano Zampini PetscInt cols[2]; 2573339f8db1SStefano Zampini 2574339f8db1SStefano Zampini cols[0] = idxs[i]; 25754f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 2576339f8db1SStefano Zampini vals[0] = 1.; 2577b0f5fe93SStefano Zampini vals[1] = 1.; 25784f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2579339f8db1SStefano Zampini } 25804f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 25814f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 25824f1b2e48SStefano Zampini array[nzs-1] = 1.; 25834f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 25844f1b2e48SStefano Zampini /* store local idxs for p0 */ 25854f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 25864f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 2587339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 25884f1b2e48SStefano Zampini } 2589339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2590339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2591a3df083aSStefano Zampini /* project if needed */ 2592a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 25931dd7afcfSStefano Zampini Mat M; 25941dd7afcfSStefano Zampini 25951dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 2596339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25971dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25981dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2599a3df083aSStefano Zampini } 26004f1b2e48SStefano Zampini /* store global idxs for p0 */ 26014f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2602339f8db1SStefano Zampini } 2603ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 26044f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 2605b0f5fe93SStefano Zampini 2606b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 2607b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 260827b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 260927b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2610339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 2611339f8db1SStefano Zampini PetscFunctionReturn(0); 2612339f8db1SStefano Zampini } 2613339f8db1SStefano Zampini 2614015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 2615efc2fbd9SStefano Zampini { 2616efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2617de9d7bd0SStefano Zampini PetscScalar *array; 2618efc2fbd9SStefano Zampini PetscErrorCode ierr; 2619efc2fbd9SStefano Zampini 2620efc2fbd9SStefano Zampini PetscFunctionBegin; 2621efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 2622efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 26234f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 2624efc2fbd9SStefano Zampini } 2625de9d7bd0SStefano Zampini if (get) { 2626efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 26274f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 26284f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 2629efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 2630de9d7bd0SStefano Zampini } else { 2631de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 2632de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2633de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 2634de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 2635efc2fbd9SStefano Zampini } 2636efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2637efc2fbd9SStefano Zampini } 2638efc2fbd9SStefano Zampini 2639c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 2640c263805aSStefano Zampini { 2641c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2642c263805aSStefano Zampini PetscErrorCode ierr; 2643c263805aSStefano Zampini 2644c263805aSStefano Zampini PetscFunctionBegin; 2645c263805aSStefano Zampini /* TODO: add error checking 2646c263805aSStefano Zampini - avoid nested pop (or push) calls. 2647c263805aSStefano Zampini - cannot push before pop. 26481c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 2649c263805aSStefano Zampini */ 26504f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 2651efc2fbd9SStefano Zampini PetscFunctionReturn(0); 2652efc2fbd9SStefano Zampini } 2653c263805aSStefano Zampini if (pop) { 2654a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 26554f1b2e48SStefano Zampini IS is_p0; 26564f1b2e48SStefano Zampini MatReuse reuse; 2657c263805aSStefano Zampini 2658c263805aSStefano Zampini /* extract B_0 */ 26594f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 26604f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 26614f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 26624f1b2e48SStefano Zampini } 26634f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 26647dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 2665c263805aSStefano Zampini /* remove rows and cols from local problem */ 2666c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 266797d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 26684f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 26694f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 2670a3df083aSStefano Zampini } else { 2671a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 2672a3df083aSStefano Zampini PetscScalar *vals; 2673a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 2674a3df083aSStefano Zampini 2675a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 2676a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 2677a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 26780b5adadeSStefano Zampini PetscInt *nnz; 2679a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 2680a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 2681a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2682331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 2683331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2684331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 2685331e053bSStefano Zampini nnz[i] = n - nnz[i]; 2686331e053bSStefano Zampini } 2687331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 2688331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2689331e053bSStefano Zampini } 2690a3df083aSStefano Zampini 2691a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 2692a3df083aSStefano Zampini PetscScalar *array; 2693a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 2694a3df083aSStefano Zampini 2695a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 2696a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 2697a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2698a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 2699a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 2700a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 2701a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 2702a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 2703a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 2704a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 2705a3df083aSStefano Zampini cum = 0; 2706a3df083aSStefano Zampini for (j=0;j<n;j++) { 270722db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 2708a3df083aSStefano Zampini vals[cum] = array[j]; 2709a3df083aSStefano Zampini idxs_ins[cum] = j; 2710a3df083aSStefano Zampini cum++; 2711a3df083aSStefano Zampini } 2712a3df083aSStefano Zampini } 2713a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 2714a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 2715a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 2716a3df083aSStefano Zampini } 2717a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2718a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2719a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 2720a3df083aSStefano Zampini } 2721c263805aSStefano Zampini } else { /* push */ 2722a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 27234f1b2e48SStefano Zampini PetscInt i; 27244f1b2e48SStefano Zampini 27254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 27264f1b2e48SStefano Zampini PetscScalar *B0_vals; 27274f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 27284f1b2e48SStefano Zampini 27294f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 27304f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 27317b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 27324f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 27334f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 27344f1b2e48SStefano Zampini } 2735c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2736c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2737a3df083aSStefano Zampini } else { 2738a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 2739a3df083aSStefano Zampini } 2740c263805aSStefano Zampini } 2741c263805aSStefano Zampini PetscFunctionReturn(0); 2742c263805aSStefano Zampini } 2743c263805aSStefano Zampini 274408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 2745b1b3d7a2SStefano Zampini { 2746b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 274708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 274808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 274908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 275008122e43SStefano Zampini PetscScalar *work,lwork; 275108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 275208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 275308122e43SStefano Zampini PetscReal *eigs,thresh; 27541b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 2755f6f667cfSStefano Zampini PetscBool allocated_S_St; 275608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 275708122e43SStefano Zampini PetscReal *rwork; 275808122e43SStefano Zampini #endif 2759b1b3d7a2SStefano Zampini PetscErrorCode ierr; 2760b1b3d7a2SStefano Zampini 2761b1b3d7a2SStefano Zampini PetscFunctionBegin; 2762b334f244SStefano Zampini if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data"); 2763af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 2764d750e150Sstefano_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); 276506a4e24aSStefano Zampini 2766fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2767fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2768fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2769fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 27701575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2771fd14bc51SStefano Zampini } 2772fd14bc51SStefano Zampini 2773e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 2774e496cd5dSStefano 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); 2775e496cd5dSStefano Zampini } 2776e496cd5dSStefano Zampini 277708122e43SStefano Zampini /* max size of subsets */ 277808122e43SStefano Zampini mss = 0; 277908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 278008122e43SStefano Zampini PetscInt subset_size; 2781862806e4SStefano Zampini 278208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 278308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 278408122e43SStefano Zampini } 278508122e43SStefano Zampini 278608122e43SStefano Zampini /* min/max and threshold */ 278708122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 2788f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 278908122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 2790f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 2791f6f667cfSStefano Zampini if (nmin) { 2792f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 2793f6f667cfSStefano Zampini } 279408122e43SStefano Zampini 279508122e43SStefano Zampini /* allocate lapack workspace */ 279608122e43SStefano Zampini cum = cum2 = 0; 279708122e43SStefano Zampini maxneigs = 0; 279808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 279908122e43SStefano Zampini PetscInt n,subset_size; 2800f6f667cfSStefano Zampini 280108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 280208122e43SStefano Zampini n = PetscMin(subset_size,nmax); 28039162d606SStefano Zampini cum += subset_size; 28049162d606SStefano Zampini cum2 += subset_size*n; 280508122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 280608122e43SStefano Zampini } 280708122e43SStefano Zampini if (mss) { 28089ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 280908122e43SStefano Zampini PetscBLASInt B_itype = 1; 281008122e43SStefano Zampini PetscBLASInt B_N = mss; 28114c6709b3SStefano Zampini PetscReal zero = 0.0; 28124c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 281308122e43SStefano Zampini 281408122e43SStefano Zampini B_lwork = -1; 281508122e43SStefano Zampini S = NULL; 281608122e43SStefano Zampini St = NULL; 2817a58a30b4SStefano Zampini eigs = NULL; 2818a58a30b4SStefano Zampini eigv = NULL; 2819a58a30b4SStefano Zampini B_iwork = NULL; 2820a58a30b4SStefano Zampini B_ifail = NULL; 2821d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2822d1710679SStefano Zampini rwork = NULL; 2823d1710679SStefano Zampini #endif 28248bec7fa6SStefano Zampini thresh = 1.0; 282508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 282608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 282708122e43SStefano 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)); 282808122e43SStefano Zampini #else 282908122e43SStefano 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)); 283008122e43SStefano Zampini #endif 283108122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 283208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 283308122e43SStefano Zampini } else { 283408122e43SStefano Zampini /* TODO */ 283508122e43SStefano Zampini } 283608122e43SStefano Zampini } else { 283708122e43SStefano Zampini lwork = 0; 283808122e43SStefano Zampini } 283908122e43SStefano Zampini 284008122e43SStefano Zampini nv = 0; 2841d62866d3SStefano 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) */ 2842d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 284308122e43SStefano Zampini } 28444c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 2845f6f667cfSStefano Zampini if (allocated_S_St) { 2846f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 2847f6f667cfSStefano Zampini } 2848f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 284908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 285008122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 285108122e43SStefano Zampini #endif 28529162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 28539162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 28549162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 285508122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 28569162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 285708122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 285808122e43SStefano Zampini 285908122e43SStefano Zampini maxneigs = 0; 286072b8c272SStefano Zampini cum = cumarray = 0; 28619162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 28629162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 2863d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 286408122e43SStefano Zampini const PetscInt *idxs; 286508122e43SStefano Zampini 2866d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 286708122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 286808122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 286908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 287008122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 28719162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 28729162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 287308122e43SStefano Zampini } 2874d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 287508122e43SStefano Zampini } 287608122e43SStefano Zampini 287708122e43SStefano Zampini if (mss) { /* multilevel */ 287808122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 287908122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 288008122e43SStefano Zampini } 288108122e43SStefano Zampini 2882ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 288308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 288408122e43SStefano Zampini const PetscInt *idxs; 28859d54b7f4SStefano Zampini PetscReal upper,lower; 2886862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 288708122e43SStefano Zampini PetscBLASInt B_N; 2888aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 288908122e43SStefano Zampini 28909d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 28919d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 28929d54b7f4SStefano Zampini lower = thresh; 28939d54b7f4SStefano Zampini } else { 28949d54b7f4SStefano Zampini upper = 1./thresh; 28959d54b7f4SStefano Zampini lower = 0.; 28969d54b7f4SStefano Zampini } 2897862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 2898ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 2899f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 2900f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 29019ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 2902aff50787SStefano Zampini PetscInt j,k; 2903aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 2904aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 2905aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 290608122e43SStefano Zampini } 290708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 2908aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 2909aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 2910aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 2911aff50787SStefano Zampini } 291208122e43SStefano Zampini } 291308122e43SStefano Zampini } else { 291408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 291508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 291608122e43SStefano Zampini } 29178bec7fa6SStefano Zampini } else { 2918f6f667cfSStefano Zampini S = Sarray + cumarray; 2919f6f667cfSStefano Zampini St = Starray + cumarray; 29208bec7fa6SStefano Zampini } 2921aff50787SStefano Zampini /* see if we can save some work */ 2922b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 2923aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 2924aff50787SStefano Zampini } 2925aff50787SStefano Zampini 2926b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 2927aff50787SStefano Zampini B_neigs = 0; 2928aff50787SStefano Zampini } else { 29299ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 293008122e43SStefano Zampini PetscBLASInt B_itype = 1; 2931f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 29324c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 29339552c7c7SStefano Zampini PetscInt nmin_s; 2934b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 293508122e43SStefano Zampini 2936fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2937eee23b56SStefano 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]]); 2938fd14bc51SStefano Zampini } 2939d16cbb6bSStefano Zampini 2940b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 2941b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 2942b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 2943b7ab4a40SStefano Zampini } 2944b7ab4a40SStefano Zampini 294508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2946b7ab4a40SStefano Zampini if (compute_range) { 2947d16cbb6bSStefano Zampini 2948d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 294908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 29509d54b7f4SStefano 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)); 295108122e43SStefano Zampini #else 29529d54b7f4SStefano 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)); 295308122e43SStefano Zampini #endif 2954b7ab4a40SStefano Zampini } else if (!same_data) { 2955d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 2956d16cbb6bSStefano Zampini B_IL = 1; 2957d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 29589d54b7f4SStefano 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)); 2959d16cbb6bSStefano Zampini #else 29609d54b7f4SStefano 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)); 2961d16cbb6bSStefano Zampini #endif 2962b03ebc13SStefano Zampini } else { /* same_data is true, so just get the adaptive functional requested by the user */ 2963b7ab4a40SStefano Zampini PetscInt k; 2964b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 2965b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 2966b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 2967b7ab4a40SStefano Zampini nmin = nmax; 2968b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 2969b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 2970b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 2971b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 2972b7ab4a40SStefano Zampini } 2973d16cbb6bSStefano Zampini } 297408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 297508122e43SStefano Zampini if (B_ierr) { 29766c4ed002SBarry 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); 29776c4ed002SBarry 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); 29786c4ed002SBarry 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); 297908122e43SStefano Zampini } 298008122e43SStefano Zampini 298108122e43SStefano Zampini if (B_neigs > nmax) { 2982fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2983fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 2984fd14bc51SStefano Zampini } 29859d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 298608122e43SStefano Zampini B_neigs = nmax; 298708122e43SStefano Zampini } 298808122e43SStefano Zampini 29899552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 29909552c7c7SStefano Zampini if (B_neigs < nmin_s) { 299108122e43SStefano Zampini PetscBLASInt B_neigs2; 299208122e43SStefano Zampini 29939d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2994f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 29959d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 29969d54b7f4SStefano Zampini } else { 29979d54b7f4SStefano Zampini B_IL = B_neigs + 1; 29989d54b7f4SStefano Zampini B_IU = nmin_s; 29999d54b7f4SStefano Zampini } 3000fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3001fd14bc51SStefano 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); 3002fd14bc51SStefano Zampini } 30039ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 30041ae86dd6SStefano Zampini PetscInt j,k; 300508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 30061ae86dd6SStefano Zampini for (k=j;k<subset_size;k++) { 30071ae86dd6SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 30081ae86dd6SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 300908122e43SStefano Zampini } 301008122e43SStefano Zampini } 301108122e43SStefano Zampini } else { 301208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 301308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 301408122e43SStefano Zampini } 301508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 301608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 30179d54b7f4SStefano 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)); 301808122e43SStefano Zampini #else 30199d54b7f4SStefano 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)); 302008122e43SStefano Zampini #endif 302108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 302208122e43SStefano Zampini B_neigs += B_neigs2; 302308122e43SStefano Zampini } 302408122e43SStefano Zampini if (B_ierr) { 30256c4ed002SBarry 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); 30266c4ed002SBarry 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); 30276c4ed002SBarry 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); 302808122e43SStefano Zampini } 3029fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3030ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 303108122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 303208122e43SStefano Zampini if (eigs[j] == 0.0) { 3033ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 303408122e43SStefano Zampini } else { 30359d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3036ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 30379d54b7f4SStefano Zampini } else { 30389d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 30399d54b7f4SStefano Zampini } 3040fd14bc51SStefano Zampini } 304108122e43SStefano Zampini } 304208122e43SStefano Zampini } 304308122e43SStefano Zampini } else { 304408122e43SStefano Zampini /* TODO */ 304508122e43SStefano Zampini } 3046aff50787SStefano Zampini } 30476c3e6151SStefano Zampini /* change the basis back to the original one */ 30486c3e6151SStefano Zampini if (sub_schurs->change) { 304972b8c272SStefano Zampini Mat change,phi,phit; 30506c3e6151SStefano Zampini 30516c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 30526c3e6151SStefano Zampini PetscInt ii; 30536c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 30546c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 30556c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 3056684229deSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3057684229deSStefano Zampini PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]); 3058684229deSStefano Zampini PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]); 3059684229deSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3060684229deSStefano Zampini #else 30616c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 3062684229deSStefano Zampini #endif 30636c3e6151SStefano Zampini } 30646c3e6151SStefano Zampini } 30656c3e6151SStefano Zampini } 306672b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 30676c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 306872b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 30696c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 30706c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 30716c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 30726c3e6151SStefano Zampini } 30738bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 30748bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 30759162d606SStefano Zampini if (B_neigs) { 30769162d606SStefano 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); 3077fd14bc51SStefano Zampini 3078fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 30799552c7c7SStefano Zampini PetscInt ii; 30809552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 3081ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 30829552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 3083ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 3084ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3085ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 3086ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 3087ac47001eSStefano Zampini #else 3088ac47001eSStefano 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); 3089ac47001eSStefano Zampini #endif 30909552c7c7SStefano Zampini } 30919552c7c7SStefano Zampini } 3092fd14bc51SStefano Zampini } 30939162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 30949162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 30959162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 30969162d606SStefano Zampini cum++; 309708122e43SStefano Zampini } 309808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 309908122e43SStefano Zampini /* shift for next computation */ 310008122e43SStefano Zampini cumarray += subset_size*subset_size; 310108122e43SStefano Zampini } 3102fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 3103fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3104fd14bc51SStefano Zampini } 310508122e43SStefano Zampini 310608122e43SStefano Zampini if (mss) { 310708122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 310808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 3109f6f667cfSStefano Zampini /* destroy matrices (junk) */ 3110f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 3111f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 311208122e43SStefano Zampini } 3113f6f667cfSStefano Zampini if (allocated_S_St) { 3114f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 3115f6f667cfSStefano Zampini } 3116f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 311708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 311808122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 311908122e43SStefano Zampini #endif 312008122e43SStefano Zampini if (pcbddc->dbg_flag) { 31211b968477SStefano Zampini PetscInt maxneigs_r; 3122b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 31239b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 312408122e43SStefano Zampini } 312508122e43SStefano Zampini PetscFunctionReturn(0); 312608122e43SStefano Zampini } 3127b1b3d7a2SStefano Zampini 3128c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 3129c8587f34SStefano Zampini { 31308629588bSStefano Zampini PetscScalar *coarse_submat_vals; 3131c8587f34SStefano Zampini PetscErrorCode ierr; 3132c8587f34SStefano Zampini 3133c8587f34SStefano Zampini PetscFunctionBegin; 3134f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 31355e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 3136c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 3137c8587f34SStefano Zampini 3138684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 31390fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 3140684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3141c8587f34SStefano Zampini 31428629588bSStefano Zampini /* 31438629588bSStefano Zampini Setup local correction and local part of coarse basis. 31448629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 31458629588bSStefano Zampini */ 314647f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 31478629588bSStefano Zampini 31488629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 31498629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 31508629588bSStefano Zampini 31518629588bSStefano Zampini /* free */ 31528629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 3153c8587f34SStefano Zampini PetscFunctionReturn(0); 3154c8587f34SStefano Zampini } 3155c8587f34SStefano Zampini 3156674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 3157674ae819SStefano Zampini { 3158674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3159674ae819SStefano Zampini PetscErrorCode ierr; 3160674ae819SStefano Zampini 3161674ae819SStefano Zampini PetscFunctionBegin; 3162674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 316330368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 3164674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 3165785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3166674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 3167f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3168f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3169785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 317063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 317163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 3172674ae819SStefano Zampini PetscFunctionReturn(0); 3173674ae819SStefano Zampini } 3174674ae819SStefano Zampini 3175674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 3176674ae819SStefano Zampini { 3177674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 31784f1b2e48SStefano Zampini PetscInt i; 3179674ae819SStefano Zampini PetscErrorCode ierr; 3180674ae819SStefano Zampini 3181674ae819SStefano Zampini PetscFunctionBegin; 31821e0482f5SStefano Zampini ierr = MatDestroy(&pcbddc->nedcG);CHKERRQ(ierr); 31831e0482f5SStefano Zampini ierr = ISDestroy(&pcbddc->nedclocal);CHKERRQ(ierr); 3184a13144ffSStefano Zampini ierr = MatDestroy(&pcbddc->discretegradient);CHKERRQ(ierr); 3185b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3186674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 318716909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 31881dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 3189674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3190669cc0f4SStefano Zampini ierr = MatDestroy(&pcbddc->divudotp);CHKERRQ(ierr); 3191fa23a32eSStefano Zampini ierr = ISDestroy(&pcbddc->divudotp_vl2l);CHKERRQ(ierr); 31929326c5c6Sstefano_zampini ierr = PCBDDCGraphDestroy(&pcbddc->mat_graph);CHKERRQ(ierr); 31934f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 31944f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 31954f1b2e48SStefano Zampini } 3196e68a0315Sstefano_zampini pcbddc->n_local_subs = 0; 31974f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 3198e68a0315Sstefano_zampini ierr = PCBDDCSubSchursDestroy(&pcbddc->sub_schurs);CHKERRQ(ierr); 3199c703fcc7SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 32008af8fcf9SStefano Zampini pcbddc->recompute_topography = PETSC_TRUE; 3201674ae819SStefano Zampini PetscFunctionReturn(0); 3202674ae819SStefano Zampini } 3203674ae819SStefano Zampini 3204674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 3205674ae819SStefano Zampini { 3206674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3207674ae819SStefano Zampini PetscErrorCode ierr; 3208674ae819SStefano Zampini 3209674ae819SStefano Zampini PetscFunctionBegin; 3210674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 321158da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 3212ca92afb2SStefano Zampini PetscScalar *array; 321306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 321406656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 321558da7f69SStefano Zampini } 3216674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3217674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 321815aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 321915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3220674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 3221674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 3222674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 322306656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 3224674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3225674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 32268ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 3227674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 3228674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 3229674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 32309326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_D);CHKERRQ(ierr); 32319326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 32329326c5c6Sstefano_zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3233f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 3234727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 32350e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3236f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 323770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 323881d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 32390369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 32401dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 32414f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 32428b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 3243ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 3244ca92afb2SStefano Zampini PetscInt i; 3245ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 3246ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 3247ca92afb2SStefano Zampini } 3248ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 3249ca92afb2SStefano Zampini } 32504f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 3251674ae819SStefano Zampini PetscFunctionReturn(0); 3252674ae819SStefano Zampini } 3253674ae819SStefano Zampini 3254f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 32556bfb1811SStefano Zampini { 32566bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 32576bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 32586bfb1811SStefano Zampini VecType impVecType; 32594f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 32606bfb1811SStefano Zampini PetscErrorCode ierr; 32616bfb1811SStefano Zampini 32626bfb1811SStefano Zampini PetscFunctionBegin; 32634f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 3264b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 32656bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 3266e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 3267e7b262bdSStefano Zampini /* R nodes */ 3268e7b262bdSStefano Zampini old_size = -1; 3269e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 3270e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 3271e7b262bdSStefano Zampini } 3272e7b262bdSStefano Zampini if (n_R != old_size) { 3273e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 3274e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 32756bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 32766bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 32776bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 32786bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 3279e7b262bdSStefano Zampini } 3280e7b262bdSStefano Zampini /* local primal dofs */ 3281e7b262bdSStefano Zampini old_size = -1; 3282e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 3283e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 3284e7b262bdSStefano Zampini } 3285e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 3286e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 328783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 3288e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 32896bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 3290e7b262bdSStefano Zampini } 3291e7b262bdSStefano Zampini /* local explicit constraints */ 3292e7b262bdSStefano Zampini old_size = -1; 3293e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 3294e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 3295e7b262bdSStefano Zampini } 3296e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 3297e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 329883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 329983b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 330083b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 330183b7ccabSStefano Zampini } 33026bfb1811SStefano Zampini PetscFunctionReturn(0); 33036bfb1811SStefano Zampini } 33046bfb1811SStefano Zampini 330547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 330688ebb749SStefano Zampini { 330725084f0cSStefano Zampini PetscErrorCode ierr; 330825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 330988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 331088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3311d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 331225084f0cSStefano Zampini /* submatrices of local problem */ 331380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 331406656605SStefano Zampini /* submatrices of local coarse problem */ 331506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 331625084f0cSStefano Zampini /* working matrices */ 331706656605SStefano Zampini Mat C_CR; 331825084f0cSStefano Zampini /* additional working stuff */ 331906656605SStefano Zampini PC pc_R; 3320c58f9fdbSStefano Zampini Mat F,Brhs = NULL; 33215cbda25cSStefano Zampini Vec dummy_vec; 3322c58f9fdbSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction,sparserhs; 332325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 332406656605SStefano Zampini PetscScalar *work; 332506656605SStefano Zampini PetscInt *idx_V_B; 3326ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 332706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 3328ffd830a3SStefano Zampini 332925084f0cSStefano Zampini /* some shortcuts to scalars */ 333006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 333188ebb749SStefano Zampini 333288ebb749SStefano Zampini PetscFunctionBegin; 33339a962809SStefano 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"); 3334ffd830a3SStefano Zampini 3335ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 3336b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 33374f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 3338b371cd4fSStefano Zampini n_B = pcis->n_B; 3339b371cd4fSStefano Zampini n_D = pcis->n - n_B; 334088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 334188ebb749SStefano Zampini 334288ebb749SStefano Zampini /* vertices in boundary numbering */ 3343785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 33440e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 33456c4ed002SBarry 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); 334688ebb749SStefano Zampini 334706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 3348019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 334906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 335006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 335106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 335206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 335306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 335406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 335506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 335606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 335706656605SStefano Zampini 335806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 335906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 336006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 336106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 336206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 3363ffd830a3SStefano Zampini lda_rhs = n_R; 3364a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 336506656605SStefano Zampini if (isLU || isILU || isCHOL) { 336606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 3367b334f244SStefano Zampini } else if (sub_schurs && sub_schurs->reuse_solver) { 3368df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3369d62866d3SStefano Zampini MatFactorType type; 3370d62866d3SStefano Zampini 3371df4d28bfSStefano Zampini F = reuse_solver->F; 33726816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 3373d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 3374ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 337522db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 337606656605SStefano Zampini } else { 337706656605SStefano Zampini F = NULL; 337806656605SStefano Zampini } 337906656605SStefano Zampini 3380c58f9fdbSStefano Zampini /* determine if we can use a sparse right-hand side */ 3381c58f9fdbSStefano Zampini sparserhs = PETSC_FALSE; 3382c58f9fdbSStefano Zampini if (F) { 3383c58f9fdbSStefano Zampini const MatSolverPackage solver; 3384c58f9fdbSStefano Zampini 3385c58f9fdbSStefano Zampini ierr = MatFactorGetSolverPackage(F,&solver);CHKERRQ(ierr); 3386c58f9fdbSStefano Zampini ierr = PetscStrcmp(solver,MATSOLVERMUMPS,&sparserhs);CHKERRQ(ierr); 3387c58f9fdbSStefano Zampini } 3388c58f9fdbSStefano Zampini 3389ffd830a3SStefano Zampini /* allocate workspace */ 3390ffd830a3SStefano Zampini n = 0; 3391ffd830a3SStefano Zampini if (n_constraints) { 3392ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 3393ffd830a3SStefano Zampini } 3394ffd830a3SStefano Zampini if (n_vertices) { 3395ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 3396ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 3397ffd830a3SStefano Zampini } 33982a3a6641Sstefano_zampini if (!pcbddc->symmetric_primal) { 33992a3a6641Sstefano_zampini n = PetscMax(2*lda_rhs*pcbddc->local_primal_size,n); 34002a3a6641Sstefano_zampini } 3401ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 3402ffd830a3SStefano Zampini 34035cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 34045cbda25cSStefano Zampini dummy_vec = NULL; 34055cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 34065cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 34075cbda25cSStefano Zampini } 34085cbda25cSStefano Zampini 340988ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 341088ebb749SStefano Zampini if (n_constraints) { 341172b8c272SStefano Zampini Mat M1,M2,M3,C_B; 341206656605SStefano Zampini IS is_aux; 341380677318SStefano Zampini PetscScalar *array,*array2; 341406656605SStefano Zampini 3415f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 341680677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 341788ebb749SStefano Zampini 341825084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 341925084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 34207dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 34217dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 342288ebb749SStefano Zampini 342380677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 342480677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 3425c58f9fdbSStefano Zampini if (!sparserhs) { 3426ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 342788ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 342806656605SStefano Zampini const PetscScalar *row_cmat_values; 342906656605SStefano Zampini const PetscInt *row_cmat_indices; 343006656605SStefano Zampini PetscInt size_of_constraint,j; 343188ebb749SStefano Zampini 343206656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 343306656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 3434ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 343506656605SStefano Zampini } 343606656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 343706656605SStefano Zampini } 3438c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&Brhs);CHKERRQ(ierr); 3439c58f9fdbSStefano Zampini } else { 3440c58f9fdbSStefano Zampini Mat tC_CR; 3441c58f9fdbSStefano Zampini 3442c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 3443c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 3444c58f9fdbSStefano Zampini PetscScalar *aa; 3445c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 3446c58f9fdbSStefano Zampini PetscBool done; 3447c58f9fdbSStefano Zampini 3448c58f9fdbSStefano Zampini ierr = MatGetRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3449c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed");CHKERRQ(ierr); 3450c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(C_CR,&aa);CHKERRQ(ierr); 3451c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_constraints,lda_rhs,ii,jj,aa,&tC_CR);CHKERRQ(ierr); 3452c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3453c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed");CHKERRQ(ierr); 3454c58f9fdbSStefano Zampini } else { 3455c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)C_CR);CHKERRQ(ierr); 3456c58f9fdbSStefano Zampini tC_CR = C_CR; 3457c58f9fdbSStefano Zampini } 3458c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tC_CR,&Brhs);CHKERRQ(ierr); 3459c58f9fdbSStefano Zampini ierr = MatDestroy(&tC_CR);CHKERRQ(ierr); 3460c58f9fdbSStefano Zampini } 3461ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 346206656605SStefano Zampini if (F) { 3463a3df083aSStefano Zampini if (need_benign_correction) { 3464df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3465a3df083aSStefano Zampini 346672b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 346772b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 3468a3df083aSStefano Zampini } 3469c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,local_auxmat2_R);CHKERRQ(ierr); 3470a3df083aSStefano Zampini if (need_benign_correction) { 3471a3df083aSStefano Zampini PetscScalar *marr; 3472df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3473a3df083aSStefano Zampini 3474a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 34755cbda25cSStefano Zampini if (lda_rhs != n_R) { 34765cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 34775cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 34785cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 34795cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 34805cbda25cSStefano Zampini } 34815cbda25cSStefano Zampini } else { 3482a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 3483a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 34845cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3485a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3486a3df083aSStefano Zampini } 34875cbda25cSStefano Zampini } 3488a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 3489a3df083aSStefano Zampini } 349006656605SStefano Zampini } else { 349180677318SStefano Zampini PetscScalar *marr; 349280677318SStefano Zampini 349380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 349406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 3495ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3496ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 349706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 349806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 349906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 350006656605SStefano Zampini } 350180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 350206656605SStefano Zampini } 3503c58f9fdbSStefano Zampini if (sparserhs) { 3504c58f9fdbSStefano Zampini ierr = MatScale(C_CR,-1.0);CHKERRQ(ierr); 3505c58f9fdbSStefano Zampini } 3506c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 350780677318SStefano Zampini if (!pcbddc->switch_static) { 350880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 350980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 351080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 351180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 3512ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 351380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 351480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 351580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 351680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 351780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 351880677318SStefano Zampini } 351980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 352080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 352172b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 352280677318SStefano Zampini } else { 3523ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3524ffd830a3SStefano Zampini IS dummy; 3525ffd830a3SStefano Zampini 3526ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 35277dae84e0SHong Zhang ierr = MatCreateSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 3528ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 3529ffd830a3SStefano Zampini } else { 353080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 353180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 3532ffd830a3SStefano Zampini } 353325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 353480677318SStefano Zampini } 353580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 353680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 353780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 353806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 353906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 354080677318SStefano Zampini if (isCHOL) { 354180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 354280677318SStefano Zampini } else { 354325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 354480677318SStefano Zampini } 354580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 354606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 354725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 354825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 354925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 355080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 355172b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 355272b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 355306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 355406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 3555f4ddd8eeSStefano Zampini } 3556fc227af8SStefano Zampini 3557fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 355888ebb749SStefano Zampini if (n_vertices) { 355906656605SStefano Zampini IS is_aux; 3560c58f9fdbSStefano Zampini PetscBool isseqaij; 35613a50541eSStefano Zampini 3562b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 35636816873aSStefano Zampini IS tis; 35646816873aSStefano Zampini 35656816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 35666816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 35676816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 35686816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 35696816873aSStefano Zampini } else { 35703a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 35716816873aSStefano Zampini } 35727dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 35737dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 3574c58f9fdbSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3575c58f9fdbSStefano Zampini if (!isseqaij) { /* MatMatMult(A_VR,A_RRmA_RV) below will raise an error */ 3576c58f9fdbSStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 3577c58f9fdbSStefano Zampini } 35787dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 357925084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 358088ebb749SStefano Zampini } 358188ebb749SStefano Zampini 358288ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 3583f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 358406656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 358506656605SStefano Zampini if (pcbddc->coarse_phi_D) { 358606656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 358706656605SStefano Zampini } 3588f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 358906656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 359006656605SStefano Zampini PetscScalar *marray; 359106656605SStefano Zampini 359206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 359306656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 3594f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3595f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3596f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3597f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 3598f4ddd8eeSStefano Zampini } 3599f4ddd8eeSStefano Zampini } 360006656605SStefano Zampini 3601f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 3602a6e023c1Sstefano_zampini PetscScalar *marr; 360388ebb749SStefano Zampini 3604a6e023c1Sstefano_zampini /* memory size */ 360506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 3606a6e023c1Sstefano_zampini if (pcbddc->switch_static || pcbddc->dbg_flag) n += n_D*pcbddc->local_primal_size; 3607a6e023c1Sstefano_zampini if (!pcbddc->symmetric_primal) n *= 2; 3608a6e023c1Sstefano_zampini ierr = PetscCalloc1(n,&marr);CHKERRQ(ierr); 3609a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 3610a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 36118eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3612a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 3613a6e023c1Sstefano_zampini marr += n_D*pcbddc->local_primal_size; 361488ebb749SStefano Zampini } 36153301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 3616a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 3617a6e023c1Sstefano_zampini marr += n_B*pcbddc->local_primal_size; 36188eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3619a6e023c1Sstefano_zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 362088ebb749SStefano Zampini } 362188ebb749SStefano Zampini } else { 3622c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 3623c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 36241b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 3625c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 3626c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 3627c0553b1fSStefano Zampini } 362888ebb749SStefano Zampini } 362906656605SStefano Zampini } 3630019a44ceSStefano Zampini 363106656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 36324f1b2e48SStefano Zampini p0_lidx_I = NULL; 36334f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 3634d12edf2fSStefano Zampini const PetscInt *idxs; 3635d12edf2fSStefano Zampini 3636d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 36374f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 36384f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 36394f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 36404f1b2e48SStefano Zampini } 3641d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 3642d12edf2fSStefano Zampini } 3643d16cbb6bSStefano Zampini 364406656605SStefano Zampini /* vertices */ 364506656605SStefano Zampini if (n_vertices) { 3646c58f9fdbSStefano Zampini PetscBool restoreavr = PETSC_FALSE; 364716f15bc4SStefano Zampini 3648af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 364904708bb6SStefano Zampini 365016f15bc4SStefano Zampini if (n_R) { 365114393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 365206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 365316f15bc4SStefano Zampini PetscScalar *x,*y; 365406656605SStefano Zampini 365521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 365614393ed6SStefano Zampini if (need_benign_correction) { 365714393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 365814393ed6SStefano Zampini IS is_p0; 365914393ed6SStefano Zampini PetscInt *idxs_p0,n; 366014393ed6SStefano Zampini 366114393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 366214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 366314393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 3664af25d912SStefano 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); 366514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 366614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 36677dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 366814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 366914393ed6SStefano Zampini } 367014393ed6SStefano Zampini 3671c58f9fdbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3672c58f9fdbSStefano Zampini if (!sparserhs || need_benign_correction) { 3673ffd830a3SStefano Zampini if (lda_rhs == n_R) { 3674af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3675ffd830a3SStefano Zampini } else { 3676ca92afb2SStefano Zampini PetscScalar *av,*array; 3677ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 3678ca92afb2SStefano Zampini PetscInt n; 3679ca92afb2SStefano Zampini PetscBool flg_row; 3680ffd830a3SStefano Zampini 3681ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 3682ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 36839d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 3684ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3685ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 3686ca92afb2SStefano Zampini for (i=0;i<n;i++) { 3687ca92afb2SStefano Zampini PetscInt j; 3688ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 3689ffd830a3SStefano Zampini } 3690ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 3691ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3692ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 3693ffd830a3SStefano Zampini } 3694a3df083aSStefano Zampini if (need_benign_correction) { 3695df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3696a3df083aSStefano Zampini PetscScalar *marr; 3697a3df083aSStefano Zampini 3698a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 369914393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 370014393ed6SStefano Zampini 370114393ed6SStefano Zampini | 0 0 0 | (V) 370214393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 370314393ed6SStefano Zampini | 0 0 -1 | (p0) 370414393ed6SStefano Zampini 370514393ed6SStefano Zampini */ 3706df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 370714393ed6SStefano Zampini const PetscScalar *vals; 370814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 370914393ed6SStefano Zampini PetscInt n,j,nz; 371014393ed6SStefano Zampini 3711df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3712df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 371314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 371414393ed6SStefano Zampini for (j=0;j<n;j++) { 371514393ed6SStefano Zampini PetscScalar val = vals[j]; 371614393ed6SStefano Zampini PetscInt k,col = idxs[j]; 371714393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 371814393ed6SStefano Zampini } 371914393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3720df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 372114393ed6SStefano Zampini } 372272b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 372372b8c272SStefano Zampini } 3724c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RV);CHKERRQ(ierr); 3725c58f9fdbSStefano Zampini Brhs = A_RV; 3726c58f9fdbSStefano Zampini } else { 3727c58f9fdbSStefano Zampini Mat tA_RVT,A_RVT; 3728c58f9fdbSStefano Zampini 3729c58f9fdbSStefano Zampini if (!pcbddc->symmetric_primal) { 3730c58f9fdbSStefano Zampini ierr = MatTranspose(A_RV,MAT_INITIAL_MATRIX,&A_RVT);CHKERRQ(ierr); 3731c58f9fdbSStefano Zampini } else { 3732c58f9fdbSStefano Zampini restoreavr = PETSC_TRUE; 3733c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 3734c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_VR);CHKERRQ(ierr); 3735c58f9fdbSStefano Zampini A_RVT = A_VR; 3736c58f9fdbSStefano Zampini } 3737c58f9fdbSStefano Zampini if (lda_rhs != n_R) { 3738c58f9fdbSStefano Zampini PetscScalar *aa; 3739c58f9fdbSStefano Zampini PetscInt r,*ii,*jj; 3740c58f9fdbSStefano Zampini PetscBool done; 3741c58f9fdbSStefano Zampini 3742c58f9fdbSStefano Zampini ierr = MatGetRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3743c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed");CHKERRQ(ierr); 3744c58f9fdbSStefano Zampini ierr = MatSeqAIJGetArray(A_RVT,&aa);CHKERRQ(ierr); 3745c58f9fdbSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_vertices,lda_rhs,ii,jj,aa,&tA_RVT);CHKERRQ(ierr); 3746c58f9fdbSStefano Zampini ierr = MatRestoreRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);CHKERRQ(ierr); 3747c58f9fdbSStefano Zampini if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed");CHKERRQ(ierr); 3748c58f9fdbSStefano Zampini } else { 3749c58f9fdbSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RVT);CHKERRQ(ierr); 3750c58f9fdbSStefano Zampini tA_RVT = A_RVT; 3751c58f9fdbSStefano Zampini } 3752c58f9fdbSStefano Zampini ierr = MatCreateTranspose(tA_RVT,&Brhs);CHKERRQ(ierr); 3753c58f9fdbSStefano Zampini ierr = MatDestroy(&tA_RVT);CHKERRQ(ierr); 3754c58f9fdbSStefano Zampini ierr = MatDestroy(&A_RVT);CHKERRQ(ierr); 3755c58f9fdbSStefano Zampini } 375672b8c272SStefano Zampini if (F) { 375714393ed6SStefano Zampini /* need to correct the rhs */ 375872b8c272SStefano Zampini if (need_benign_correction) { 375972b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 376072b8c272SStefano Zampini PetscScalar *marr; 376172b8c272SStefano Zampini 3762c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&marr);CHKERRQ(ierr); 37635cbda25cSStefano Zampini if (lda_rhs != n_R) { 37645cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 37655cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 37665cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 37675cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 37685cbda25cSStefano Zampini } 37695cbda25cSStefano Zampini } else { 3770a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3771a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 37725cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 3773a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3774a3df083aSStefano Zampini } 37755cbda25cSStefano Zampini } 3776c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&marr);CHKERRQ(ierr); 3777a3df083aSStefano Zampini } 3778c58f9fdbSStefano Zampini ierr = MatMatSolve(F,Brhs,A_RRmA_RV);CHKERRQ(ierr); 3779c58f9fdbSStefano Zampini if (restoreavr) { 3780c58f9fdbSStefano Zampini ierr = MatScale(A_VR,-1.0);CHKERRQ(ierr); 3781c58f9fdbSStefano Zampini } 378214393ed6SStefano Zampini /* need to correct the solution */ 3783a3df083aSStefano Zampini if (need_benign_correction) { 3784df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3785a3df083aSStefano Zampini PetscScalar *marr; 3786a3df083aSStefano Zampini 3787a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 37885cbda25cSStefano Zampini if (lda_rhs != n_R) { 37895cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 37905cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 37915cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 37925cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 37935cbda25cSStefano Zampini } 37945cbda25cSStefano Zampini } else { 3795a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 3796a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 37975cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 3798a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 3799a3df083aSStefano Zampini } 38005cbda25cSStefano Zampini } 3801a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 3802a3df083aSStefano Zampini } 380306656605SStefano Zampini } else { 3804c58f9fdbSStefano Zampini ierr = MatDenseGetArray(Brhs,&y);CHKERRQ(ierr); 380506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 3806ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 3807ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 380806656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 380906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 381006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 381106656605SStefano Zampini } 3812c58f9fdbSStefano Zampini ierr = MatDenseRestoreArray(Brhs,&y);CHKERRQ(ierr); 381306656605SStefano Zampini } 381480677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 3815c58f9fdbSStefano Zampini ierr = MatDestroy(&Brhs);CHKERRQ(ierr); 3816ffd830a3SStefano Zampini /* S_VV and S_CV */ 381706656605SStefano Zampini if (n_constraints) { 381806656605SStefano Zampini Mat B; 381980677318SStefano Zampini 3820ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 382180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 3822ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 3823ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 382480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 382580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 382680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 382780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 382880677318SStefano Zampini } 3829ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 383080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 383180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3832ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 383380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 383406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 3835ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 3836ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 383706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 383806656605SStefano Zampini } 3839ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3840ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 3841ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 3842ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 3843ffd830a3SStefano Zampini } 384406656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 384514393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 384614393ed6SStefano Zampini if (need_benign_correction) { 3847df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 384814393ed6SStefano Zampini PetscScalar *marr,*sums; 384914393ed6SStefano Zampini 385014393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 3851f913dca9SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr);CHKERRQ(ierr); 3852df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 385314393ed6SStefano Zampini const PetscScalar *vals; 385414393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 385514393ed6SStefano Zampini PetscInt n,j,nz; 385614393ed6SStefano Zampini 3857df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 3858df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 385914393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 386014393ed6SStefano Zampini PetscInt k; 386114393ed6SStefano Zampini sums[j] = 0.; 386214393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 386314393ed6SStefano Zampini } 386414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 386514393ed6SStefano Zampini for (j=0;j<n;j++) { 386614393ed6SStefano Zampini PetscScalar val = vals[j]; 386714393ed6SStefano Zampini PetscInt k; 386814393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 386914393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 387014393ed6SStefano Zampini } 387114393ed6SStefano Zampini } 387214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 3873df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 387414393ed6SStefano Zampini } 387514393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 3876f913dca9SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr);CHKERRQ(ierr); 387714393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 387814393ed6SStefano Zampini } 387980677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 388006656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 388106656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 388206656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 388306656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 388406656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 388506656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 388606656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3887d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 3888019a44ceSStefano Zampini } else { 3889d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3890d16cbb6bSStefano Zampini } 389121eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 3892d16cbb6bSStefano Zampini 389306656605SStefano Zampini /* coarse basis functions */ 389406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 389516f15bc4SStefano Zampini PetscScalar *y; 389616f15bc4SStefano Zampini 3897ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 389806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 389906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 390006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 390106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 390206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 390306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 390406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 390506656605SStefano Zampini 390606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 39074f1b2e48SStefano Zampini PetscInt j; 39084f1b2e48SStefano Zampini 390906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 391006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 391106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 391206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 391306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 39144f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 391506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 391606656605SStefano Zampini } 391706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 391806656605SStefano Zampini } 391904708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 392004708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 392106656605SStefano Zampini } 39225cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 392306656605SStefano Zampini 392406656605SStefano Zampini if (n_constraints) { 392506656605SStefano Zampini Mat B; 392606656605SStefano Zampini 3927ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 392806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 392980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 393006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 393106656605SStefano Zampini if (n_vertices) { 393280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 393380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 393480677318SStefano Zampini } else { 393580677318SStefano Zampini Mat S_VCt; 393680677318SStefano Zampini 3937ffd830a3SStefano Zampini if (lda_rhs != n_R) { 3938ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 393972b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 3940ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 3941ffd830a3SStefano Zampini } 394280677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 394380677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 394480677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 394580677318SStefano Zampini } 394606656605SStefano Zampini } 394706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 394806656605SStefano Zampini /* coarse basis functions */ 394906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 395006656605SStefano Zampini PetscScalar *y; 395106656605SStefano Zampini 3952ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 395306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 395406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 395506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 395606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 395706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 395806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 395906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 39604f1b2e48SStefano Zampini PetscInt j; 39614f1b2e48SStefano Zampini 396206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 396306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 396406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 396506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 396606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 39674f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 396806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 396906656605SStefano Zampini } 397006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 397106656605SStefano Zampini } 397206656605SStefano Zampini } 397380677318SStefano Zampini if (n_constraints) { 397480677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 397580677318SStefano Zampini } 39764f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 397772b8c272SStefano Zampini 397872b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 397972b8c272SStefano Zampini if (pcbddc->benign_n) { 398072b8c272SStefano Zampini Mat B0_B,B0_BPHI; 398172b8c272SStefano Zampini IS is_dummy; 398272b8c272SStefano Zampini PetscScalar *data; 398372b8c272SStefano Zampini PetscInt j; 398472b8c272SStefano Zampini 398572b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 39867dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 398772b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 398872b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 398986c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 399072b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 399172b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 399272b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 399372b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 399472b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 399572b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 399672b8c272SStefano Zampini } 399772b8c272SStefano Zampini } 399872b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 399972b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 400072b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 400172b8c272SStefano Zampini } 4002019a44ceSStefano Zampini 400306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 40043301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 4005ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 4006ffd830a3SStefano Zampini PetscScalar *marray; 400706656605SStefano Zampini 400806656605SStefano Zampini if (n_constraints) { 4009ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 401006656605SStefano Zampini 4011abc8f43dSstefano_zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 401206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 4013ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 401416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 401506656605SStefano Zampini if (n_vertices) { 4016ffd830a3SStefano Zampini Mat S_VCT; 401706656605SStefano Zampini 401806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 4019ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 402016f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 402106656605SStefano Zampini } 4022ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 40235b782168SStefano Zampini } else { 40245b782168SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);CHKERRQ(ierr); 402506656605SStefano Zampini } 402616f15bc4SStefano Zampini if (n_vertices && n_R) { 4027ffd830a3SStefano Zampini PetscScalar *av,*marray; 4028ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 4029ffd830a3SStefano Zampini PetscInt n; 4030ffd830a3SStefano Zampini PetscBool flg_row; 403106656605SStefano Zampini 4032ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 4033af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 4034ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4035ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 4036ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4037ffd830a3SStefano Zampini for (i=0;i<n;i++) { 4038ffd830a3SStefano Zampini PetscInt j; 4039ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 4040ffd830a3SStefano Zampini } 4041ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4042ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4043ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 404406656605SStefano Zampini } 404506656605SStefano Zampini 4046ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 4047abc8f43dSstefano_zampini if (n_vertices) { 4048ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 4049ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 4050ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 4051ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 405206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 405306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 405406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 405506656605SStefano Zampini } 4056ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 4057abc8f43dSstefano_zampini } 40585b782168SStefano Zampini if (B_C) { 4059ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 4060ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 4061ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 4062ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 4063ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 4064ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 4065ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 406606656605SStefano Zampini } 4067ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 40685b782168SStefano Zampini } 406906656605SStefano Zampini /* coarse basis functions */ 407006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 407106656605SStefano Zampini PetscScalar *y; 407206656605SStefano Zampini 4073ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 407406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 407506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 407606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 407706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 407806656605SStefano Zampini if (i<n_vertices) { 407906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 408006656605SStefano Zampini } 408106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 408206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 408306656605SStefano Zampini 408406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 408506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 408606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 408706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 408906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 409006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 409106656605SStefano Zampini } 409206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 409306656605SStefano Zampini } 4094ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 4095ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 409606656605SStefano Zampini } 4097a6e023c1Sstefano_zampini 4098d62866d3SStefano Zampini /* free memory */ 409988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 410006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 410106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 410206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 410306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 4104d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 4105d62866d3SStefano Zampini if (n_vertices) { 4106d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 4107d62866d3SStefano Zampini } 4108d62866d3SStefano Zampini if (n_constraints) { 4109d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 4110d62866d3SStefano Zampini } 411188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 411288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 411388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 4114d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 411588ebb749SStefano Zampini Mat coarse_sub_mat; 411625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 411788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 411888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 411988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 41208bec7fa6SStefano Zampini Mat C_B,CPHI; 41218bec7fa6SStefano Zampini IS is_dummy; 41228bec7fa6SStefano Zampini Vec mones; 412388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 412488ebb749SStefano Zampini PetscReal real_value; 412588ebb749SStefano Zampini 4126a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 4127a3df083aSStefano Zampini Mat A; 4128a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 41297dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 41307dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 41317dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 41327dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4133a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 4134a3df083aSStefano Zampini } else { 413588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 413688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 413788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 413888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 4139a3df083aSStefano Zampini } 414088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 414188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 4142ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 414388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 414488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 414588ebb749SStefano Zampini } 414688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 414788ebb749SStefano Zampini 414825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41493301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 415025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4151ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 415288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 415388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 415488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 415588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 415688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 415788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 415888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 415988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 416088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 416188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 416288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 416388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 416488ebb749SStefano Zampini } else { 416588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 416688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 416788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 416888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 416988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 417088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 417188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 417288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 417388ebb749SStefano Zampini } 417488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 417588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 417688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 4177511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 41784f1b2e48SStefano Zampini if (pcbddc->benign_n) { 4179fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 4180d12edf2fSStefano Zampini PetscScalar *data,*data2; 41814f1b2e48SStefano Zampini PetscInt j; 4182d12edf2fSStefano Zampini 41834f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 41847dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 4185d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 418686c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 4187d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 4188d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 41894f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 41904f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 4191d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 41924f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 41934f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 41944f1b2e48SStefano Zampini } 4195d12edf2fSStefano Zampini } 4196d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 4197d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 4198d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 4199d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 4200d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 4201d12edf2fSStefano Zampini } 4202d12edf2fSStefano Zampini #if 0 4203d12edf2fSStefano Zampini { 4204d12edf2fSStefano Zampini PetscViewer viewer; 4205d12edf2fSStefano Zampini char filename[256]; 4206ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4207d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4208d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4209ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 4210ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 4211ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 4212d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 421372b8c272SStefano Zampini if (save_change) { 421472b8c272SStefano Zampini Mat phi_B; 421572b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 421672b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 421772b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 421872b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 421972b8c272SStefano Zampini } else { 4220ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 4221ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 422272b8c272SStefano Zampini } 4223ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 4224ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 4225ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 4226ffd830a3SStefano Zampini } 4227ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 4228ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 4229ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 4230ffd830a3SStefano Zampini } 423172b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 4232ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 4233ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 4234ffd830a3SStefano Zampini } 4235d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4236d12edf2fSStefano Zampini } 4237d12edf2fSStefano Zampini #endif 423881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 42398bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 42401575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 424106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 42428bec7fa6SStefano Zampini 42438bec7fa6SStefano Zampini /* check constraints */ 4244a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 42457dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 42464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 42478bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4248a00504b5SStefano Zampini } else { 4249a00504b5SStefano Zampini PetscScalar *data; 4250a00504b5SStefano Zampini Mat tmat; 4251a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4252a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 4253a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 4254a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4255a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4256a00504b5SStefano Zampini } 42578bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 42588bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 42598bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 42608bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4261bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 4262ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 4263bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 4264bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 4265bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 4266bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 4267bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 426888ebb749SStefano Zampini } 42698bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 42708bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 42718bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 42728bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 427325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 427488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 427588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 427688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 427788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 427888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 427988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 428088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 428188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 428288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 428388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 4284ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 428588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 428688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 428788ebb749SStefano Zampini } 428888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 428988ebb749SStefano Zampini } 42908629588bSStefano Zampini /* get back data */ 42918629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 429288ebb749SStefano Zampini PetscFunctionReturn(0); 429388ebb749SStefano Zampini } 429488ebb749SStefano Zampini 42957dae84e0SHong Zhang PetscErrorCode MatCreateSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 4296aa0d41d4SStefano Zampini { 4297d65f70fdSStefano Zampini Mat *work_mat; 4298d65f70fdSStefano Zampini IS isrow_s,iscol_s; 4299d65f70fdSStefano Zampini PetscBool rsorted,csorted; 4300c43ebad9SStefano Zampini PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL; 4301aa0d41d4SStefano Zampini PetscErrorCode ierr; 4302aa0d41d4SStefano Zampini 4303aa0d41d4SStefano Zampini PetscFunctionBegin; 4304d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 4305d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 4306d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 4307d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 4308aa0d41d4SStefano Zampini 4309d65f70fdSStefano Zampini if (!rsorted) { 4310906d46d4SStefano Zampini const PetscInt *idxs; 4311906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 4312aa0d41d4SStefano Zampini 4313d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 4314d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 4315d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4316d65f70fdSStefano Zampini idxs_perm_r[i] = i; 4317aa0d41d4SStefano Zampini } 4318d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 4319d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 4320d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4321d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 4322aa0d41d4SStefano Zampini } 4323d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 4324d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 4325d65f70fdSStefano Zampini } else { 4326d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 4327d65f70fdSStefano Zampini isrow_s = isrow; 4328aa0d41d4SStefano Zampini } 4329906d46d4SStefano Zampini 4330d65f70fdSStefano Zampini if (!csorted) { 4331d65f70fdSStefano Zampini if (isrow == iscol) { 4332d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 4333d65f70fdSStefano Zampini iscol_s = isrow_s; 4334d65f70fdSStefano Zampini } else { 4335d65f70fdSStefano Zampini const PetscInt *idxs; 4336d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 4337906d46d4SStefano Zampini 4338d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 4339d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 4340d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4341d65f70fdSStefano Zampini idxs_perm_c[i] = i; 4342d65f70fdSStefano Zampini } 4343d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 4344d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 4345d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4346d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 4347d65f70fdSStefano Zampini } 4348d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 4349d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 4350d65f70fdSStefano Zampini } 4351d65f70fdSStefano Zampini } else { 4352d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 4353d65f70fdSStefano Zampini iscol_s = iscol; 4354d65f70fdSStefano Zampini } 4355d65f70fdSStefano Zampini 43567dae84e0SHong Zhang ierr = MatCreateSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4357d65f70fdSStefano Zampini 4358d65f70fdSStefano Zampini if (!rsorted || !csorted) { 4359906d46d4SStefano Zampini Mat new_mat; 4360d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 4361906d46d4SStefano Zampini 4362d65f70fdSStefano Zampini if (!rsorted) { 4363d65f70fdSStefano Zampini PetscInt *idxs_r,i; 4364d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 4365d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 4366d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 4367906d46d4SStefano Zampini } 4368d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 4369d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 4370d65f70fdSStefano Zampini } else { 4371d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 4372906d46d4SStefano Zampini } 4373d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 4374d65f70fdSStefano Zampini 4375d65f70fdSStefano Zampini if (!csorted) { 4376d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 4377d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 4378d65f70fdSStefano Zampini is_perm_c = is_perm_r; 4379d65f70fdSStefano Zampini } else { 4380d65f70fdSStefano Zampini PetscInt *idxs_c,i; 4381f913dca9SStefano Zampini if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present"); 4382d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 4383d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 4384d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 4385d65f70fdSStefano Zampini } 4386d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 4387d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 4388d65f70fdSStefano Zampini } 4389d65f70fdSStefano Zampini } else { 4390d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 4391d65f70fdSStefano Zampini } 4392d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 4393d65f70fdSStefano Zampini 4394d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 4395d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 4396d65f70fdSStefano Zampini work_mat[0] = new_mat; 4397d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 4398d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 4399d65f70fdSStefano Zampini } 4400d65f70fdSStefano Zampini 4401d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 4402d65f70fdSStefano Zampini *B = work_mat[0]; 4403d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 4404d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 4405d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 4406d65f70fdSStefano Zampini PetscFunctionReturn(0); 4407d65f70fdSStefano Zampini } 4408d65f70fdSStefano Zampini 44095e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 4410aa0d41d4SStefano Zampini { 4411aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 44125e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4413022d8d2bSstefano_zampini Mat new_mat,lA; 44145e8657edSStefano Zampini IS is_local,is_global; 4415d65f70fdSStefano Zampini PetscInt local_size; 4416d65f70fdSStefano Zampini PetscBool isseqaij; 4417aa0d41d4SStefano Zampini PetscErrorCode ierr; 4418aa0d41d4SStefano Zampini 4419aa0d41d4SStefano Zampini PetscFunctionBegin; 4420aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 44215e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 44225e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 4423b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 4424aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 44257dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 4426aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 4427906d46d4SStefano Zampini 4428906d46d4SStefano Zampini /* check */ 4429906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4430906d46d4SStefano Zampini Vec x,x_change; 4431906d46d4SStefano Zampini PetscReal error; 4432906d46d4SStefano Zampini 44335e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 4434906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 44355e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 4436e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4437e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4438d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 443988428137SStefano Zampini if (!pcbddc->change_interior) { 444088428137SStefano Zampini const PetscScalar *x,*y,*v; 444188428137SStefano Zampini PetscReal lerror = 0.; 444288428137SStefano Zampini PetscInt i; 444388428137SStefano Zampini 444488428137SStefano Zampini ierr = VecGetArrayRead(matis->x,&x);CHKERRQ(ierr); 444588428137SStefano Zampini ierr = VecGetArrayRead(matis->y,&y);CHKERRQ(ierr); 444688428137SStefano Zampini ierr = VecGetArrayRead(matis->counter,&v);CHKERRQ(ierr); 444788428137SStefano Zampini for (i=0;i<local_size;i++) 444888428137SStefano Zampini if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror) 444988428137SStefano Zampini lerror = PetscAbsScalar(x[i]-y[i]); 445088428137SStefano Zampini ierr = VecRestoreArrayRead(matis->x,&x);CHKERRQ(ierr); 445188428137SStefano Zampini ierr = VecRestoreArrayRead(matis->y,&y);CHKERRQ(ierr); 445288428137SStefano Zampini ierr = VecRestoreArrayRead(matis->counter,&v);CHKERRQ(ierr); 445388428137SStefano Zampini ierr = MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4454637e8532SStefano Zampini if (error > PETSC_SMALL) { 4455637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4456637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on I: %1.6e\n",error); 4457637e8532SStefano Zampini } else { 4458637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on I: %1.6e\n",error); 4459637e8532SStefano Zampini } 4460637e8532SStefano Zampini } 446188428137SStefano Zampini } 4462e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4463e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4464906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4465906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4466637e8532SStefano Zampini if (error > PETSC_SMALL) { 4467637e8532SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) { 4468637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 4469637e8532SStefano Zampini } else { 4470637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on N: %1.6e\n",error); 4471637e8532SStefano Zampini } 4472637e8532SStefano Zampini } 4473906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4474906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4475906d46d4SStefano Zampini } 4476906d46d4SStefano Zampini 4477022d8d2bSstefano_zampini /* lA is present if we are setting up an inner BDDC for a saddle point FETI-DP */ 4478022d8d2bSstefano_zampini ierr = PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject*)&lA);CHKERRQ(ierr); 4479022d8d2bSstefano_zampini 448022d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 44819b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 448222d5777bSStefano Zampini if (isseqaij) { 4483a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4484a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 4485022d8d2bSstefano_zampini if (lA) { 4486022d8d2bSstefano_zampini Mat work; 4487022d8d2bSstefano_zampini ierr = MatPtAP(lA,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 4488022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 4489022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 4490022d8d2bSstefano_zampini } 4491aa0d41d4SStefano Zampini } else { 4492a00504b5SStefano Zampini Mat work_mat; 44931cf9b237SStefano Zampini 4494a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4495aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4496a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 44971d82a3b6SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 4498022d8d2bSstefano_zampini if (lA) { 4499022d8d2bSstefano_zampini Mat work; 4500022d8d2bSstefano_zampini ierr = MatConvert(lA,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 4501022d8d2bSstefano_zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&work);CHKERRQ(ierr); 4502022d8d2bSstefano_zampini ierr = PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);CHKERRQ(ierr); 4503022d8d2bSstefano_zampini ierr = MatDestroy(&work);CHKERRQ(ierr); 4504022d8d2bSstefano_zampini } 4505aa0d41d4SStefano Zampini } 45063301b35fSStefano Zampini if (matis->A->symmetric_set) { 45073301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 4508e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 45093301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 4510e496cd5dSStefano Zampini #endif 45113301b35fSStefano Zampini } 4512d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 4513aa0d41d4SStefano Zampini PetscFunctionReturn(0); 4514aa0d41d4SStefano Zampini } 4515aa0d41d4SStefano Zampini 45168ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 4517a64d13efSStefano Zampini { 4518a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 4519a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4520d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 452153892102SStefano Zampini PetscInt *idx_R_local=NULL; 45223a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 45233a50541eSStefano Zampini PetscInt vbs,bs; 45246816873aSStefano Zampini PetscBT bitmask=NULL; 4525a64d13efSStefano Zampini PetscErrorCode ierr; 4526a64d13efSStefano Zampini 4527a64d13efSStefano Zampini PetscFunctionBegin; 4528b23d619eSStefano Zampini /* 4529b23d619eSStefano Zampini No need to setup local scatters if 4530b23d619eSStefano Zampini - primal space is unchanged 4531b23d619eSStefano Zampini AND 4532b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 4533b23d619eSStefano Zampini AND 4534b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 4535b23d619eSStefano Zampini */ 4536b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 4537f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 4538f4ddd8eeSStefano Zampini } 4539f4ddd8eeSStefano Zampini /* destroy old objects */ 4540f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 4541f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 4542f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 4543a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 4544b371cd4fSStefano Zampini n_B = pcis->n_B; 4545b371cd4fSStefano Zampini n_D = pcis->n - n_B; 4546b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 45473a50541eSStefano Zampini 4548a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 45496816873aSStefano Zampini 455053892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 4551b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4552854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 4553a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 4554a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 45550e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 4556a64d13efSStefano Zampini } 4557a64d13efSStefano Zampini 4558a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 45594641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 45606816873aSStefano Zampini idx_R_local[n_R++] = i; 4561a64d13efSStefano Zampini } 4562a64d13efSStefano Zampini } 4563df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 4564df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 45656816873aSStefano Zampini 4566df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4567df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 45686816873aSStefano Zampini } 45693a50541eSStefano Zampini 45703a50541eSStefano Zampini /* Block code */ 45713a50541eSStefano Zampini vbs = 1; 45723a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 45733a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 45743a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 45753a50541eSStefano Zampini PetscInt *vary; 4576b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 4577785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 45783a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 4579d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 4580d3df7717SStefano 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 */ 45810e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 4582d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 45833a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 45843a50541eSStefano Zampini is_blocked = PETSC_FALSE; 45853a50541eSStefano Zampini break; 45863a50541eSStefano Zampini } 45873a50541eSStefano Zampini } 4588d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 4589d3df7717SStefano Zampini } else { 4590d3df7717SStefano Zampini /* Verify directly the R set */ 4591d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 4592d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 4593d3df7717SStefano Zampini for (j=1; j<bs; j++) { 4594d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 4595d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 4596d3df7717SStefano Zampini break; 4597d3df7717SStefano Zampini } 4598d3df7717SStefano Zampini } 4599d3df7717SStefano Zampini } 4600d3df7717SStefano Zampini } 46013a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 46023a50541eSStefano Zampini vbs = bs; 46033a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 46043a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 46053a50541eSStefano Zampini } 46063a50541eSStefano Zampini } 46073a50541eSStefano Zampini } 46083a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 4609b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4610df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 461153892102SStefano Zampini 4612df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4613df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 461453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 4615df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 461653892102SStefano Zampini } else { 46173a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 461853892102SStefano Zampini } 4619a64d13efSStefano Zampini 4620a64d13efSStefano Zampini /* print some info if requested */ 4621a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 4622a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4623a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46241575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4625a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 4626a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 46274f1b2e48SStefano 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); 4628a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4629a64d13efSStefano Zampini } 4630a64d13efSStefano Zampini 4631a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 4632b334f244SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) { 46336816873aSStefano Zampini IS is_aux1,is_aux2; 46346816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 46356816873aSStefano Zampini 46363a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4637854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 4638854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 4639a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 46404641a718SStefano Zampini for (i=0; i<n_D; i++) { 46414641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 46424641a718SStefano Zampini } 4643a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4644a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 46454641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 46464641a718SStefano Zampini aux_array1[j++] = i; 4647a64d13efSStefano Zampini } 4648a64d13efSStefano Zampini } 4649a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4650a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4651a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 46524641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 46534641a718SStefano Zampini aux_array2[j++] = i; 4654a64d13efSStefano Zampini } 4655a64d13efSStefano Zampini } 4656a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4657a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 4658a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 4659a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4660a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 4661a64d13efSStefano Zampini 46628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 4663785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 4664a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 46654641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 46664641a718SStefano Zampini aux_array1[j++] = i; 4667a64d13efSStefano Zampini } 4668a64d13efSStefano Zampini } 4669a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 4670a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 4671a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 4672a64d13efSStefano Zampini } 46734641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 46743a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 4675d62866d3SStefano Zampini } else { 4676df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 46776816873aSStefano Zampini IS tis; 46786816873aSStefano Zampini PetscInt schur_size; 46796816873aSStefano Zampini 4680df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 46816816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 4682df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 46836816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 46846816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 46856816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 46866816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 46876816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 4688d62866d3SStefano Zampini } 4689d62866d3SStefano Zampini } 4690a64d13efSStefano Zampini PetscFunctionReturn(0); 4691a64d13efSStefano Zampini } 4692a64d13efSStefano Zampini 4693304d26faSStefano Zampini 4694684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 4695304d26faSStefano Zampini { 4696304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4697304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4698304d26faSStefano Zampini PC pc_temp; 4699304d26faSStefano Zampini Mat A_RR; 4700f4ddd8eeSStefano Zampini MatReuse reuse; 4701304d26faSStefano Zampini PetscScalar m_one = -1.0; 4702304d26faSStefano Zampini PetscReal value; 470304708bb6SStefano Zampini PetscInt n_D,n_R; 4704c7017625SStefano Zampini PetscBool check_corr[2],issbaij; 4705304d26faSStefano Zampini PetscErrorCode ierr; 4706e604994aSStefano Zampini /* prefixes stuff */ 4707312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 4708e604994aSStefano Zampini size_t len; 4709304d26faSStefano Zampini 4710304d26faSStefano Zampini PetscFunctionBegin; 4711304d26faSStefano Zampini 4712e604994aSStefano Zampini /* compute prefixes */ 4713e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 4714e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 4715e604994aSStefano Zampini if (!pcbddc->current_level) { 4716e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4717e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4718e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4719e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4720e604994aSStefano Zampini } else { 4721e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4722312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4723e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4724e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 4725312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4726312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 472734d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 472834d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4729e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 4730e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 4731e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 4732e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 4733e604994aSStefano Zampini } 4734e604994aSStefano Zampini 4735304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 4736684f6988SStefano Zampini if (dirichlet) { 4737d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4738450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 47399a962809SStefano Zampini if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 4740450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 4741a3df083aSStefano Zampini Mat A_IIn; 4742a3df083aSStefano Zampini 4743a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 4744a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 4745a3df083aSStefano Zampini pcis->A_II = A_IIn; 4746a3df083aSStefano Zampini } 4747450f8f5eSStefano Zampini } 47483301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 47493301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 4750964fefecSStefano Zampini } 4751ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 4752964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 4753304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 4754304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 4755304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 4756304d26faSStefano Zampini /* default */ 4757304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 4758e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 47599577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 4760304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 47619577ea80SStefano Zampini if (issbaij) { 47629577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 47639577ea80SStefano Zampini } else { 4764304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 47659577ea80SStefano Zampini } 4766304d26faSStefano Zampini /* Allow user's customization */ 4767304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 4768304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4769304d26faSStefano Zampini } 4770d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 4771b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4772df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4773d62866d3SStefano Zampini 4774df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 4775d5574798SStefano Zampini } 4776304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4777304d26faSStefano Zampini if (!n_D) { 4778304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 4779304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4780304d26faSStefano Zampini } 4781304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 4782304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 4783304d26faSStefano Zampini /* set ksp_D into pcis data */ 4784304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 4785304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 4786304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 4787684f6988SStefano Zampini } 4788304d26faSStefano Zampini 4789304d26faSStefano Zampini /* NEUMANN PROBLEM */ 4790684f6988SStefano Zampini A_RR = 0; 4791684f6988SStefano Zampini if (neumann) { 4792d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 479304708bb6SStefano Zampini PetscInt ibs,mbs; 479404708bb6SStefano Zampini PetscBool issbaij; 479504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4796f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 47978ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 4798f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 4799f4ddd8eeSStefano Zampini PetscInt nn_R; 480081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 4801f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4802f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 4803f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 4804f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 4805f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4806f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4807f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 4808727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 4809f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4810f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4811f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 4812f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 4813f4ddd8eeSStefano Zampini } 4814f4ddd8eeSStefano Zampini } 4815f4ddd8eeSStefano Zampini /* last check */ 4816d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 4817f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4818f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4819f4ddd8eeSStefano Zampini } 4820f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 4821f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 4822f4ddd8eeSStefano Zampini } 4823a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 4824af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 4825af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 482604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 482704708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 482804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 482904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 483004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4831af732b37SStefano Zampini } else { 4832511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 48336816873aSStefano Zampini } 483404708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 483504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 483604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 483704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483804708bb6SStefano Zampini } else { 4839511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 484004708bb6SStefano Zampini } 484104708bb6SStefano Zampini } 4842a00504b5SStefano Zampini /* extract A_RR */ 4843b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4844a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4845a00504b5SStefano Zampini 4846a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 484716e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4848a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 484916e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 485016e386b8SStefano Zampini } else { 48517dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 4852a00504b5SStefano Zampini } 4853a00504b5SStefano Zampini } else { 4854a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4855a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 4856a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 4857a00504b5SStefano Zampini } 4858a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 48597dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 486016e386b8SStefano Zampini } 48613301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 48623301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 48636816873aSStefano Zampini } 4864f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 4865304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 4866304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 4867304d26faSStefano Zampini /* default */ 4868304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 4869e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 4870304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 48719577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 48729577ea80SStefano Zampini if (issbaij) { 48739577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 48749577ea80SStefano Zampini } else { 4875304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 48769577ea80SStefano Zampini } 4877304d26faSStefano Zampini /* Allow user's customization */ 4878304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 4879304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4880304d26faSStefano Zampini } 4881304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 4882304d26faSStefano Zampini if (!n_R) { 4883304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 4884304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 4885304d26faSStefano Zampini } 48865cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 4887df4d28bfSStefano Zampini /* Reuse solver if it is present */ 4888b334f244SStefano Zampini if (sub_schurs && sub_schurs->reuse_solver) { 4889df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4890d62866d3SStefano Zampini 4891df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 4892d62866d3SStefano Zampini } 4893304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 4894304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 4895684f6988SStefano Zampini } 4896304d26faSStefano Zampini 4897684f6988SStefano Zampini if (pcbddc->dbg_flag) { 4898684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 48991575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 4900684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4901684f6988SStefano Zampini } 4902c7017625SStefano Zampini 4903c7017625SStefano Zampini /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */ 4904c7017625SStefano Zampini check_corr[0] = check_corr[1] = PETSC_FALSE; 4905c7017625SStefano Zampini if (pcbddc->NullSpace_corr[0]) { 4906c7017625SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);CHKERRQ(ierr); 4907c7017625SStefano Zampini } 4908c7017625SStefano Zampini if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) { 4909c7017625SStefano Zampini check_corr[0] = PETSC_TRUE; 4910c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);CHKERRQ(ierr); 4911c7017625SStefano Zampini } 4912c7017625SStefano Zampini if (neumann && pcbddc->NullSpace_corr[2]) { 4913c7017625SStefano Zampini check_corr[1] = PETSC_TRUE; 4914c7017625SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);CHKERRQ(ierr); 4915c7017625SStefano Zampini } 4916c7017625SStefano Zampini 4917c7017625SStefano Zampini /* check Dirichlet and Neumann solvers */ 4918c7017625SStefano Zampini if (pcbddc->dbg_flag) { 4919684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 49200fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 49210fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 49220fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 49230fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 49240fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 4925e604994aSStefano 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); 4926c7017625SStefano Zampini if (check_corr[0]) { 4927c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_TRUE);CHKERRQ(ierr); 4928c7017625SStefano Zampini } 4929304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4930304d26faSStefano Zampini } 4931684f6988SStefano Zampini if (neumann) { /* Neumann */ 49320fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 49330fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 49340fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 49350fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 49360fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 4937e604994aSStefano 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); 4938c7017625SStefano Zampini if (check_corr[1]) { 4939c7017625SStefano Zampini ierr = PCBDDCNullSpaceCheckCorrection(pc,PETSC_FALSE);CHKERRQ(ierr); 4940c7017625SStefano Zampini } 4941304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4942304d26faSStefano Zampini } 4943684f6988SStefano Zampini } 49445cbda25cSStefano Zampini /* free Neumann problem's matrix */ 49455cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 4946304d26faSStefano Zampini PetscFunctionReturn(0); 4947304d26faSStefano Zampini } 4948304d26faSStefano Zampini 494980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 4950674ae819SStefano Zampini { 4951674ae819SStefano Zampini PetscErrorCode ierr; 4952674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 4953be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 4954b334f244SStefano Zampini PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE; 4955674ae819SStefano Zampini 4956674ae819SStefano Zampini PetscFunctionBegin; 4957b334f244SStefano Zampini if (!reuse_solver) { 495880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 495920c7b377SStefano Zampini } 496080677318SStefano Zampini if (!pcbddc->switch_static) { 496180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 496280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 496380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 496420c7b377SStefano Zampini } 4965b334f244SStefano Zampini if (!reuse_solver) { 496680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 496780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 496820c7b377SStefano Zampini } else { 4969df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4970be83ff47SStefano Zampini 4971df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4972df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 497320c7b377SStefano Zampini } 4974be83ff47SStefano Zampini } else { 497580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 497980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 498080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 498180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 498280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 498380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4984674ae819SStefano Zampini } 4985674ae819SStefano Zampini } 4986b334f244SStefano Zampini if (!reuse_solver || pcbddc->switch_static) { 498780677318SStefano Zampini if (applytranspose) { 498880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 498980677318SStefano Zampini } else { 499080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 499180677318SStefano Zampini } 4992be83ff47SStefano Zampini } else { 4993df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 4994be83ff47SStefano Zampini 4995be83ff47SStefano Zampini if (applytranspose) { 4996df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4997be83ff47SStefano Zampini } else { 4998df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 4999be83ff47SStefano Zampini } 5000be83ff47SStefano Zampini } 500180677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 500280677318SStefano Zampini if (!pcbddc->switch_static) { 5003b334f244SStefano Zampini if (!reuse_solver) { 500480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 500580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5006be83ff47SStefano Zampini } else { 5007df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 5008be83ff47SStefano Zampini 5009df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5010df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5011be83ff47SStefano Zampini } 501280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 501380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 501480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 501580677318SStefano Zampini } 501680677318SStefano Zampini } else { 501780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 501880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 501980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 502280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 502380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 502480677318SStefano Zampini } 502580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 502880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5029674ae819SStefano Zampini } 5030674ae819SStefano Zampini PetscFunctionReturn(0); 5031674ae819SStefano Zampini } 5032674ae819SStefano Zampini 5033dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 5034dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 5035674ae819SStefano Zampini { 5036674ae819SStefano Zampini PetscErrorCode ierr; 5037674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 5038674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 5039674ae819SStefano Zampini const PetscScalar zero = 0.0; 5040674ae819SStefano Zampini 5041674ae819SStefano Zampini PetscFunctionBegin; 5042dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 50434fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5044dc359a40SStefano Zampini if (applytranspose) { 5045674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 50468eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 5047dc359a40SStefano Zampini } else { 5048674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 5049674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 505015aaf578SStefano Zampini } 50514fee134fSStefano Zampini } else { 50524fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 50534fee134fSStefano Zampini } 5054efc2fbd9SStefano Zampini 5055efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 50564f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5057efc2fbd9SStefano Zampini PetscScalar *array; 50584f1b2e48SStefano Zampini PetscInt j; 5059efc2fbd9SStefano Zampini 5060efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 50614f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 5062efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5063efc2fbd9SStefano Zampini } 5064efc2fbd9SStefano Zampini 506512edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 506612edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 506712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 506812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 506912edc857SStefano Zampini 50709f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 507112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 507251694757SStefano Zampini Mat coarse_mat; 5073964fefecSStefano Zampini Vec rhs,sol; 507451694757SStefano Zampini MatNullSpace nullsp; 507527b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 5076964fefecSStefano Zampini 507727b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 507827b6a85dSStefano Zampini PC coarse_pc; 507927b6a85dSStefano Zampini 508027b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 508127b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 508227b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 508327b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 508427b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 508527b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 50863bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 508727b6a85dSStefano Zampini } 508827b6a85dSStefano Zampini } 5089964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 5090964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 509151694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 509251694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 509351694757SStefano Zampini if (nullsp) { 509451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 509551694757SStefano Zampini } 509612edc857SStefano Zampini if (applytranspose) { 50979a962809SStefano Zampini if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented"); 5098964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 50992701bc32SStefano Zampini } else { 51001f4df5f7SStefano Zampini if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */ 51012701bc32SStefano Zampini PC coarse_pc; 51022701bc32SStefano Zampini 51032701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 51042701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 51053e589ea0SStefano Zampini ierr = PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);CHKERRQ(ierr); 51062701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 510712edc857SStefano Zampini } else { 5108964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 510912edc857SStefano Zampini } 51102701bc32SStefano Zampini } 51111d82a3b6SStefano Zampini /* we don't need the benign correction at coarser levels anymore */ 511227b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 511327b6a85dSStefano Zampini PC coarse_pc; 511427b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 511527b6a85dSStefano Zampini 511627b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 511727b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 511827b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 51193bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 512027b6a85dSStefano Zampini } 512151694757SStefano Zampini if (nullsp) { 512251694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 512351694757SStefano Zampini } 512412edc857SStefano Zampini } 5125674ae819SStefano Zampini 5126674ae819SStefano Zampini /* Local solution on R nodes */ 51274fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 512880677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 51299f00e9b4SStefano Zampini } 51309f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 51319f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 513212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5133674ae819SStefano Zampini 51344fee134fSStefano Zampini /* Sum contributions from the two levels */ 51354fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 5136dc359a40SStefano Zampini if (applytranspose) { 5137dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 5138dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5139dc359a40SStefano Zampini } else { 5140674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 51418eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 5142dc359a40SStefano Zampini } 5143efc2fbd9SStefano Zampini /* store p0 */ 51444f1b2e48SStefano Zampini if (pcbddc->benign_n) { 5145efc2fbd9SStefano Zampini PetscScalar *array; 51464f1b2e48SStefano Zampini PetscInt j; 5147efc2fbd9SStefano Zampini 5148efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 51494f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 5150efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 5151efc2fbd9SStefano Zampini } 51524fee134fSStefano Zampini } else { /* expand the coarse solution */ 51534fee134fSStefano Zampini if (applytranspose) { 51544fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 51554fee134fSStefano Zampini } else { 51564fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 51574fee134fSStefano Zampini } 51584fee134fSStefano Zampini } 5159674ae819SStefano Zampini PetscFunctionReturn(0); 5160674ae819SStefano Zampini } 5161674ae819SStefano Zampini 516212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 5163674ae819SStefano Zampini { 5164674ae819SStefano Zampini PetscErrorCode ierr; 5165674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 516658da7f69SStefano Zampini PetscScalar *array; 516712edc857SStefano Zampini Vec from,to; 5168674ae819SStefano Zampini 5169674ae819SStefano Zampini PetscFunctionBegin; 517012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 517112edc857SStefano Zampini from = pcbddc->coarse_vec; 517212edc857SStefano Zampini to = pcbddc->vec1_P; 517312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 517412edc857SStefano Zampini Vec tvec; 517558da7f69SStefano Zampini 517658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 517758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 517812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 517958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 518058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 518158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 518212edc857SStefano Zampini } 518312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 518412edc857SStefano Zampini from = pcbddc->vec1_P; 518512edc857SStefano Zampini to = pcbddc->coarse_vec; 518612edc857SStefano Zampini } 518712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 5188674ae819SStefano Zampini PetscFunctionReturn(0); 5189674ae819SStefano Zampini } 5190674ae819SStefano Zampini 519112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 5192674ae819SStefano Zampini { 5193674ae819SStefano Zampini PetscErrorCode ierr; 5194674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 519558da7f69SStefano Zampini PetscScalar *array; 519612edc857SStefano Zampini Vec from,to; 5197674ae819SStefano Zampini 5198674ae819SStefano Zampini PetscFunctionBegin; 519912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 520012edc857SStefano Zampini from = pcbddc->coarse_vec; 520112edc857SStefano Zampini to = pcbddc->vec1_P; 520212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 520312edc857SStefano Zampini from = pcbddc->vec1_P; 520412edc857SStefano Zampini to = pcbddc->coarse_vec; 520512edc857SStefano Zampini } 520612edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 520712edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 520812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 520912edc857SStefano Zampini Vec tvec; 521058da7f69SStefano Zampini 521112edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 521258da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 521358da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 521458da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 521558da7f69SStefano Zampini } 521658da7f69SStefano Zampini } else { 521758da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 521858da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 521912edc857SStefano Zampini } 522012edc857SStefano Zampini } 5221674ae819SStefano Zampini PetscFunctionReturn(0); 5222674ae819SStefano Zampini } 5223674ae819SStefano Zampini 5224984c4197SStefano Zampini /* uncomment for testing purposes */ 5225984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 5226674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 5227674ae819SStefano Zampini { 5228674ae819SStefano Zampini PetscErrorCode ierr; 5229674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 5230674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5231674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5232984c4197SStefano Zampini /* one and zero */ 5233984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 5234984c4197SStefano Zampini /* space to store constraints and their local indices */ 52359162d606SStefano Zampini PetscScalar *constraints_data; 52369162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 52379162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 52389162d606SStefano Zampini PetscInt *constraints_n; 5239984c4197SStefano Zampini /* iterators */ 5240b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 5241984c4197SStefano Zampini /* BLAS integers */ 5242e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 5243e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 5244c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 5245727cdba6SStefano Zampini /* reuse */ 52460e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 52470e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 5248984c4197SStefano Zampini /* change of basis */ 5249b3d85658SStefano Zampini PetscBool qr_needed; 52509162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 5251984c4197SStefano Zampini /* auxiliary stuff */ 525264efe560SStefano Zampini PetscInt *nnz,*is_indices; 52538a0068c3SStefano Zampini PetscInt ncc; 5254984c4197SStefano Zampini /* some quantities */ 525545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 5256a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 5257984c4197SStefano Zampini 5258674ae819SStefano Zampini PetscFunctionBegin; 52598e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 52608e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 52618e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 526216909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 5263088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 5264088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 52650e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 52660e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 52670e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 52680e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 52690e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 5270088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5271cf5a6209SStefano Zampini 5272cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 52739162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 5274cf5a6209SStefano Zampini MatNullSpace nearnullsp; 5275cf5a6209SStefano Zampini const Vec *nearnullvecs; 5276cf5a6209SStefano Zampini Vec *localnearnullsp; 5277cf5a6209SStefano Zampini PetscScalar *array; 5278cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 5279cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 5280674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 5281b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 5282674ae819SStefano Zampini PetscScalar *work; 5283674ae819SStefano Zampini PetscReal *singular_vals; 5284674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5285674ae819SStefano Zampini PetscReal *rwork; 5286674ae819SStefano Zampini #endif 5287674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5288674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 5289674ae819SStefano Zampini #else 5290964fefecSStefano Zampini PetscBLASInt dummy_int=1; 5291964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 5292674ae819SStefano Zampini #endif 5293674ae819SStefano Zampini 5294674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 5295d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 5296e4d548c7SStefano Zampini /* print some info */ 52975c643e28SStefano Zampini if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) { 5298e4d548c7SStefano Zampini PetscInt nv; 5299e4d548c7SStefano Zampini 5300c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5301e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 5302e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5303e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5304e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 5305e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 5306e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 5307e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5308e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5309e4d548c7SStefano Zampini } 5310e4d548c7SStefano Zampini 5311d06fc5fdSStefano Zampini /* free unneeded index sets */ 5312d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 5313d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 5314674ae819SStefano Zampini } 5315d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 5316d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5317d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5318d06fc5fdSStefano Zampini } 5319d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5320d06fc5fdSStefano Zampini n_ISForEdges = 0; 5321d06fc5fdSStefano Zampini } 5322d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 5323d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5324d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5325d06fc5fdSStefano Zampini } 5326d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5327d06fc5fdSStefano Zampini n_ISForFaces = 0; 5328d06fc5fdSStefano Zampini } 532970022509SStefano Zampini 5330674ae819SStefano Zampini /* check if near null space is attached to global mat */ 5331674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 5332674ae819SStefano Zampini if (nearnullsp) { 5333674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 5334f4ddd8eeSStefano Zampini /* remove any stored info */ 5335f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 5336f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5337f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 5338f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 5339f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 5340473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 5341f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 5342f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 5343f4ddd8eeSStefano Zampini } 5344984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 5345984c4197SStefano Zampini nnsp_size = 0; 5346674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 5347674ae819SStefano Zampini } 5348984c4197SStefano Zampini /* get max number of constraints on a single cc */ 5349984c4197SStefano Zampini max_constraints = nnsp_size; 5350984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 5351984c4197SStefano Zampini 5352674ae819SStefano Zampini /* 5353674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 53549162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 53559162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 53569162d606SStefano Zampini There can be multiple constraints per connected component 5357674ae819SStefano Zampini */ 5358674ae819SStefano Zampini n_vertices = 0; 5359674ae819SStefano Zampini if (ISForVertices) { 5360674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 5361674ae819SStefano Zampini } 53629162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 53639162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 53649162d606SStefano Zampini 53659162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 53669162d606SStefano Zampini total_counts *= max_constraints; 5367674ae819SStefano Zampini total_counts += n_vertices; 53684641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 53699162d606SStefano Zampini 5370674ae819SStefano Zampini total_counts = 0; 5371674ae819SStefano Zampini max_size_of_constraint = 0; 5372674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 53739162d606SStefano Zampini IS used_is; 5374674ae819SStefano Zampini if (i<n_ISForEdges) { 53759162d606SStefano Zampini used_is = ISForEdges[i]; 5376674ae819SStefano Zampini } else { 53779162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 5378674ae819SStefano Zampini } 53799162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 5380674ae819SStefano Zampini total_counts += j; 5381674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 5382674ae819SStefano Zampini } 53839162d606SStefano 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); 53849162d606SStefano Zampini 5385984c4197SStefano Zampini /* get local part of global near null space vectors */ 5386785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 5387984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5388984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 5389e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5390e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5391984c4197SStefano Zampini } 5392674ae819SStefano Zampini 5393242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 5394242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 5395a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 5396242a89d7SStefano Zampini 5397984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 5398a773dcb8SStefano Zampini if (!skip_lapack) { 5399674ae819SStefano Zampini PetscScalar temp_work; 5400911cabfeSStefano Zampini 5401674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5402984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 5403785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 5404785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 5405785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 5406674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5407785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 5408674ae819SStefano Zampini #endif 5409674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5410c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 5411c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 5412674ae819SStefano Zampini lwork = -1; 5413674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5414674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5415c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 5416674ae819SStefano Zampini #else 5417c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 5418674ae819SStefano Zampini #endif 5419674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5420984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 5421674ae819SStefano Zampini #else /* on missing GESVD */ 5422674ae819SStefano Zampini /* SVD */ 5423674ae819SStefano Zampini PetscInt max_n,min_n; 5424674ae819SStefano Zampini max_n = max_size_of_constraint; 5425984c4197SStefano Zampini min_n = max_constraints; 5426984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 5427674ae819SStefano Zampini min_n = max_size_of_constraint; 5428984c4197SStefano Zampini max_n = max_constraints; 5429674ae819SStefano Zampini } 5430785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 5431674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5432785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 5433674ae819SStefano Zampini #endif 5434674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 5435674ae819SStefano Zampini lwork = -1; 5436e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 5437e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 5438b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 5439674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5440674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 54419162d606SStefano 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)); 5442674ae819SStefano Zampini #else 54439162d606SStefano 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)); 5444674ae819SStefano Zampini #endif 5445674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5446984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 5447984c4197SStefano Zampini #endif /* on missing GESVD */ 5448674ae819SStefano Zampini /* Allocate optimal workspace */ 5449674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 5450854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 5451674ae819SStefano Zampini } 5452674ae819SStefano Zampini /* Now we can loop on constraining sets */ 5453674ae819SStefano Zampini total_counts = 0; 54549162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 54559162d606SStefano Zampini constraints_data_ptr[0] = 0; 5456674ae819SStefano Zampini /* vertices */ 54579162d606SStefano Zampini if (n_vertices) { 5458674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 54599162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 5460674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 54619162d606SStefano Zampini constraints_n[total_counts] = 1; 54629162d606SStefano Zampini constraints_data[total_counts] = 1.0; 54639162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 54649162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 5465674ae819SStefano Zampini total_counts++; 5466674ae819SStefano Zampini } 5467674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5468674ae819SStefano Zampini n_vertices = total_counts; 5469674ae819SStefano Zampini } 5470984c4197SStefano Zampini 5471674ae819SStefano Zampini /* edges and faces */ 54729162d606SStefano Zampini total_counts_cc = total_counts; 5473911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 54749162d606SStefano Zampini IS used_is; 54759162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 54769162d606SStefano Zampini 5477911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 54789162d606SStefano Zampini used_is = ISForEdges[ncc]; 5479984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 5480674ae819SStefano Zampini } else { 54819162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 5482984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 5483674ae819SStefano Zampini } 5484674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 54859162d606SStefano Zampini 54869162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 54879162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 5488984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 5489984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 5490674ae819SStefano Zampini if (nnsp_has_cnst) { 54915b08dc53SStefano Zampini PetscScalar quad_value; 54929162d606SStefano Zampini 54939162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 54949162d606SStefano Zampini idxs_copied = PETSC_TRUE; 54959162d606SStefano Zampini 5496a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 5497674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 5498a773dcb8SStefano Zampini } else { 5499a773dcb8SStefano Zampini quad_value = 1.0; 5500a773dcb8SStefano Zampini } 5501674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 55029162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 5503674ae819SStefano Zampini } 55049162d606SStefano Zampini temp_constraints++; 5505674ae819SStefano Zampini total_counts++; 5506674ae819SStefano Zampini } 5507674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 5508984c4197SStefano Zampini PetscReal real_value; 55099162d606SStefano Zampini PetscScalar *ptr_to_data; 55109162d606SStefano Zampini 5511984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 55129162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 5513674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 55149162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 5515674ae819SStefano Zampini } 5516984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 5517984c4197SStefano Zampini /* check if array is null on the connected component */ 5518e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 55199162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 55205b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 5521674ae819SStefano Zampini temp_constraints++; 5522674ae819SStefano Zampini total_counts++; 55239162d606SStefano Zampini if (!idxs_copied) { 55249162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 55259162d606SStefano Zampini idxs_copied = PETSC_TRUE; 5526674ae819SStefano Zampini } 5527674ae819SStefano Zampini } 55289162d606SStefano Zampini } 55299162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 553045a1bb75SStefano Zampini valid_constraints = temp_constraints; 5531eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 5532a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 55339162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 55349162d606SStefano Zampini 55359162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5536a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 55379162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 5538a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 55399162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 5540a773dcb8SStefano Zampini } else { /* perform SVD */ 5541984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 55429162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 5543674ae819SStefano Zampini 5544674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5545984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 5546984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 5547984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 5548984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 5549984c4197SStefano Zampini from that computed using LAPACKgesvd 5550984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 5551984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 5552984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 5553674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 5554e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 5555984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5556674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 5557674ae819SStefano Zampini for (k=0;k<j+1;k++) { 55589162d606SStefano 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)); 5559674ae819SStefano Zampini } 5560674ae819SStefano Zampini } 5561e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 5562e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5563e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 5564674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 5565c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 5566674ae819SStefano Zampini #else 5567c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 5568674ae819SStefano Zampini #endif 5569674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5570984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 5571984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 5572674ae819SStefano Zampini j = 0; 5573984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 5574674ae819SStefano Zampini total_counts = total_counts-j; 557545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 5576e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 5577c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5578c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5579c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 5580c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5581c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 5582c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 5583674ae819SStefano Zampini if (j<temp_constraints) { 5584984c4197SStefano Zampini PetscInt ii; 5585984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 5586674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 55879162d606SStefano 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)); 5588674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5589984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 5590674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 55919162d606SStefano 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]; 5592674ae819SStefano Zampini } 5593674ae819SStefano Zampini } 5594674ae819SStefano Zampini } 5595674ae819SStefano Zampini #else /* on missing GESVD */ 5596e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 5597e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 5598b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5599674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 5600674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 56019162d606SStefano 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)); 5602674ae819SStefano Zampini #else 56039162d606SStefano 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)); 5604674ae819SStefano Zampini #endif 5605984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 5606674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5607984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 5608e310c8b4SStefano Zampini k = temp_constraints; 5609e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 5610674ae819SStefano Zampini j = 0; 5611e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 561245a1bb75SStefano Zampini valid_constraints = k-j; 5613911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 5614984c4197SStefano Zampini #endif /* on missing GESVD */ 5615674ae819SStefano Zampini } 5616a773dcb8SStefano Zampini } 56179162d606SStefano Zampini /* update pointers information */ 56189162d606SStefano Zampini if (valid_constraints) { 56199162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 56209162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 56219162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 56229162d606SStefano Zampini /* set change_of_basis flag */ 562345a1bb75SStefano Zampini if (boolforchange) { 5624b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 56259162d606SStefano Zampini } 5626b3d85658SStefano Zampini total_counts_cc++; 562745a1bb75SStefano Zampini } 562845a1bb75SStefano Zampini } 5629984c4197SStefano Zampini /* free workspace */ 56308f1c130eSStefano Zampini if (!skip_lapack) { 5631984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 5632984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 5633984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 5634984c4197SStefano Zampini #endif 5635984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 5636984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 5637984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 5638984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 5639984c4197SStefano Zampini #endif 5640984c4197SStefano Zampini } 5641984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 5642984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 5643984c4197SStefano Zampini } 5644984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 5645cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 5646cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 5647cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 5648cf5a6209SStefano Zampini } 5649cf5a6209SStefano Zampini if (n_ISForFaces) { 5650cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 5651cf5a6209SStefano Zampini } 5652cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 5653cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 5654cf5a6209SStefano Zampini } 5655cf5a6209SStefano Zampini if (n_ISForEdges) { 5656cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 5657cf5a6209SStefano Zampini } 5658cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 565908122e43SStefano Zampini } else { 566008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 5661984c4197SStefano Zampini 566208122e43SStefano Zampini total_counts = 0; 566308122e43SStefano Zampini n_vertices = 0; 5664d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 5665d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 566608122e43SStefano Zampini } 566708122e43SStefano Zampini max_constraints = 0; 56689162d606SStefano Zampini total_counts_cc = 0; 566908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 567008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 56719162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 567208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 567308122e43SStefano Zampini } 56749162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 56759162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 56769162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 56779162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 567874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 56799162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 56809162d606SStefano Zampini total_counts_cc = 0; 56819162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 56829162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 56839162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 568408122e43SStefano Zampini } 568508122e43SStefano Zampini } 56869162d606SStefano Zampini #if 0 56879162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 56889162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 56899162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 56909162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 56919162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 56929162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 56939162d606SStefano Zampini } 56949162d606SStefano Zampini printf("\n"); 56959162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 56969162d606SStefano Zampini } 56971b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 56988bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 56991b968477SStefano Zampini } 57001b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 57018bec7fa6SStefano 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]); 57021b968477SStefano Zampini } 570308122e43SStefano Zampini #endif 570408122e43SStefano Zampini 57058bec7fa6SStefano Zampini max_size_of_constraint = 0; 57069162d606SStefano 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]); 57079162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 570808122e43SStefano Zampini /* Change of basis */ 5709b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 571008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 571108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 571208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 5713b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 571408122e43SStefano Zampini } 571508122e43SStefano Zampini } 571608122e43SStefano Zampini } 571708122e43SStefano Zampini } 5718984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 57194f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 572008122e43SStefano Zampini 57219162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 57229162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 5723eee23b56SStefano 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); 5724674ae819SStefano Zampini 5725674ae819SStefano Zampini /* Create constraint matrix */ 5726674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 572716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 5728984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 5729984c4197SStefano Zampini 5730984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 5731a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 5732a717540cSStefano Zampini qr_needed = PETSC_FALSE; 573374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 5734984c4197SStefano Zampini total_primal_vertices=0; 5735b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 57369162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 57379162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 573872b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 57399162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 5740b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 574164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 57429162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 57439162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 5744a717540cSStefano Zampini } 5745b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 574691af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 5747a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 5748a717540cSStefano Zampini qr_needed = PETSC_TRUE; 5749a717540cSStefano Zampini } 5750fa434743SStefano Zampini } else { 5751b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 5752fa434743SStefano Zampini } 5753a717540cSStefano Zampini } 5754b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 5755b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 5756674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 575770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 57584f1b2e48SStefano 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); 57590e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 57600e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 5761984c4197SStefano Zampini 5762984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 576374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 5764785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 5765984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 576674d5cdf7SStefano Zampini 5767984c4197SStefano Zampini j = total_primal_vertices; 576874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 5769b3d85658SStefano Zampini cum = total_primal_vertices; 57709162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 57714641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 5772b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 5773b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 5774b3d85658SStefano Zampini cum++; 57759162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 577674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 577774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 577874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 577974d5cdf7SStefano Zampini } 57809162d606SStefano Zampini j += constraints_n[i]; 5781674ae819SStefano Zampini } 5782674ae819SStefano Zampini } 5783674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 5784674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 5785088faed8SStefano Zampini 5786674ae819SStefano Zampini /* set values in constraint matrix */ 5787984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 57880e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5789674ae819SStefano Zampini } 5790984c4197SStefano Zampini total_counts = total_primal_vertices; 57919162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 57924641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 57939162d606SStefano Zampini PetscInt *cols; 57949162d606SStefano Zampini 57959162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 57969162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 57979162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 57989162d606SStefano Zampini PetscInt row = total_counts+k; 57999162d606SStefano Zampini PetscScalar *vals; 58009162d606SStefano Zampini 58019162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 58029162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 58039162d606SStefano Zampini } 58049162d606SStefano Zampini total_counts += constraints_n[i]; 5805674ae819SStefano Zampini } 5806674ae819SStefano Zampini } 5807674ae819SStefano Zampini /* assembling */ 5808674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5809674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5810088faed8SStefano Zampini 5811984c4197SStefano Zampini /* 58126a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5813984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 5814f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 5815984c4197SStefano Zampini */ 5816674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 5817674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 5818026de310SStefano Zampini /* dual and primal dofs on a single cc */ 5819984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 5820984c4197SStefano Zampini /* working stuff for GEQRF */ 582181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 5822984c4197SStefano Zampini PetscBLASInt lqr_work; 5823984c4197SStefano Zampini /* working stuff for UNGQR */ 5824984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 5825984c4197SStefano Zampini PetscBLASInt lgqr_work; 5826984c4197SStefano Zampini /* working stuff for TRTRS */ 5827984c4197SStefano Zampini PetscScalar *trs_rhs; 58283f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 5829984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 5830984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 5831984c4197SStefano Zampini PetscScalar *start_vals; 5832984c4197SStefano Zampini /* working stuff for values insertion */ 58334641a718SStefano Zampini PetscBT is_primal; 583464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 5835906d46d4SStefano Zampini /* matrix sizes */ 5836906d46d4SStefano Zampini PetscInt global_size,local_size; 5837906d46d4SStefano Zampini /* temporary change of basis */ 5838906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 5839cf5a6209SStefano Zampini /* extra space for debugging */ 5840cf5a6209SStefano Zampini PetscScalar *dbg_work; 5841984c4197SStefano Zampini 5842906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 5843906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 584416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 5845bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 5846906d46d4SStefano Zampini /* nonzeros for local mat */ 5847bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 58481dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5849bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 58501dd7afcfSStefano Zampini } else { 58511dd7afcfSStefano Zampini const PetscInt *ii; 58521dd7afcfSStefano Zampini PetscInt n; 58531dd7afcfSStefano Zampini PetscBool flg_row; 58541dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 58551dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 58561dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 58571dd7afcfSStefano Zampini } 58589162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 5859a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 58609162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 5861a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 58629162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 5863a717540cSStefano Zampini } else { 58649162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 58659162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 5866a717540cSStefano Zampini } 5867a717540cSStefano Zampini } 5868a717540cSStefano Zampini } 5869906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 5870bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 58711dd7afcfSStefano Zampini /* Set interior change in the matrix */ 58721dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 5873bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 5874906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 5875a717540cSStefano Zampini } 58761dd7afcfSStefano Zampini } else { 58771dd7afcfSStefano Zampini const PetscInt *ii,*jj; 58781dd7afcfSStefano Zampini PetscScalar *aa; 58791dd7afcfSStefano Zampini PetscInt n; 58801dd7afcfSStefano Zampini PetscBool flg_row; 58811dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 58821dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 58831dd7afcfSStefano Zampini for (i=0;i<n;i++) { 58841dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 58851dd7afcfSStefano Zampini } 58861dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 58871dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 58881dd7afcfSStefano Zampini } 5889a717540cSStefano Zampini 5890a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5891a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 5892a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5893a717540cSStefano Zampini } 5894a717540cSStefano Zampini 5895a717540cSStefano Zampini 5896a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 5897a717540cSStefano Zampini /* 5898a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 5899a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 5900a717540cSStefano Zampini 5901a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 5902a717540cSStefano Zampini 5903a6b551f4SStefano 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) 5904a6b551f4SStefano Zampini 5905a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 5906a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 5907a717540cSStefano Zampini | ... | 5908a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 5909a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 5910a717540cSStefano Zampini 5911a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 5912a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 5913a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 5914a6b551f4SStefano Zampini 5915a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 5916a717540cSStefano Zampini */ 5917a717540cSStefano Zampini if (qr_needed) { 5918984c4197SStefano Zampini /* space to store Q */ 5919854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 59204e64d54eSstefano_zampini /* array to store scaling factors for reflectors */ 59214e64d54eSstefano_zampini ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 5922984c4197SStefano Zampini /* first we issue queries for optimal work */ 59233f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 59243f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 59253f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5926984c4197SStefano Zampini lqr_work = -1; 59273f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 5928984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 5929984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 5930785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 5931984c4197SStefano Zampini lgqr_work = -1; 59323f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 59333f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 59343f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 59353f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 59363f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 59373f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 5938984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 5939984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 5940785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 5941984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 5942785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 5943a717540cSStefano Zampini /* allocating workspace for check */ 5944a717540cSStefano Zampini if (pcbddc->dbg_flag) { 5945cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 5946a717540cSStefano Zampini } 5947a717540cSStefano Zampini } 5948984c4197SStefano Zampini /* array to store whether a node is primal or not */ 59494641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 5950473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 59510e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 5952eee23b56SStefano 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); 595339e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 595439e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 595539e2fb2aSStefano Zampini } 595639e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 5957984c4197SStefano Zampini 5958a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 59599162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 59609162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 59614641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 5962984c4197SStefano Zampini /* get constraint info */ 59639162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 5964984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 5965984c4197SStefano Zampini 5966984c4197SStefano Zampini if (pcbddc->dbg_flag) { 59679162d606SStefano 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); 5968674ae819SStefano Zampini } 5969984c4197SStefano Zampini 5970fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 5971a717540cSStefano Zampini 5972a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 5973a717540cSStefano Zampini if (pcbddc->dbg_flag) { 59749162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5975a717540cSStefano Zampini } 5976984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 59779162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 5978984c4197SStefano Zampini 5979984c4197SStefano Zampini /* compute QR decomposition of constraints */ 59803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 59813f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 59823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 5983674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59843f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 5985984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 5986674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5987984c4197SStefano Zampini 5988984c4197SStefano Zampini /* explictly compute R^-T */ 5989984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 5990984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 59913f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 59923f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 59933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 59943f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 5995984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 59963f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 5997984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 5998984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 5999984c4197SStefano Zampini 6000a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 60013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 60023f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 60033f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 60043f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6005984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 60063f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 6007984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 6008984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6009984c4197SStefano Zampini 6010984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 6011984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 6012984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 60133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 60143f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 60153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 60163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 60173f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 60183f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 6019984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 60209162d606SStefano 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)); 6021984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 60229162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 6023984c4197SStefano Zampini 6024984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 60259162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 6026984c4197SStefano Zampini /* insert cols for primal dofs */ 6027984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 6028984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 60299162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6030906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6031984c4197SStefano Zampini } 6032984c4197SStefano Zampini /* insert cols for dual dofs */ 6033984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 60349162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 6035984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 60369162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6037906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 6038984c4197SStefano Zampini j++; 6039674ae819SStefano Zampini } 6040674ae819SStefano Zampini } 6041984c4197SStefano Zampini 6042984c4197SStefano Zampini /* check change of basis */ 6043984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6044984c4197SStefano Zampini PetscInt ii,jj; 6045984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 6046c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 6047c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 6048c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 6049c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 6050c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 6051c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 6052984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 6053cf5a6209SStefano 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)); 6054984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 6055984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6056984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6057cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 6058cf5a6209SStefano 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; 6059674ae819SStefano Zampini } 6060674ae819SStefano Zampini } 6061984c4197SStefano Zampini if (!valid_qr) { 606222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 6063984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 6064984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 6065cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 6066cf5a6209SStefano 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])); 6067674ae819SStefano Zampini } 6068cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 6069cf5a6209SStefano 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])); 6070984c4197SStefano Zampini } 6071984c4197SStefano Zampini } 6072984c4197SStefano Zampini } 6073674ae819SStefano Zampini } else { 607422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 6075674ae819SStefano Zampini } 6076674ae819SStefano Zampini } 6077a717540cSStefano Zampini } else { /* simple transformation block */ 6078a717540cSStefano Zampini PetscInt row,col; 6079a6b551f4SStefano Zampini PetscScalar val,norm; 6080a6b551f4SStefano Zampini 6081a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 60829162d606SStefano 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)); 6083a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 60849162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 60859162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 6086bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 60879162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 6088906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 60899162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 6090a717540cSStefano Zampini } else { 6091a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 60929162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 6093a717540cSStefano Zampini if (row != col) { 60949162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 6095a717540cSStefano Zampini } else { 60969162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 6097a717540cSStefano Zampini } 6098906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 6099a717540cSStefano Zampini } 6100a717540cSStefano Zampini } 6101a717540cSStefano Zampini } 610298a51de6SStefano Zampini if (pcbddc->dbg_flag) { 610322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 6104a717540cSStefano Zampini } 6105674ae819SStefano Zampini } 6106984c4197SStefano Zampini } else { 6107984c4197SStefano Zampini if (pcbddc->dbg_flag) { 61089162d606SStefano 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); 6109674ae819SStefano Zampini } 6110674ae819SStefano Zampini } 6111674ae819SStefano Zampini } 6112a717540cSStefano Zampini 6113a717540cSStefano Zampini /* free workspace */ 6114a717540cSStefano Zampini if (qr_needed) { 6115984c4197SStefano Zampini if (pcbddc->dbg_flag) { 6116cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 6117984c4197SStefano Zampini } 6118984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 6119984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 6120984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 6121984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 6122984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 6123674ae819SStefano Zampini } 6124a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 6125906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6126906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6127906d46d4SStefano Zampini 6128906d46d4SStefano Zampini /* assembling of global change of variable */ 612988c03ad3SStefano Zampini if (!pcbddc->fake_change) { 6130bbb9e6c6SStefano Zampini Mat tmat; 613116f15bc4SStefano Zampini PetscInt bs; 613216f15bc4SStefano Zampini 6133906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 6134906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 6135bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 6136bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 6137bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6138bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 613916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 614016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 6141906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 6142bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 6143bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 6144bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6145bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6146bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6147e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6148e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6149bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 6150bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 615188c03ad3SStefano Zampini 6152906d46d4SStefano Zampini /* check */ 6153906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 6154906d46d4SStefano Zampini PetscReal error; 6155906d46d4SStefano Zampini Vec x,x_change; 6156906d46d4SStefano Zampini 6157906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 6158906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 6159906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 6160906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 6161e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6162e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6163bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 6164e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6165e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6166906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 6167906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 6168906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 6169637e8532SStefano Zampini if (error > PETSC_SMALL) { 6170637e8532SStefano Zampini SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e\n",error); 6171637e8532SStefano Zampini } 6172906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 6173906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 6174906d46d4SStefano Zampini } 6175b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 6176b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 6177b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6178bf3a8328SStefano Zampini 61799a962809SStefano 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); 6180b334f244SStefano Zampini if (sub_schurs && sub_schurs->S_Ej_all) { 6181ac632422SStefano Zampini Mat S_new,tmat; 6182bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 6183bbb9e6c6SStefano Zampini 6184bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 61857dae84e0SHong Zhang ierr = MatCreateSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 6186bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6187bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 6188bf3a8328SStefano Zampini IS is_V; 6189b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 6190b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 6191b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 6192b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 6193b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 6194bf3a8328SStefano Zampini } 6195bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 6196ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6197b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 6198ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6199bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6200bf3a8328SStefano Zampini const PetscScalar *array; 6201bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 6202bf3a8328SStefano Zampini PetscInt i,n_V; 6203bf3a8328SStefano Zampini 6204b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6205b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 6206b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6207b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6208b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 6209b087196eSStefano Zampini for (i=0;i<n_V;i++) { 6210b087196eSStefano Zampini PetscScalar val; 6211b087196eSStefano Zampini PetscInt idx; 6212b087196eSStefano Zampini 6213b087196eSStefano Zampini idx = idxs_V[i]; 6214b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 6215b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 6216b087196eSStefano Zampini } 6217b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6218b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6219bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 6220bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 6221bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 6222bf3a8328SStefano Zampini } 6223ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 6224ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6225ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 6226ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 6227b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 6228ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 6229bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 6230b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 6231bf3a8328SStefano Zampini } 6232ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 6233ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 6234ac632422SStefano Zampini } 6235b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 623688c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 6237b96c3477SStefano Zampini } 6238c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 6239b334f244SStefano Zampini if (sub_schurs && sub_schurs->change) { 6240c9db6a07SStefano Zampini PetscInt i; 6241c9db6a07SStefano Zampini 6242c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6243c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 6244c9db6a07SStefano Zampini } 6245c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 6246c9db6a07SStefano Zampini } 6247b96c3477SStefano Zampini } 624816909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 624916909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 625016909a7fSStefano Zampini } else { 6251906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 625216909a7fSStefano Zampini } 62531dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 625427b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 625572b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 625672b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 625772b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 625872b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 625972b8c272SStefano Zampini } 62601dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 626127b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 6262b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 6263b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 6264906d46d4SStefano Zampini } else { 62651dd7afcfSStefano Zampini Mat benign_global = NULL; 626627b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 62671dd7afcfSStefano Zampini Mat tmat; 62681dd7afcfSStefano Zampini 62691dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 62701dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 62711dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 62721dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 62731dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 62741dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 62751dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 62761dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 62771dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 62781dd7afcfSStefano Zampini if (pcbddc->benign_change) { 62791dd7afcfSStefano Zampini Mat M; 62801dd7afcfSStefano Zampini 62811dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 62821dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 62831dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 62841dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 6285906d46d4SStefano Zampini } else { 62861dd7afcfSStefano Zampini Mat eye; 62871dd7afcfSStefano Zampini PetscScalar *array; 62881dd7afcfSStefano Zampini 62891dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 62901dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 62911dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 62921dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 6293906d46d4SStefano Zampini } 62941dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 62951dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 62961dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 62971dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 62981dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 62991dd7afcfSStefano Zampini } 63001dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 63011dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 63021dd7afcfSStefano Zampini } 63031dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 63041dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 63051dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 630627b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 63071dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 63081dd7afcfSStefano Zampini } 63091dd7afcfSStefano Zampini } 631016909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 631116909a7fSStefano Zampini IS is_global; 631216909a7fSStefano Zampini const PetscInt *gidxs; 631316909a7fSStefano Zampini 631416909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 631516909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 631616909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 63177dae84e0SHong Zhang ierr = MatCreateSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 631816909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 631916909a7fSStefano Zampini } 63201dd7afcfSStefano Zampini } 63211dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 63221dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 6323b9b85e73SStefano Zampini } 6324a717540cSStefano Zampini 632572b8c272SStefano Zampini if (!pcbddc->fake_change) { 63264f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 63274f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 63284f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 63294f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 6330019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 6331019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 6332019a44ceSStefano Zampini pcbddc->local_primal_size++; 6333019a44ceSStefano Zampini } 6334019a44ceSStefano Zampini 6335019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 6336727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 6337727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 63389f47a83aSStefano 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); 6339c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 63400e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 63419f47a83aSStefano 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); 6342727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 6343727cdba6SStefano Zampini } 63440e6343abSStefano Zampini } 6345727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 6346b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 634772b8c272SStefano Zampini } 634872b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 6349727cdba6SStefano Zampini 6350a717540cSStefano Zampini /* flush dbg viewer */ 6351b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 6352b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6353b8ffe317SStefano Zampini } 6354a717540cSStefano Zampini 6355e310c8b4SStefano Zampini /* free workspace */ 6356a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 63574641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 635808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 63599162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 63609162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 636108122e43SStefano Zampini } else { 63629162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 63639162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 63649162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 636508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 636608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 63679162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 63689162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 636908122e43SStefano Zampini } 6370674ae819SStefano Zampini PetscFunctionReturn(0); 6371674ae819SStefano Zampini } 6372674ae819SStefano Zampini 6373674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 6374674ae819SStefano Zampini { 637571582508SStefano Zampini ISLocalToGlobalMapping map; 6376674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 6377674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 637866da6bd7Sstefano_zampini PetscInt i,N; 637966da6bd7Sstefano_zampini PetscBool rcsr = PETSC_FALSE; 638066da6bd7Sstefano_zampini PetscErrorCode ierr; 6381674ae819SStefano Zampini 6382674ae819SStefano Zampini PetscFunctionBegin; 63838af8fcf9SStefano Zampini if (pcbddc->recompute_topography) { 6384b03ebc13SStefano Zampini pcbddc->graphanalyzed = PETSC_FALSE; 63858e61c736SStefano Zampini /* Reset previously computed graph */ 63868e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 6387674ae819SStefano Zampini /* Init local Graph struct */ 63887fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 638971582508SStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);CHKERRQ(ierr); 6390be12c134Sstefano_zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);CHKERRQ(ierr); 6391674ae819SStefano Zampini 63927a0e7b2cSstefano_zampini if (pcbddc->user_primal_vertices_local && !pcbddc->user_primal_vertices) { 63937a0e7b2cSstefano_zampini ierr = PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 63947a0e7b2cSstefano_zampini } 6395575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 63969a962809SStefano 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); 63979577ea80SStefano Zampini 6398674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 639966da6bd7Sstefano_zampini if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) { 64004d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 64014d379d7bSStefano Zampini PetscInt nvtxs; 6402e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 6403674ae819SStefano Zampini 64042fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 64052fffb893SStefano Zampini if (flg_row) { 64064d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 6407b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 64082fffb893SStefano Zampini } 64092fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 641066da6bd7Sstefano_zampini rcsr = PETSC_TRUE; 6411674ae819SStefano Zampini } 64129b28b941SStefano Zampini if (pcbddc->dbg_flag) { 64139b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6414674ae819SStefano Zampini } 6415674ae819SStefano Zampini 6416674ae819SStefano Zampini /* Setup of Graph */ 64174b2aedd3SStefano Zampini pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */ 641814f95afaSStefano 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); 6419674ae819SStefano Zampini 64204f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 64214f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 64224f1b2e48SStefano Zampini PetscInt *local_subs; 64234f1b2e48SStefano Zampini 64244f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 64254f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 64264f1b2e48SStefano Zampini const PetscInt *idxs; 64274f1b2e48SStefano Zampini PetscInt nl,j; 64284f1b2e48SStefano Zampini 64294f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 64304f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 643171582508SStefano Zampini for (j=0;j<nl;j++) local_subs[idxs[j]] = i; 64324f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 64334f1b2e48SStefano Zampini } 64344f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 64354f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 64364f1b2e48SStefano Zampini } 64378af8fcf9SStefano Zampini } 64384f1b2e48SStefano Zampini 6439cac5312eSStefano Zampini if (!pcbddc->graphanalyzed) { 6440674ae819SStefano Zampini /* Graph's connected components analysis */ 6441674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 644271582508SStefano Zampini pcbddc->graphanalyzed = PETSC_TRUE; 64438af8fcf9SStefano Zampini } 644466da6bd7Sstefano_zampini if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0; 6445674ae819SStefano Zampini PetscFunctionReturn(0); 6446674ae819SStefano Zampini } 6447674ae819SStefano Zampini 64489a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 64499a7d3425SStefano Zampini { 64509a7d3425SStefano Zampini PetscInt i,j; 64519a7d3425SStefano Zampini PetscScalar *alphas; 64529a7d3425SStefano Zampini PetscErrorCode ierr; 64539a7d3425SStefano Zampini 64549a7d3425SStefano Zampini PetscFunctionBegin; 6455785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 64569a7d3425SStefano Zampini for (i=0;i<n;i++) { 64579a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 6458669cc0f4SStefano Zampini ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],alphas);CHKERRQ(ierr); 6459669cc0f4SStefano Zampini for (j=0;j<n-i-1;j++) alphas[j] = PetscConj(-alphas[j]); 6460669cc0f4SStefano Zampini ierr = VecMAXPY(vecs[j],n-i-1,alphas,vecs+i);CHKERRQ(ierr); 64619a7d3425SStefano Zampini } 64629a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 64639a7d3425SStefano Zampini PetscFunctionReturn(0); 64649a7d3425SStefano Zampini } 64659a7d3425SStefano Zampini 6466bb360cb4SStefano Zampini PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 6467e7931f94SStefano Zampini { 646857de7509SStefano Zampini Mat A; 6469e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 6470e7931f94SStefano Zampini PetscMPIInt size,rank,color; 647152e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 647252e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 6473bb360cb4SStefano Zampini PetscInt im_active,active_procs,N,n,i,j,threshold = 2; 647457de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 647527b6a85dSStefano Zampini PetscInt xadj_count,*count; 647627b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 647727b6a85dSStefano Zampini PetscSubcomm psubcomm; 647827b6a85dSStefano Zampini MPI_Comm subcomm; 647952e5ac9dSStefano Zampini PetscErrorCode ierr; 6480a57a6d2fSStefano Zampini 6481e7931f94SStefano Zampini PetscFunctionBegin; 648257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 648357de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 6484fbfcfee5SBarry 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); 648557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 648657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 648757de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 648857de7509SStefano Zampini 648957de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 649057de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 649157de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 649257de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 649357de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 6494bb360cb4SStefano Zampini im_active = !!n; 649557de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 649657de7509SStefano Zampini void_procs = size - active_procs; 649757de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 649857de7509SStefano Zampini if (void_procs) { 649957de7509SStefano Zampini PetscInt ncand; 650057de7509SStefano Zampini 650157de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 650257de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 650357de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 650457de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 650557de7509SStefano Zampini if (!procs_candidates[i]) { 650657de7509SStefano Zampini procs_candidates[ncand++] = i; 650757de7509SStefano Zampini } 650857de7509SStefano Zampini } 650957de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 651057de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 651157de7509SStefano Zampini } 651257de7509SStefano Zampini 6513bb360cb4SStefano Zampini /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix 651414f0bfb9SStefano Zampini number of subdomains requested 1 -> send to master or first candidate in voids */ 6515bb360cb4SStefano Zampini ierr = MatGetSize(mat,&N,NULL);CHKERRQ(ierr); 6516bb360cb4SStefano Zampini if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) { 651714f0bfb9SStefano Zampini PetscInt issize,isidx,dest; 651814f0bfb9SStefano Zampini if (*n_subdomains == 1) dest = 0; 651914f0bfb9SStefano Zampini else dest = rank; 652057de7509SStefano Zampini if (im_active) { 652157de7509SStefano Zampini issize = 1; 652257de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 652314f0bfb9SStefano Zampini isidx = procs_candidates[dest]; 652457de7509SStefano Zampini } else { 652514f0bfb9SStefano Zampini isidx = dest; 652657de7509SStefano Zampini } 652757de7509SStefano Zampini } else { 652857de7509SStefano Zampini issize = 0; 652957de7509SStefano Zampini isidx = -1; 653057de7509SStefano Zampini } 6531bb360cb4SStefano Zampini if (*n_subdomains != 1) *n_subdomains = active_procs; 653257de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 6533daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 653457de7509SStefano Zampini PetscFunctionReturn(0); 653557de7509SStefano Zampini } 6536c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 6537c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 653827b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 6539e7931f94SStefano Zampini 6540e7931f94SStefano Zampini /* Get info on mapping */ 65413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6542e7931f94SStefano Zampini 6543e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 6544785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 6545e7931f94SStefano Zampini xadj[0] = 0; 6546e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 6547785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 6548785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 6549bb360cb4SStefano Zampini ierr = PetscCalloc1(n,&count);CHKERRQ(ierr); 655027b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 655127b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 655227b6a85dSStefano Zampini count[shared[i][j]] += 1; 6553e7931f94SStefano Zampini 655427b6a85dSStefano Zampini xadj_count = 0; 65552b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 655627b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 655727b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 6558d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 6559d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 6560d023bfaeSStefano Zampini xadj_count++; 656127b6a85dSStefano Zampini break; 656227b6a85dSStefano Zampini } 6563e7931f94SStefano Zampini } 6564e7931f94SStefano Zampini } 6565d023bfaeSStefano Zampini xadj[1] = xadj_count; 656627b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 65673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 6568e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 6569e7931f94SStefano Zampini 65703837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 6571e7931f94SStefano Zampini 657227b6a85dSStefano Zampini /* Restrict work on active processes only */ 657327b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 657427b6a85dSStefano Zampini if (void_procs) { 657527b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 657627b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 657727b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 657827b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 657927b6a85dSStefano Zampini } else { 658027b6a85dSStefano Zampini psubcomm = NULL; 658127b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 658227b6a85dSStefano Zampini } 658327b6a85dSStefano Zampini 658427b6a85dSStefano Zampini v_wgt = NULL; 658527b6a85dSStefano Zampini if (!color) { 6586e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 6587e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 6588e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6589c8587f34SStefano Zampini } else { 659052e5ac9dSStefano Zampini Mat subdomain_adj; 659152e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 659252e5ac9dSStefano Zampini MatPartitioning partitioner; 659327b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 659452e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 659557de7509SStefano Zampini PetscMPIInt size; 6596b0c7d250SStefano Zampini PetscBool aggregate; 6597b0c7d250SStefano Zampini 659827b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 659927b6a85dSStefano Zampini if (void_procs) { 660027b6a85dSStefano Zampini PetscInt prank = rank; 6601785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 660227b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 6603e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 6604e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 6605c8587f34SStefano Zampini } 6606e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 660727b6a85dSStefano Zampini } else { 660827b6a85dSStefano Zampini oldranks = NULL; 660927b6a85dSStefano Zampini } 6610b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 661127b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 6612b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 6613b0c7d250SStefano Zampini PetscMPIInt nrank; 6614b0c7d250SStefano Zampini PetscScalar *vals; 6615b0c7d250SStefano Zampini 661627b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 6617b0c7d250SStefano Zampini lrows = 0; 6618b0c7d250SStefano Zampini if (nrank<redprocs) { 6619b0c7d250SStefano Zampini lrows = size/redprocs; 6620b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 6621b0c7d250SStefano Zampini } 662227b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 6623b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 6624b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6625b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 6626b0c7d250SStefano Zampini row = nrank; 6627b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 6628b0c7d250SStefano Zampini cols = adjncy; 6629b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 6630b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 6631b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 6632b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 6633b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 663452e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 663552e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 663652e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 6637b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 663827b6a85dSStefano Zampini if (use_vwgt) { 663927b6a85dSStefano Zampini Vec v; 664027b6a85dSStefano Zampini const PetscScalar *array; 664127b6a85dSStefano Zampini PetscInt nl; 664227b6a85dSStefano Zampini 664327b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 6644bb360cb4SStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)n,INSERT_VALUES);CHKERRQ(ierr); 664527b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 664627b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 664727b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 664827b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 664927b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 665022db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 665127b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 665227b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 665327b6a85dSStefano Zampini } 6654b0c7d250SStefano Zampini } else { 665527b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 665627b6a85dSStefano Zampini if (use_vwgt) { 665727b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 6658bb360cb4SStefano Zampini v_wgt[0] = n; 665927b6a85dSStefano Zampini } 6660b0c7d250SStefano Zampini } 666122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 6662e7931f94SStefano Zampini 6663e7931f94SStefano Zampini /* Partition */ 666427b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 6665e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 666627b6a85dSStefano Zampini if (v_wgt) { 6667e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 6668c8587f34SStefano Zampini } 666957de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 667057de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 6671e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 6672e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 667322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 6674e7931f94SStefano Zampini 667552e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 66766583bcc1SStefano Zampini ierr = ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 667752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 667852e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 667957de7509SStefano Zampini if (!aggregate) { 668057de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 668127b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 668227b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 668327b6a85dSStefano Zampini #endif 668457de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 668527b6a85dSStefano Zampini } else if (oldranks) { 6686b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 668727b6a85dSStefano Zampini } else { 668827b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 668957de7509SStefano Zampini } 669028143c3dSStefano Zampini } else { 6691b0c7d250SStefano Zampini PetscInt idxs[1]; 6692b0c7d250SStefano Zampini PetscMPIInt tag; 6693b0c7d250SStefano Zampini MPI_Request *reqs; 6694b0c7d250SStefano Zampini 6695b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 6696b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 6697b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 669827b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 669928143c3dSStefano Zampini } 670027b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 6701b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6702b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 670357de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 670427b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 670527b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 670627b6a85dSStefano Zampini #endif 670757de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 670827b6a85dSStefano Zampini } else if (oldranks) { 6709b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 671027b6a85dSStefano Zampini } else { 671127b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 6712e7931f94SStefano Zampini } 671357de7509SStefano Zampini } 671452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 6715e7931f94SStefano Zampini /* clean up */ 6716e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 671752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 6718e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 6719e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 6720e7931f94SStefano Zampini } 672127b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 672257de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 6723e7931f94SStefano Zampini 6724e7931f94SStefano Zampini /* assemble parallel IS for sends */ 6725e7931f94SStefano Zampini i = 1; 672627b6a85dSStefano Zampini if (!color) i=0; 672757de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 6728e7931f94SStefano Zampini PetscFunctionReturn(0); 6729e7931f94SStefano Zampini } 6730e7931f94SStefano Zampini 6731e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 6732e7931f94SStefano Zampini 67331e0482f5SStefano 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[]) 6734e7931f94SStefano Zampini { 673570cf5478SStefano Zampini Mat local_mat; 6736e7931f94SStefano Zampini IS is_sends_internal; 67379d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 67381ae86dd6SStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs; 67399d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 6740e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 6741e7931f94SStefano Zampini PetscInt* l2gmap_indices; 6742e7931f94SStefano Zampini const PetscInt* is_indices; 6743e7931f94SStefano Zampini MatType new_local_type; 6744e7931f94SStefano Zampini /* buffers */ 6745e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 674628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 67479d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 6748e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 67491ae86dd6SStefano Zampini PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs; 6750e7931f94SStefano Zampini /* MPI */ 675128143c3dSStefano Zampini MPI_Comm comm,comm_n; 675228143c3dSStefano Zampini PetscSubcomm subcomm; 6753e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 675428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 675528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 67561ae86dd6SStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest; 67571ae86dd6SStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs; 67581ae86dd6SStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs; 6759e7931f94SStefano Zampini PetscErrorCode ierr; 6760e7931f94SStefano Zampini 6761e7931f94SStefano Zampini PetscFunctionBegin; 676257de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 6763e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 6764fbfcfee5SBarry Smith if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME); 676557de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 676657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 676757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 676857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 676957de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 67701ae86dd6SStefano Zampini PetscValidLogicalCollectiveInt(mat,nvecs,10); 67711ae86dd6SStefano Zampini if (nvecs) { 67721ae86dd6SStefano Zampini if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported"); 67731ae86dd6SStefano Zampini PetscValidHeaderSpecific(nnsp_vec[0],VEC_CLASSID,11); 67741ae86dd6SStefano Zampini } 677557de7509SStefano Zampini /* further checks */ 6776e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 6777e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 6778e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 6779e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 6780e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 678157de7509SStefano Zampini if (reuse && *mat_n) { 678270cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 678357de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 678470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 678528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 678670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 678770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 678870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 678970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 679070cf5478SStefano Zampini } 6791e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 6792e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 679357de7509SStefano Zampini 6794e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 6795e7931f94SStefano Zampini if (!is_sends) { 679628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 6797bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 6798c8587f34SStefano Zampini } else { 6799e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 6800e7931f94SStefano Zampini is_sends_internal = is_sends; 6801c8587f34SStefano Zampini } 6802e7931f94SStefano Zampini 6803e7931f94SStefano Zampini /* get comm */ 6804a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 6805e7931f94SStefano Zampini 6806e7931f94SStefano Zampini /* compute number of sends */ 6807e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 6808e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 6809e7931f94SStefano Zampini 6810e7931f94SStefano Zampini /* compute number of receives */ 6811e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 6812785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 6813e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 6814e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6815e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 6816e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 6817e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 6818e7931f94SStefano Zampini 681928143c3dSStefano Zampini /* restrict comm if requested */ 682028143c3dSStefano Zampini subcomm = 0; 682128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 682228143c3dSStefano Zampini if (restrict_comm) { 6823779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 6824779c1cceSStefano Zampini 682528143c3dSStefano Zampini color = 0; 682653a05cb3SStefano Zampini if (restrict_full) { 682753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 682853a05cb3SStefano Zampini } else { 682953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 683053a05cb3SStefano Zampini } 6831b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 683228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 683328143c3dSStefano Zampini /* check if reuse has been requested */ 683457de7509SStefano Zampini if (reuse) { 683528143c3dSStefano Zampini if (*mat_n) { 683628143c3dSStefano Zampini PetscMPIInt subcommsize2; 683728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 683828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 683928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 684028143c3dSStefano Zampini } else { 684128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 684228143c3dSStefano Zampini } 684328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 6844779c1cceSStefano Zampini PetscMPIInt rank; 6845779c1cceSStefano Zampini 6846779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 684728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 684828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 684928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 6850306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 685128143c3dSStefano Zampini } 685228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 685328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 685428143c3dSStefano Zampini } else { 685528143c3dSStefano Zampini comm_n = comm; 685628143c3dSStefano Zampini } 685728143c3dSStefano Zampini 6858e7931f94SStefano Zampini /* prepare send/receive buffers */ 6859785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 6860e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 6861785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 6862e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 686328143c3dSStefano Zampini if (nis) { 6864854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 686528143c3dSStefano Zampini } 6866e7931f94SStefano Zampini 686728143c3dSStefano Zampini /* Get data from local matrices */ 68686c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 6869e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 6870e7931f94SStefano Zampini /* 6871e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 6872e7931f94SStefano Zampini send_buffer_idxs should contain: 6873e7931f94SStefano Zampini - MatType_PRIVATE type 6874e7931f94SStefano Zampini - PetscInt size_of_l2gmap 6875e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 6876e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 6877e7931f94SStefano Zampini */ 68786c4ed002SBarry Smith else { 6879e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 68803bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 6881854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 6882e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 6883e7931f94SStefano Zampini send_buffer_idxs[1] = i; 68843bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6885e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 68863bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 6887e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 6888e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6889e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 6890e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 6891c8587f34SStefano Zampini } 6892c8587f34SStefano Zampini } 6893e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 689428143c3dSStefano Zampini /* additional is (if any) */ 689528143c3dSStefano Zampini if (nis) { 689628143c3dSStefano Zampini PetscMPIInt psum; 689728143c3dSStefano Zampini PetscInt j; 689828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 689928143c3dSStefano Zampini PetscInt plen; 690028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 690128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 690228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 690328143c3dSStefano Zampini } 6904854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 690528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 690628143c3dSStefano Zampini PetscInt plen; 690728143c3dSStefano Zampini const PetscInt *is_array_idxs; 690828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 690928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 691028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 691128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 691228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 691328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 691428143c3dSStefano Zampini } 691528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 691628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 691728143c3dSStefano Zampini } 691828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 691928143c3dSStefano Zampini } 69203b3b1effSJed Brown ierr = MatISRestoreLocalMat(mat,&local_mat);CHKERRQ(ierr); 692128143c3dSStefano Zampini 6922e7931f94SStefano Zampini buf_size_idxs = 0; 6923e7931f94SStefano Zampini buf_size_vals = 0; 692428143c3dSStefano Zampini buf_size_idxs_is = 0; 69251ae86dd6SStefano Zampini buf_size_vecs = 0; 6926e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6927e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 6928e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 692928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 69301ae86dd6SStefano Zampini if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i]; 6931e7931f94SStefano Zampini } 6932785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 6933785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 693495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 69351ae86dd6SStefano Zampini ierr = PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);CHKERRQ(ierr); 6936e7931f94SStefano Zampini 6937e7931f94SStefano Zampini /* get new tags for clean communications */ 6938e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 6939e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 694028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 69411ae86dd6SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);CHKERRQ(ierr); 6942e7931f94SStefano Zampini 6943e7931f94SStefano Zampini /* allocate for requests */ 6944785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 6945785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 694695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 69471ae86dd6SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_vecs);CHKERRQ(ierr); 6948785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 6949785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 695095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 69511ae86dd6SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_vecs);CHKERRQ(ierr); 6952e7931f94SStefano Zampini 6953e7931f94SStefano Zampini /* communications */ 6954e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 6955e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 695628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 69571ae86dd6SStefano Zampini ptr_vecs = recv_buffer_vecs; 6958e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 6959e7931f94SStefano Zampini source_dest = onodes[i]; 6960e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 6961e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 6962e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6963e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 696428143c3dSStefano Zampini if (nis) { 696557de7509SStefano Zampini source_dest = onodes_is[i]; 696628143c3dSStefano 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); 696728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 696828143c3dSStefano Zampini } 69691ae86dd6SStefano Zampini if (nvecs) { 69701ae86dd6SStefano Zampini source_dest = onodes[i]; 69711ae86dd6SStefano Zampini ierr = MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);CHKERRQ(ierr); 69721ae86dd6SStefano Zampini ptr_vecs += olengths_idxs[i]-2; 69731ae86dd6SStefano Zampini } 6974e7931f94SStefano Zampini } 6975e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 6976e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 6977e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 6978e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 697928143c3dSStefano Zampini if (nis) { 698028143c3dSStefano 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); 698128143c3dSStefano Zampini } 69821ae86dd6SStefano Zampini if (nvecs) { 69831ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 69841ae86dd6SStefano 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); 69851ae86dd6SStefano Zampini } 6986e7931f94SStefano Zampini } 6987e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 6988e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 6989e7931f94SStefano Zampini 6990e7931f94SStefano Zampini /* assemble new l2g map */ 6991e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 6992e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 69939d30be91SStefano Zampini new_local_rows = 0; 6994e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 69959d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 6996e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 6997e7931f94SStefano Zampini } 69989d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 6999e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 70009d30be91SStefano Zampini new_local_rows = 0; 7001e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 70029d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 70039d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 7004e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7005e7931f94SStefano Zampini } 70069d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 70079d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 7008e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 7009e7931f94SStefano Zampini 7010e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 7011e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 7012e7931f94SStefano 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) */ 7013e7931f94SStefano Zampini if (n_recvs) { 701428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 7015e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 7016e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7017e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 7018e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 7019e7931f94SStefano Zampini break; 7020e7931f94SStefano Zampini } 7021e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7022e7931f94SStefano Zampini } 7023e7931f94SStefano Zampini switch (new_local_type_private) { 702428143c3dSStefano Zampini case MATDENSE_PRIVATE: 7025e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7026e7931f94SStefano Zampini bs = 1; 7027e7931f94SStefano Zampini break; 7028e7931f94SStefano Zampini case MATAIJ_PRIVATE: 7029e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 7030e7931f94SStefano Zampini bs = 1; 7031e7931f94SStefano Zampini break; 7032e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 7033e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 7034e7931f94SStefano Zampini break; 7035e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 7036e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 7037e7931f94SStefano Zampini break; 7038e7931f94SStefano Zampini default: 7039fbfcfee5SBarry Smith SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,PETSC_FUNCTION_NAME); 7040e7931f94SStefano Zampini break; 7041e7931f94SStefano Zampini } 7042ed8ed4edSstefano_zampini } else { /* by default, new_local_type is seqaij */ 7043ed8ed4edSstefano_zampini new_local_type = MATSEQAIJ; 704428143c3dSStefano Zampini bs = 1; 7045e7931f94SStefano Zampini } 7046e7931f94SStefano Zampini 704770cf5478SStefano Zampini /* create MATIS object if needed */ 704857de7509SStefano Zampini if (!reuse) { 7049e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 7050e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 705170cf5478SStefano Zampini } else { 705270cf5478SStefano Zampini /* it also destroys the local matrices */ 705357de7509SStefano Zampini if (*mat_n) { 705470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 705557de7509SStefano Zampini } else { /* this is a fake object */ 705657de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 705757de7509SStefano Zampini } 705870cf5478SStefano Zampini } 705970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 7060e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 70619d30be91SStefano Zampini 70629d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 70639d30be91SStefano Zampini 70649d30be91SStefano Zampini /* Global to local map of received indices */ 70659d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 70669d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 70679d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 70689d30be91SStefano Zampini 70699d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 70709d30be91SStefano Zampini buf_size_idxs = 0; 70719d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 70729d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 70739d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 70749d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 70759d30be91SStefano Zampini } 70769d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 70779d30be91SStefano Zampini 70789d30be91SStefano Zampini /* set preallocation */ 70799d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 70809d30be91SStefano Zampini if (!newisdense) { 70819d30be91SStefano Zampini PetscInt *new_local_nnz=0; 70829d30be91SStefano Zampini 70839d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 70849d30be91SStefano Zampini if (n_recvs) { 70859d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 70869d30be91SStefano Zampini } 70879d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 70889d30be91SStefano Zampini PetscInt j; 70899d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 70909d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 70919d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 70929d30be91SStefano Zampini } 70939d30be91SStefano Zampini } else { 70949d30be91SStefano Zampini /* TODO */ 70959d30be91SStefano Zampini } 70969d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 70979d30be91SStefano Zampini } 70989d30be91SStefano Zampini if (new_local_nnz) { 70999d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 71009d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 71019d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 71029d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 71039d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 71049d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 71059d30be91SStefano Zampini } else { 71069d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 71079d30be91SStefano Zampini } 71089d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 71099d30be91SStefano Zampini } else { 71109d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 71119d30be91SStefano Zampini } 7112e7931f94SStefano Zampini 7113e7931f94SStefano Zampini /* set values */ 7114e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 71159d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 7116e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 7117e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 7118e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 71199d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 7120e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7121e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 7122e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 712328143c3dSStefano Zampini } else { 712428143c3dSStefano Zampini /* TODO */ 7125e7931f94SStefano Zampini } 7126e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 7127e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 7128e7931f94SStefano Zampini } 7129e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7130e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 71313b3b1effSJed Brown ierr = MatISRestoreLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 713270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 713370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 71349d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 7135e7931f94SStefano Zampini 7136dfd14d43SStefano Zampini #if 0 713728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 7138e7931f94SStefano Zampini Vec lvec,rvec; 7139e7931f94SStefano Zampini PetscReal infty_error; 7140e7931f94SStefano Zampini 71412a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 7142e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 7143e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 7144e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 714570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 7146e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 7147e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 7148e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 7149e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 7150e7931f94SStefano Zampini } 715128143c3dSStefano Zampini #endif 7152e7931f94SStefano Zampini 715328143c3dSStefano Zampini /* assemble new additional is (if any) */ 715428143c3dSStefano Zampini if (nis) { 715528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 715628143c3dSStefano Zampini 715728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7158854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 715928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 716028143c3dSStefano Zampini psum = 0; 716128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 716228143c3dSStefano Zampini for (j=0;j<nis;j++) { 716328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 716428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 716528143c3dSStefano Zampini psum += plen; 716628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 716728143c3dSStefano Zampini } 716828143c3dSStefano Zampini } 7169854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 7170854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 717128143c3dSStefano Zampini for (i=1;i<nis;i++) { 717228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 717328143c3dSStefano Zampini } 717428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 717528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 717628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 717728143c3dSStefano Zampini for (j=0;j<nis;j++) { 717828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 717928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 718028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 718128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 718228143c3dSStefano Zampini } 718328143c3dSStefano Zampini } 718428143c3dSStefano Zampini for (i=0;i<nis;i++) { 718528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 718628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 718728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 718828143c3dSStefano Zampini } 718928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 719028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 719128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 719228143c3dSStefano Zampini } 7193e7931f94SStefano Zampini /* free workspace */ 719428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 7195e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7196e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 7197e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 7198e7931f94SStefano Zampini if (isdense) { 7199e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 7200e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 72013b3b1effSJed Brown ierr = MatISRestoreLocalMat(mat,&local_mat);CHKERRQ(ierr); 7202e7931f94SStefano Zampini } else { 7203e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 7204e7931f94SStefano Zampini } 720528143c3dSStefano Zampini if (nis) { 720628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 720728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 720828143c3dSStefano Zampini } 72091ae86dd6SStefano Zampini 72101ae86dd6SStefano Zampini if (nvecs) { 72111ae86dd6SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 72121ae86dd6SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 72131ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72141ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 72151ae86dd6SStefano Zampini ierr = VecCreate(comm_n,&nnsp_vec[0]);CHKERRQ(ierr); 72161ae86dd6SStefano Zampini ierr = VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);CHKERRQ(ierr); 72171ae86dd6SStefano Zampini ierr = VecSetType(nnsp_vec[0],VECSTANDARD);CHKERRQ(ierr); 72181ae86dd6SStefano Zampini /* set values */ 72191ae86dd6SStefano Zampini ptr_vals = recv_buffer_vecs; 72201ae86dd6SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 72211ae86dd6SStefano Zampini ierr = VecGetArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72221ae86dd6SStefano Zampini for (i=0;i<n_recvs;i++) { 72231ae86dd6SStefano Zampini PetscInt j; 72241ae86dd6SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 72251ae86dd6SStefano Zampini send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j); 72261ae86dd6SStefano Zampini } 72271ae86dd6SStefano Zampini ptr_idxs += olengths_idxs[i]; 72281ae86dd6SStefano Zampini ptr_vals += olengths_idxs[i]-2; 72291ae86dd6SStefano Zampini } 72301ae86dd6SStefano Zampini ierr = VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);CHKERRQ(ierr); 72311ae86dd6SStefano Zampini ierr = VecAssemblyBegin(nnsp_vec[0]);CHKERRQ(ierr); 72321ae86dd6SStefano Zampini ierr = VecAssemblyEnd(nnsp_vec[0]);CHKERRQ(ierr); 72331ae86dd6SStefano Zampini } 72341ae86dd6SStefano Zampini 72351ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_vecs);CHKERRQ(ierr); 72361ae86dd6SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 7237e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 7238e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 72391ae86dd6SStefano Zampini ierr = PetscFree(recv_req_vecs);CHKERRQ(ierr); 724028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 7241e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 7242e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 72431ae86dd6SStefano Zampini ierr = PetscFree(send_req_vecs);CHKERRQ(ierr); 724428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 7245e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 7246e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 7247e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 7248e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 7249e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 725028143c3dSStefano Zampini if (nis) { 725128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 725228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 725328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 725428143c3dSStefano Zampini } 725528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 725628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 725728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 725828143c3dSStefano Zampini for (i=0;i<nis;i++) { 725928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 726028143c3dSStefano Zampini } 72611ae86dd6SStefano Zampini if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */ 72621ae86dd6SStefano Zampini ierr = VecDestroy(&nnsp_vec[0]);CHKERRQ(ierr); 72631ae86dd6SStefano Zampini } 726453a05cb3SStefano Zampini *mat_n = NULL; 726528143c3dSStefano Zampini } 7266e7931f94SStefano Zampini PetscFunctionReturn(0); 7267e7931f94SStefano Zampini } 7268a57a6d2fSStefano Zampini 726912edc857SStefano Zampini /* temporary hack into ksp private data structure */ 7270af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 727112edc857SStefano Zampini 7272c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 7273c8587f34SStefano Zampini { 7274c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 7275c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 727620a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 72771ae86dd6SStefano Zampini Mat coarsedivudotp = NULL; 72781e0482f5SStefano Zampini Mat coarseG,t_coarse_mat_is; 72799881197aSStefano Zampini MatNullSpace CoarseNullSpace = NULL; 728020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 72816e683305SStefano Zampini IS coarse_is,*isarray; 72826e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 728330368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 7284f9eb5b7dSStefano Zampini PC pc_temp; 7285c8587f34SStefano Zampini PCType coarse_pc_type; 7286c8587f34SStefano Zampini KSPType coarse_ksp_type; 7287f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 72884f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 72891e0482f5SStefano Zampini PetscInt ncoarse,nedcfield; 729068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 729122bc73bbSStefano Zampini PetscScalar *array; 729257de7509SStefano Zampini MatReuse coarse_mat_reuse; 729357de7509SStefano Zampini PetscBool restr, full_restr, have_void; 72947de4f681Sstefano_zampini PetscMPIInt commsize; 72959881197aSStefano Zampini PetscErrorCode ierr; 7296fdc09c96SStefano Zampini 7297c8587f34SStefano Zampini PetscFunctionBegin; 7298c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 729968457ee5SStefano 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 */ 7300fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 73015a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 73027de4f681Sstefano_zampini 73037de4f681Sstefano_zampini pcbddc->new_primal_space = PETSC_TRUE; 7304fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 7305f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 7306f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 7307f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 7308fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 730951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 731051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 7311727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 7312fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7313fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 7314fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7315f4ddd8eeSStefano Zampini } 7316fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 7317fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 7318f4ddd8eeSStefano Zampini } 731970cf5478SStefano Zampini /* reset any subassembling information */ 732057de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 732170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 732257de7509SStefano Zampini } 73236e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 7324fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 7325f4ddd8eeSStefano Zampini } 732657de7509SStefano Zampini /* assemble coarse matrix */ 732757de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 732857de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 732957de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 733057de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 733118a45a71SStefano Zampini } else { 733257de7509SStefano Zampini coarse_mat = NULL; 733357de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 73346e683305SStefano Zampini } 7335e7931f94SStefano Zampini 7336abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 7337abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 7338abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 7339abbbba34SStefano Zampini 7340abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 734122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 734222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 734322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 734422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 7345e176bc59SStefano 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); 73466e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 73476e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 73486e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 7349abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 7350abbbba34SStefano Zampini 735157de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 735257de7509SStefano Zampini im_active = !!(pcis->n); 735357de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 735457de7509SStefano Zampini 735514f0bfb9SStefano Zampini /* determine number of processes partecipating to coarse solver and compute subassembling pattern */ 735657de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 735757de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 73587de4f681Sstefano_zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&commsize);CHKERRQ(ierr); 735957de7509SStefano Zampini coarse_mat_is = NULL; 736057de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 736157de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 73621ae86dd6SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc); 736357de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 736457de7509SStefano Zampini if (multilevel_requested) { 736557de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 736657de7509SStefano Zampini restr = PETSC_FALSE; 736757de7509SStefano Zampini full_restr = PETSC_FALSE; 736857de7509SStefano Zampini } else { 736957de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 737057de7509SStefano Zampini restr = PETSC_TRUE; 737157de7509SStefano Zampini full_restr = PETSC_TRUE; 737257de7509SStefano Zampini } 73737de4f681Sstefano_zampini if (!pcbddc->coarse_size || commsize == 1) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE; 737457de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 737557de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 7376a198735bSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 7377bb360cb4SStefano Zampini if (multilevel_requested) { 7378bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7379bb360cb4SStefano Zampini } else { 7380bb360cb4SStefano Zampini ierr = PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 7381bb360cb4SStefano Zampini } 7382a198735bSStefano Zampini } else { 73837de4f681Sstefano_zampini PetscMPIInt rank; 7384a198735bSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 73857de4f681Sstefano_zampini have_void = (active_procs == (PetscInt)commsize) ? PETSC_FALSE : PETSC_TRUE; 7386a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 7387a198735bSStefano Zampini } 738857de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 738957de7509SStefano Zampini PetscInt psum; 739057de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 739157de7509SStefano Zampini else psum = 0; 739257de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 73937de4f681Sstefano_zampini if (ncoarse < commsize) have_void = PETSC_TRUE; 739457de7509SStefano Zampini } 739557de7509SStefano Zampini /* determine if we can go multilevel */ 739657de7509SStefano Zampini if (multilevel_requested) { 739757de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 739857de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 739957de7509SStefano Zampini } 740057de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 740157de7509SStefano Zampini 7402e4d548c7SStefano Zampini /* dump subassembling pattern */ 7403e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 7404e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 7405e4d548c7SStefano Zampini } 7406e4d548c7SStefano Zampini 74076e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 74081e0482f5SStefano Zampini nedcfield = -1; 74091e0482f5SStefano Zampini if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal)) { /* protects from unneded computations */ 74106e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 74116e683305SStefano Zampini const PetscInt *idxs; 74126e683305SStefano Zampini ISLocalToGlobalMapping tmap; 74136e683305SStefano Zampini 74146e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 74150be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 74166e683305SStefano Zampini /* allocate space for temporary storage */ 7417854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 7418854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 74196e683305SStefano Zampini /* allocate for IS array */ 74206e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 74211e0482f5SStefano Zampini if (pcbddc->nedclocal) { 74221e0482f5SStefano Zampini if (pcbddc->nedfield > -1) { 74231e0482f5SStefano Zampini nedcfield = pcbddc->nedfield; 74241e0482f5SStefano Zampini } else { 74251e0482f5SStefano Zampini nedcfield = 0; 74261e0482f5SStefano Zampini if (nisdofs) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen (%d)",nisdofs); 74271e0482f5SStefano Zampini nisdofs = 1; 74281e0482f5SStefano Zampini } 74291e0482f5SStefano Zampini } 74306e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 743127b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 743230368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 7433854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 74346e683305SStefano Zampini /* dofs splitting */ 74356e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 74366e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 74371e0482f5SStefano Zampini if (nedcfield != i) { 74386e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 74396e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 74406e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 74416e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 74421e0482f5SStefano Zampini } else { 74431e0482f5SStefano Zampini ierr = ISGetLocalSize(pcbddc->nedclocal,&tsize);CHKERRQ(ierr); 74441e0482f5SStefano Zampini ierr = ISGetIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 74451e0482f5SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 7446eee23b56SStefano Zampini if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping coarse nedelec field! %d != %d\n",tsize,nout); 74471e0482f5SStefano Zampini ierr = ISRestoreIndices(pcbddc->nedclocal,&idxs);CHKERRQ(ierr); 74481e0482f5SStefano Zampini } 74496e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 745030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 74516e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 74526e683305SStefano Zampini } 74536e683305SStefano Zampini /* neumann boundaries */ 74546e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 74556e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 74566e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 74576e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 74586e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 74596e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 74606e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 746130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 74626e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 74636e683305SStefano Zampini } 74646e683305SStefano Zampini /* free memory */ 74656e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 74666e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 74676e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 74686e683305SStefano Zampini } else { 74696e683305SStefano Zampini nis = 0; 74706e683305SStefano Zampini nisdofs = 0; 74716e683305SStefano Zampini nisneu = 0; 747230368db7SStefano Zampini nisvert = 0; 74736e683305SStefano Zampini isarray = NULL; 74746e683305SStefano Zampini } 74756e683305SStefano Zampini /* destroy no longer needed map */ 74766e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 74776e683305SStefano Zampini 747857de7509SStefano Zampini /* subassemble */ 747957de7509SStefano Zampini if (multilevel_allowed) { 74801ae86dd6SStefano Zampini Vec vp[1]; 74811ae86dd6SStefano Zampini PetscInt nvecs = 0; 748257de7509SStefano Zampini PetscBool reuse,reuser; 74831ae86dd6SStefano Zampini 748457de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 748557de7509SStefano Zampini else reuse = PETSC_FALSE; 748657de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 74871ae86dd6SStefano Zampini vp[0] = NULL; 74881ae86dd6SStefano Zampini if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */ 74891ae86dd6SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);CHKERRQ(ierr); 74901ae86dd6SStefano Zampini ierr = VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);CHKERRQ(ierr); 74911ae86dd6SStefano Zampini ierr = VecSetType(vp[0],VECSTANDARD);CHKERRQ(ierr); 74921ae86dd6SStefano Zampini nvecs = 1; 74931ae86dd6SStefano Zampini 74941ae86dd6SStefano Zampini if (pcbddc->divudotp) { 7495a198735bSStefano Zampini Mat B,loc_divudotp; 74961ae86dd6SStefano Zampini Vec v,p; 74971ae86dd6SStefano Zampini IS dummy; 74981ae86dd6SStefano Zampini PetscInt np; 74991ae86dd6SStefano Zampini 7500a198735bSStefano Zampini ierr = MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);CHKERRQ(ierr); 7501a198735bSStefano Zampini ierr = MatGetSize(loc_divudotp,&np,NULL);CHKERRQ(ierr); 75021ae86dd6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);CHKERRQ(ierr); 75037dae84e0SHong Zhang ierr = MatCreateSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 75041ae86dd6SStefano Zampini ierr = MatCreateVecs(B,&v,&p);CHKERRQ(ierr); 75051ae86dd6SStefano Zampini ierr = VecSet(p,1.);CHKERRQ(ierr); 75061ae86dd6SStefano Zampini ierr = MatMultTranspose(B,p,v);CHKERRQ(ierr); 75071ae86dd6SStefano Zampini ierr = VecDestroy(&p);CHKERRQ(ierr); 75081ae86dd6SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 75091ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&array);CHKERRQ(ierr); 75101ae86dd6SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_P,array);CHKERRQ(ierr); 75111ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&array);CHKERRQ(ierr); 75121ae86dd6SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);CHKERRQ(ierr); 75131ae86dd6SStefano Zampini ierr = VecResetArray(pcbddc->vec1_P);CHKERRQ(ierr); 75141ae86dd6SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 75151ae86dd6SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 751674e2c79eSStefano Zampini } 75171ae86dd6SStefano Zampini } 75181ae86dd6SStefano Zampini if (reuser) { 75191e0482f5SStefano Zampini ierr = PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);CHKERRQ(ierr); 752074e2c79eSStefano Zampini } else { 75211e0482f5SStefano 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); 75221ae86dd6SStefano Zampini } 75231ae86dd6SStefano Zampini if (vp[0]) { /* vp[0] could have been placed on a different set of processes */ 75241ae86dd6SStefano Zampini PetscScalar *arraym,*arrayv; 75251ae86dd6SStefano Zampini PetscInt nl; 75261ae86dd6SStefano Zampini ierr = VecGetLocalSize(vp[0],&nl);CHKERRQ(ierr); 75271ae86dd6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);CHKERRQ(ierr); 75281ae86dd6SStefano Zampini ierr = MatDenseGetArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 75291ae86dd6SStefano Zampini ierr = VecGetArray(vp[0],&arrayv);CHKERRQ(ierr); 75301ae86dd6SStefano Zampini ierr = PetscMemcpy(arraym,arrayv,nl*sizeof(PetscScalar));CHKERRQ(ierr); 75311ae86dd6SStefano Zampini ierr = VecRestoreArray(vp[0],&arrayv);CHKERRQ(ierr); 75321ae86dd6SStefano Zampini ierr = MatDenseRestoreArray(coarsedivudotp,&arraym);CHKERRQ(ierr); 75331ae86dd6SStefano Zampini ierr = VecDestroy(&vp[0]);CHKERRQ(ierr); 7534a198735bSStefano Zampini } else { 7535a198735bSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);CHKERRQ(ierr); 75361ae86dd6SStefano Zampini } 75371ae86dd6SStefano Zampini } else { 75381e0482f5SStefano 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); 75396e683305SStefano Zampini } 754057de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 754157de7509SStefano Zampini PetscMPIInt size; 7542f913dca9SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size);CHKERRQ(ierr); 754357de7509SStefano Zampini if (!multilevel_allowed) { 754457de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 75456e683305SStefano Zampini } else { 754657de7509SStefano Zampini Mat A; 7547779c1cceSStefano Zampini 754857de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 754957de7509SStefano Zampini if (coarse_mat_is) { 755057de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 755157de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 755257de7509SStefano Zampini coarse_mat = coarse_mat_is; 755357de7509SStefano Zampini } 755457de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 755557de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 755657de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 7557779c1cceSStefano Zampini } 7558779c1cceSStefano Zampini } 755957de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 756057de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 75616e683305SStefano Zampini 75626e683305SStefano Zampini /* create local to global scatters for coarse problem */ 756368457ee5SStefano Zampini if (compute_vecs) { 75646e683305SStefano Zampini PetscInt lrows; 75656e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 756657de7509SStefano Zampini if (coarse_mat) { 756757de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 75686e683305SStefano Zampini } else { 75696e683305SStefano Zampini lrows = 0; 75706e683305SStefano Zampini } 75716e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 75726e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 75736e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 75746e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 75756e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 75766e683305SStefano Zampini } 75776e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 7578c8587f34SStefano Zampini 7579f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 7580f9eb5b7dSStefano Zampini if (multilevel_allowed) { 7581f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 7582f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 7583f9eb5b7dSStefano Zampini } else { 7584f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 7585f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 7586c8587f34SStefano Zampini } 7587c8587f34SStefano Zampini 75886e683305SStefano Zampini /* print some info if requested */ 75896e683305SStefano Zampini if (pcbddc->dbg_flag) { 75906e683305SStefano Zampini if (!multilevel_allowed) { 75916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 75926e683305SStefano Zampini if (multilevel_requested) { 75936e683305SStefano 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); 75946e683305SStefano Zampini } else if (pcbddc->max_levels) { 75956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 75966e683305SStefano Zampini } 75976e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 75986e683305SStefano Zampini } 75996e683305SStefano Zampini } 76006e683305SStefano Zampini 76011e0482f5SStefano Zampini /* communicate coarse discrete gradient */ 76021e0482f5SStefano Zampini coarseG = NULL; 76031e0482f5SStefano Zampini if (pcbddc->nedcG && multilevel_allowed) { 76041e0482f5SStefano Zampini MPI_Comm ccomm; 76051e0482f5SStefano Zampini if (coarse_mat) { 76061e0482f5SStefano Zampini ccomm = PetscObjectComm((PetscObject)coarse_mat); 76071e0482f5SStefano Zampini } else { 76081e0482f5SStefano Zampini ccomm = MPI_COMM_NULL; 76091e0482f5SStefano Zampini } 76101e0482f5SStefano Zampini ierr = MatMPIAIJRestrict(pcbddc->nedcG,ccomm,&coarseG);CHKERRQ(ierr); 76111e0482f5SStefano Zampini } 76121e0482f5SStefano Zampini 7613f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 761457de7509SStefano Zampini if (coarse_mat) { 76156a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 76166e683305SStefano Zampini if (pcbddc->dbg_flag) { 761757de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 76186e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 76196e683305SStefano Zampini } 7620f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 7621312be037SStefano Zampini char prefix[256],str_level[16]; 7622e604994aSStefano Zampini size_t len; 76231e0482f5SStefano Zampini 762457de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 7625422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 7626c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 7627f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 762857de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7629c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 76306e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 7631c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 76321e0482f5SStefano Zampini /* TODO is this logic correct? should check for coarse_mat type */ 7633c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7634e604994aSStefano Zampini /* prefix */ 7635e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 7636e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 7637e604994aSStefano Zampini if (!pcbddc->current_level) { 7638e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 7639e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 7640c8587f34SStefano Zampini } else { 7641e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 7642312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 7643312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 764434d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 7645312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 7646e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 7647e604994aSStefano Zampini } 7648e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 76493e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 76503e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 76513e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 76523e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 7653f9eb5b7dSStefano Zampini /* allow user customization */ 7654f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 76553e3c6dadSStefano Zampini } 76563e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 765751bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 76583e3c6dadSStefano Zampini if (nisdofs) { 76593e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 76603e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 76613e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 76623e3c6dadSStefano Zampini } 76633e3c6dadSStefano Zampini } 76643e3c6dadSStefano Zampini if (nisneu) { 76653e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 76663e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 7667312be037SStefano Zampini } 766830368db7SStefano Zampini if (nisvert) { 766930368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 767030368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 767130368db7SStefano Zampini } 76721e0482f5SStefano Zampini if (coarseG) { 76731e0482f5SStefano Zampini ierr = PCBDDCSetDiscreteGradient(pc_temp,coarseG,1,nedcfield,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 76741e0482f5SStefano Zampini } 7675f9eb5b7dSStefano Zampini 7676f9eb5b7dSStefano Zampini /* get some info after set from options */ 7677f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 7678f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 76794f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 7680b76f3995Sstefano_zampini /* multilevel can only be requested via -pc_bddc_levels or PCBDDCSetLevels */ 7681b76f3995Sstefano_zampini if (isbddc && !multilevel_allowed) { 7682f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 7683f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 7684f9eb5b7dSStefano Zampini } 7685b76f3995Sstefano_zampini /* multilevel cannot be done with coarse PCs different from BDDC or NN */ 7686b76f3995Sstefano_zampini if (multilevel_requested && !isbddc && !isnn) { 7687b76f3995Sstefano_zampini ierr = PCSetType(pc_temp,PCBDDC);CHKERRQ(ierr); 7688b76f3995Sstefano_zampini isbddc = PETSC_TRUE; 7689b76f3995Sstefano_zampini isnn = PETSC_FALSE; 7690b76f3995Sstefano_zampini } 769139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 76924f3a063dSStefano Zampini if (isredundant) { 76934f3a063dSStefano Zampini KSP inner_ksp; 76944f3a063dSStefano Zampini PC inner_pc; 76959326c5c6Sstefano_zampini 76964f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 76974f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 76984f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 76994f3a063dSStefano Zampini } 7700f9eb5b7dSStefano Zampini 770157de7509SStefano Zampini /* parameters which miss an API */ 770257de7509SStefano Zampini if (isbddc) { 7703720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 7704720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 770557de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 770627b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 770727b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 7708a198735bSStefano Zampini Mat coarsedivudotp_is; 7709a198735bSStefano Zampini ISLocalToGlobalMapping l2gmap,rl2g,cl2g; 7710a198735bSStefano Zampini IS row,col; 7711a198735bSStefano Zampini const PetscInt *gidxs; 7712a198735bSStefano Zampini PetscInt n,st,M,N; 7713a198735bSStefano Zampini 7714a198735bSStefano Zampini ierr = MatGetSize(coarsedivudotp,&n,NULL);CHKERRQ(ierr); 7715a198735bSStefano Zampini ierr = MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));CHKERRQ(ierr); 7716a198735bSStefano Zampini st = st-n; 7717a198735bSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);CHKERRQ(ierr); 7718a198735bSStefano Zampini ierr = MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);CHKERRQ(ierr); 7719a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(l2gmap,&n);CHKERRQ(ierr); 7720a198735bSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7721a198735bSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);CHKERRQ(ierr); 7722a198735bSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);CHKERRQ(ierr); 7723a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(row,&rl2g);CHKERRQ(ierr); 7724a198735bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(col,&cl2g);CHKERRQ(ierr); 7725a198735bSStefano Zampini ierr = ISGetSize(row,&M);CHKERRQ(ierr); 7726a198735bSStefano Zampini ierr = MatGetSize(coarse_mat,&N,NULL);CHKERRQ(ierr); 7727a198735bSStefano Zampini ierr = ISDestroy(&row);CHKERRQ(ierr); 7728a198735bSStefano Zampini ierr = ISDestroy(&col);CHKERRQ(ierr); 7729a198735bSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);CHKERRQ(ierr); 7730a198735bSStefano Zampini ierr = MatSetType(coarsedivudotp_is,MATIS);CHKERRQ(ierr); 7731a198735bSStefano Zampini ierr = MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr); 7732a198735bSStefano Zampini ierr = MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);CHKERRQ(ierr); 7733a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&rl2g);CHKERRQ(ierr); 7734a198735bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&cl2g);CHKERRQ(ierr); 7735a198735bSStefano Zampini ierr = MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);CHKERRQ(ierr); 7736a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 77378ae0ca82SStefano Zampini ierr = PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);CHKERRQ(ierr); 7738a198735bSStefano Zampini ierr = MatDestroy(&coarsedivudotp_is);CHKERRQ(ierr); 7739720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 774059e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 7741720d30f9SStefano Zampini } 7742d4d8cf7bSStefano Zampini } 77439881197aSStefano Zampini 77443301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 77455a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 77463301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 77473301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 77483301b35fSStefano Zampini } 77493301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 77503301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 77513301b35fSStefano Zampini } 77523301b35fSStefano Zampini if (pc->pmat->spd_set) { 77533301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 77543301b35fSStefano Zampini } 775527b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 775627b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 775727b6a85dSStefano Zampini } 77586e683305SStefano Zampini /* set operators */ 77595f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 77606e683305SStefano Zampini if (pcbddc->dbg_flag) { 77616e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 77626e683305SStefano Zampini } 77636e683305SStefano Zampini } 77641e0482f5SStefano Zampini ierr = MatDestroy(&coarseG);CHKERRQ(ierr); 77656e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 7766b1ecc7b1SStefano Zampini #if 0 7767b9b85e73SStefano Zampini { 7768b9b85e73SStefano Zampini PetscViewer viewer; 7769b9b85e73SStefano Zampini char filename[256]; 7770b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 7771b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 77726a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 7773b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 7774f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 7775b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 7776b9b85e73SStefano Zampini } 7777b9b85e73SStefano Zampini #endif 7778f9eb5b7dSStefano Zampini 777998a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 778098a51de6SStefano Zampini Vec crhs,csol; 778104708bb6SStefano Zampini 7782f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 7783f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 7784f347579bSStefano Zampini if (!csol) { 77852a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 7786f9eb5b7dSStefano Zampini } 7787f347579bSStefano Zampini if (!crhs) { 77882a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 7789f347579bSStefano Zampini } 7790b0f5fe93SStefano Zampini } 77911ae86dd6SStefano Zampini ierr = MatDestroy(&coarsedivudotp);CHKERRQ(ierr); 7792b0f5fe93SStefano Zampini 7793b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 7794b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 7795b0f5fe93SStefano Zampini 7796b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 77974f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 77984f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 77994f1b2e48SStefano Zampini } 7800b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 7801b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 7802b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7803b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7804b0f5fe93SStefano Zampini if (coarse_mat) { 7805b0f5fe93SStefano Zampini Vec nullv; 7806b0f5fe93SStefano Zampini PetscScalar *array,*array2; 7807b0f5fe93SStefano Zampini PetscInt nl; 7808b0f5fe93SStefano Zampini 7809b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 7810b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 7811b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7812b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 7813b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 7814b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 7815b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 7816b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 7817b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 7818b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 7819b0f5fe93SStefano Zampini } 7820b0f5fe93SStefano Zampini } 7821b0f5fe93SStefano Zampini 7822b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 7823b0f5fe93SStefano Zampini PetscBool ispreonly; 7824b0f5fe93SStefano Zampini 7825b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7826b0f5fe93SStefano Zampini PetscBool isnull; 7827b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 7828bef83e63SStefano Zampini if (isnull) { 7829b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 7830b0f5fe93SStefano Zampini } 7831bef83e63SStefano Zampini /* TODO: add local nullspaces (if any) */ 7832b0f5fe93SStefano Zampini } 7833b0f5fe93SStefano Zampini /* setup coarse ksp */ 7834b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 7835cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 7836cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 78376e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 7838c8587f34SStefano Zampini KSP check_ksp; 78392b510759SStefano Zampini KSPType check_ksp_type; 7840c8587f34SStefano Zampini PC check_pc; 78416e683305SStefano Zampini Vec check_vec,coarse_vec; 78426a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 78432b510759SStefano Zampini PetscInt its; 78446e683305SStefano Zampini PetscBool compute_eigs; 78456e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 78466e683305SStefano Zampini PetscInt neigs; 78478e185a42SStefano Zampini const char *prefix; 7848c8587f34SStefano Zampini 78492b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 78506e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 7851422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 785223ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 7853f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 7854e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 7855e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 7856e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 78572b510759SStefano Zampini if (ispreonly) { 78582b510759SStefano Zampini check_ksp_type = KSPPREONLY; 78596e683305SStefano Zampini compute_eigs = PETSC_FALSE; 78602b510759SStefano Zampini } else { 7861cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 78626e683305SStefano Zampini compute_eigs = PETSC_TRUE; 7863c8587f34SStefano Zampini } 7864c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 78656e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 78666e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 78676e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 7868a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 7869a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 7870a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 7871a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 7872c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 7873c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 7874c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 7875c8587f34SStefano Zampini /* create random vec */ 78762701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 7877c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 78786e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 7879c8587f34SStefano Zampini /* solve coarse problem */ 78806e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 7881cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 78826e683305SStefano Zampini if (compute_eigs) { 7883854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 7884854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 78856e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 78861ae86dd6SStefano Zampini if (neigs) { 78876e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 78886e683305SStefano Zampini lambda_min = eigs_r[0]; 78896e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 78902701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 78912701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 7892cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 7893cbcc2c2aSStefano Zampini } 7894c8587f34SStefano Zampini } 7895c8587f34SStefano Zampini } 78961ae86dd6SStefano Zampini } 7897cbcc2c2aSStefano Zampini 7898c8587f34SStefano Zampini /* check coarse problem residual error */ 78996e683305SStefano Zampini if (pcbddc->dbg_flag) { 79006e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 79016e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 79026e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 7903c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 79046e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 79056e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 7906779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 79076e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 79086e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 79096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 79106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 7911b0f5fe93SStefano Zampini if (CoarseNullSpace) { 7912b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 7913b0f5fe93SStefano Zampini } 79146e683305SStefano Zampini if (compute_eigs) { 79156e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 7916b03ebc13SStefano Zampini KSPConvergedReason reason; 7917deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 7918c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 7919b03ebc13SStefano Zampini ierr = KSPGetConvergedReason(check_ksp,&reason);CHKERRQ(ierr); 79206e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 7921b03ebc13SStefano 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); 79226e683305SStefano Zampini for (i=0;i<neigs;i++) { 79236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 7924c8587f34SStefano Zampini } 79256e683305SStefano Zampini } 79266e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 79276e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 79286e683305SStefano Zampini } 7929e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 79302701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 7931c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 79326e683305SStefano Zampini if (compute_eigs) { 79336e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 79346e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 7935c8587f34SStefano Zampini } 79366e683305SStefano Zampini } 79376e683305SStefano Zampini } 7938bef83e63SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 7939cbcc2c2aSStefano Zampini /* print additional info */ 7940cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 79416e683305SStefano Zampini /* waits until all processes reaches this point */ 79426e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 7943cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 7944cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7945cbcc2c2aSStefano Zampini } 7946cbcc2c2aSStefano Zampini 79472b510759SStefano Zampini /* free memory */ 7948fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 7949c8587f34SStefano Zampini PetscFunctionReturn(0); 7950c8587f34SStefano Zampini } 7951674ae819SStefano Zampini 7952f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 7953f34684f1SStefano Zampini { 7954f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 7955f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 7956f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 7957dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 7958dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 795973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 7960dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 7961f34684f1SStefano Zampini PetscErrorCode ierr; 7962f34684f1SStefano Zampini 7963f34684f1SStefano Zampini PetscFunctionBegin; 7964f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 79656c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 7966dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 79673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 7968dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7969dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 79706583bcc1SStefano Zampini ierr = ISRenumber(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 7971dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 7972dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 7973dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 79746c4ed002SBarry 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); 7975dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 7976dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7977dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 7978dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 7979dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 7980f34684f1SStefano Zampini 7981f34684f1SStefano Zampini /* check numbering */ 7982f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 7983019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 7984dc456d91SStefano Zampini PetscInt i; 7985b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 7986f34684f1SStefano Zampini 7987f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 7988f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 7989f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 79901575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 7991019a44ceSStefano Zampini /* counter */ 7992019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 7993019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 7994019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7995019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 7996019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7997019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 7998f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 7999f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 8000727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 8001f34684f1SStefano Zampini } 8002f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8003f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8004f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8005e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8006e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8007e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8008e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8009f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8010019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8011f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8012019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 80132c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 801475c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 8015b9b85e73SStefano Zampini set_error = PETSC_TRUE; 80162c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 80172c66d082SStefano 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); 8018f34684f1SStefano Zampini } 8019f34684f1SStefano Zampini } 8020019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 8021b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 8022f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8023f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 8024f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 8025f34684f1SStefano Zampini } 8026f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8027f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 8028e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8029e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8030f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 8031f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 8032b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 8033ca8b9ea9SStefano Zampini PetscInt *gidxs; 8034ca8b9ea9SStefano Zampini 8035ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 80363bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 8037f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 8038f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8039f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 8040f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 80414bc2dc4bSStefano 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); 8042f34684f1SStefano Zampini } 8043f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8044ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 8045f34684f1SStefano Zampini } 8046f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 80471575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8048302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 8049f34684f1SStefano Zampini } 80508bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 8051f34684f1SStefano Zampini /* get back data */ 8052f34684f1SStefano Zampini *coarse_size_n = coarse_size; 8053f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 8054674ae819SStefano Zampini PetscFunctionReturn(0); 8055674ae819SStefano Zampini } 8056674ae819SStefano Zampini 8057a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 8058e456f2a8SStefano Zampini { 8059e456f2a8SStefano Zampini IS localis_t; 8060a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 8061e456f2a8SStefano Zampini PetscScalar *vals; 8062e456f2a8SStefano Zampini PetscErrorCode ierr; 8063e456f2a8SStefano Zampini 8064e456f2a8SStefano Zampini PetscFunctionBegin; 8065a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 8066e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 8067854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 8068e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 8069e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 8070a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 8071a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 80721035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 8073a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 80741035eff8SStefano Zampini } 8075a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 8076e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 8077e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 8078a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 8079a7dc3881SStefano Zampini /* now compute set in local ordering */ 8080a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8081a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8082a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8083a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 8084a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8085ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8086e456f2a8SStefano Zampini lsize++; 8087e456f2a8SStefano Zampini } 8088e456f2a8SStefano Zampini } 8089854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 8090a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 8091ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 8092e456f2a8SStefano Zampini idxs[lsize++] = i; 8093e456f2a8SStefano Zampini } 8094e456f2a8SStefano Zampini } 8095a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 8096a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 8097e456f2a8SStefano Zampini *localis = localis_t; 8098e456f2a8SStefano Zampini PetscFunctionReturn(0); 8099e456f2a8SStefano Zampini } 8100906d46d4SStefano Zampini 810108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 8102b96c3477SStefano Zampini { 8103a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8104b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8105b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 8106a64f4aa4SStefano Zampini Mat S_j; 8107b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 8108b96c3477SStefano Zampini PetscBool free_used_adj; 8109b96c3477SStefano Zampini PetscErrorCode ierr; 8110b96c3477SStefano Zampini 8111b96c3477SStefano Zampini PetscFunctionBegin; 8112b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 8113b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 811408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 8115b96c3477SStefano Zampini used_xadj = NULL; 8116b96c3477SStefano Zampini used_adjncy = NULL; 8117b96c3477SStefano Zampini } else { 811808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 811908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 812008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 812108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 8122b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 8123b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 8124b96c3477SStefano Zampini } else { 81252fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 8126b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 8127b96c3477SStefano Zampini PetscInt nvtxs; 8128b96c3477SStefano Zampini 81292fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 81302fffb893SStefano Zampini if (flg_row) { 8131b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 8132b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 8133b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 8134b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 81352fffb893SStefano Zampini } else { 81362fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 81372fffb893SStefano Zampini used_xadj = NULL; 81382fffb893SStefano Zampini used_adjncy = NULL; 81392fffb893SStefano Zampini } 81402fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 8141b96c3477SStefano Zampini } 8142b96c3477SStefano Zampini } 8143d5574798SStefano Zampini 8144d5574798SStefano Zampini /* setup sub_schurs data */ 8145a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8146df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 8147df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 8148a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 814991af6908SStefano 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); 8150a64f4aa4SStefano Zampini } else { 81516816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 81524d7f8f00SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE; 8153a3df083aSStefano Zampini PetscInt benign_n; 815472b8c272SStefano Zampini Mat change = NULL; 81559d54b7f4SStefano Zampini Vec scaling = NULL; 815672b8c272SStefano Zampini IS change_primal = NULL; 8157a3df083aSStefano Zampini 81585feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 81595feab87aSStefano Zampini PetscInt n_vertices; 81605feab87aSStefano Zampini 81615feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 81622034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 81635feab87aSStefano Zampini } 816404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 816504708bb6SStefano Zampini if (!isseqaij) { 816604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 816704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 816804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 816904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 817004708bb6SStefano Zampini } else { 8171511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 817204708bb6SStefano Zampini } 817304708bb6SStefano Zampini } 8174a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 8175a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 8176ca92afb2SStefano Zampini } else { 8177a3df083aSStefano Zampini benign_n = 0; 8178ca92afb2SStefano Zampini } 8179b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 8180b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 8181b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 818272b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 818322db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 8184b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 818522db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 8186b7ab4a40SStefano Zampini } 8187b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 8188b7ab4a40SStefano 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 */ 8189b7ab4a40SStefano Zampini if (need_change) { 819088c03ad3SStefano Zampini PC_IS *pcisf; 819188c03ad3SStefano Zampini PC_BDDC *pcbddcf; 819288c03ad3SStefano Zampini PC pcf; 819388c03ad3SStefano Zampini 8194e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 819588c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 819688c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 819788c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 8198b9be95fcSstefano_zampini 819988c03ad3SStefano Zampini /* hacks */ 820088c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 820172b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 820272b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 820372b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 820472b8c272SStefano Zampini pcisf->n = pcis->n; 820572b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 820688c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 820788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 820888c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 820988c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 821088c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 821188c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 821272b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 821388c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 8214b9be95fcSstefano_zampini 8215b9be95fcSstefano_zampini /* setup constraints so that we can get information on primal vertices and change of basis (in local numbering) */ 821688c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 821772b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 821872b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 821972b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 822072b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 8221b9be95fcSstefano_zampini 822288c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 822372b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 822488c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 822588c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 822688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 822788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 822888c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 822988c03ad3SStefano Zampini pcf->ops->destroy = NULL; 8230b9be95fcSstefano_zampini pcf->ops->reset = NULL; 823188c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 823288c03ad3SStefano Zampini } 82339d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 823491af6908SStefano 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); 823572b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 823672b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 8237ca92afb2SStefano Zampini } 8238d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8239b96c3477SStefano Zampini 8240b96c3477SStefano Zampini /* free adjacency */ 8241b96c3477SStefano Zampini if (free_used_adj) { 8242b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 8243b96c3477SStefano Zampini } 8244b96c3477SStefano Zampini PetscFunctionReturn(0); 8245b96c3477SStefano Zampini } 8246b96c3477SStefano Zampini 824708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 8248b96c3477SStefano Zampini { 8249b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8250b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8251b96c3477SStefano Zampini PCBDDCGraph graph; 8252b96c3477SStefano Zampini PetscErrorCode ierr; 8253b96c3477SStefano Zampini 8254b96c3477SStefano Zampini PetscFunctionBegin; 8255b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 825608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 82573301b35fSStefano Zampini IS verticesIS,verticescomm; 82583301b35fSStefano Zampini PetscInt vsize,*idxs; 8259b96c3477SStefano Zampini 8260b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 82613301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 82623301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 82633301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 82643301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 8265c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 8266b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 8267be12c134Sstefano_zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);CHKERRQ(ierr); 8268441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 82693301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 8270b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 8271b96c3477SStefano Zampini } else { 8272b96c3477SStefano Zampini graph = pcbddc->mat_graph; 8273b96c3477SStefano Zampini } 8274e4d548c7SStefano Zampini /* print some info */ 82755c643e28SStefano Zampini if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) { 8276e4d548c7SStefano Zampini IS vertices; 8277e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 8278c8272957SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 8279e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8280e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 8281e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8282e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 8283e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 8284e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 8285e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 8286e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 8287e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 8288c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 8289e4d548c7SStefano Zampini } 8290b96c3477SStefano Zampini 8291b96c3477SStefano Zampini /* sub_schurs init */ 8292b334f244SStefano Zampini if (!pcbddc->sub_schurs) { 8293b334f244SStefano Zampini ierr = PCBDDCSubSchursCreate(&pcbddc->sub_schurs);CHKERRQ(ierr); 8294b334f244SStefano Zampini } 82958b6046baSStefano Zampini ierr = PCBDDCSubSchursInit(pcbddc->sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);CHKERRQ(ierr); 8296e62b6521Sstefano_zampini pcbddc->sub_schurs->prefix = ((PetscObject)pc)->prefix; 8297a64f4aa4SStefano Zampini 8298b96c3477SStefano Zampini /* free graph struct */ 829908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 8300b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 8301b96c3477SStefano Zampini } 8302b96c3477SStefano Zampini PetscFunctionReturn(0); 8303b96c3477SStefano Zampini } 8304fa34dd3eSStefano Zampini 8305fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 8306fa34dd3eSStefano Zampini { 8307fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 8308fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 8309fa34dd3eSStefano Zampini PetscErrorCode ierr; 8310fa34dd3eSStefano Zampini 8311fa34dd3eSStefano Zampini PetscFunctionBegin; 8312fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 8313fa34dd3eSStefano Zampini IS zerodiag = NULL; 83144f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 8315fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 83164f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 831775c01103SStefano Zampini PetscReal norm; 8318fa34dd3eSStefano Zampini PetscInt i; 8319fa34dd3eSStefano Zampini 8320fa34dd3eSStefano Zampini /* B0 and B0_B */ 8321fa34dd3eSStefano Zampini if (zerodiag) { 8322fa34dd3eSStefano Zampini IS dummy; 8323fa34dd3eSStefano Zampini 83244f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 83257dae84e0SHong Zhang ierr = MatCreateSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 8326fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 8327fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 8328fa34dd3eSStefano Zampini } 8329fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 8330fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 8331fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 8332fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8333fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8334fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8335fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8336fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 8337fa34dd3eSStefano Zampini /* S_j */ 8338fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 8339fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 8340fa34dd3eSStefano Zampini 8341fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 8342fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 8343fa34dd3eSStefano Zampini /* continuous in primal space */ 8344fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 8345fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8346fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8347fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 83484f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 83494f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 8350fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8351fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8352fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8353fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8354fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8355fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8356fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 8357fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 8358fa34dd3eSStefano Zampini 8359fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 8360fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 8361fa34dd3eSStefano Zampini /* local with Schur */ 8362fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 8363fa34dd3eSStefano Zampini if (zerodiag) { 8364fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 83654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 8366fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8367fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 8368fa34dd3eSStefano Zampini } 8369fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 8370fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8371fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8372fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8373fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8374fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 8375fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 8376fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 8377fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 8378fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8379fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8380fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8381fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 8382fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8383fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 8384fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 8385fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 8386fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 8387fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 8388fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 8389fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8390fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 8391fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 8392fa34dd3eSStefano Zampini if (zerodiag) { 8393fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 8394fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 83954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 8396fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 8397fa34dd3eSStefano Zampini } 8398fa34dd3eSStefano Zampini /* BDDC */ 8399fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 8400fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 8401fa34dd3eSStefano Zampini 8402fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 8403fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 8404fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 8405fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 84064f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 84074f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 8408fa34dd3eSStefano Zampini } 84094f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 8410fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 8411fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 8412fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 8413fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 8414fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 8415fa34dd3eSStefano Zampini } 8416fa34dd3eSStefano Zampini PetscFunctionReturn(0); 8417fa34dd3eSStefano Zampini } 84181e0482f5SStefano Zampini 84191e0482f5SStefano Zampini #include <../src/mat/impls/aij/mpi/mpiaij.h> 84201e0482f5SStefano Zampini PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B) 84211e0482f5SStefano Zampini { 84221e0482f5SStefano Zampini Mat At; 84231e0482f5SStefano Zampini IS rows; 84241e0482f5SStefano Zampini PetscInt rst,ren; 84251e0482f5SStefano Zampini PetscErrorCode ierr; 84261e0482f5SStefano Zampini PetscLayout rmap; 84271e0482f5SStefano Zampini 84281e0482f5SStefano Zampini PetscFunctionBegin; 84291e0482f5SStefano Zampini rst = ren = 0; 84301e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 84311e0482f5SStefano Zampini ierr = PetscLayoutCreate(ccomm,&rmap);CHKERRQ(ierr); 84321e0482f5SStefano Zampini ierr = PetscLayoutSetSize(rmap,A->rmap->N);CHKERRQ(ierr); 84331e0482f5SStefano Zampini ierr = PetscLayoutSetBlockSize(rmap,1);CHKERRQ(ierr); 84341e0482f5SStefano Zampini ierr = PetscLayoutSetUp(rmap);CHKERRQ(ierr); 84351e0482f5SStefano Zampini ierr = PetscLayoutGetRange(rmap,&rst,&ren);CHKERRQ(ierr); 84361e0482f5SStefano Zampini } 8437e07686f2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)A),ren-rst,rst,1,&rows);CHKERRQ(ierr); 84387dae84e0SHong Zhang ierr = MatCreateSubMatrix(A,rows,NULL,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 84391e0482f5SStefano Zampini ierr = ISDestroy(&rows);CHKERRQ(ierr); 84401e0482f5SStefano Zampini 84411e0482f5SStefano Zampini if (ccomm != MPI_COMM_NULL) { 84421e0482f5SStefano Zampini Mat_MPIAIJ *a,*b; 84431e0482f5SStefano Zampini IS from,to; 84441e0482f5SStefano Zampini Vec gvec; 84451e0482f5SStefano Zampini PetscInt lsize; 84461e0482f5SStefano Zampini 84471e0482f5SStefano Zampini ierr = MatCreate(ccomm,B);CHKERRQ(ierr); 84481e0482f5SStefano Zampini ierr = MatSetSizes(*B,ren-rst,PETSC_DECIDE,PETSC_DECIDE,At->cmap->N);CHKERRQ(ierr); 84491e0482f5SStefano Zampini ierr = MatSetType(*B,MATAIJ);CHKERRQ(ierr); 84501e0482f5SStefano Zampini ierr = PetscLayoutDestroy(&((*B)->rmap));CHKERRQ(ierr); 84511e0482f5SStefano Zampini ierr = PetscLayoutSetUp((*B)->cmap);CHKERRQ(ierr); 84521e0482f5SStefano Zampini a = (Mat_MPIAIJ*)At->data; 84531e0482f5SStefano Zampini b = (Mat_MPIAIJ*)(*B)->data; 84541e0482f5SStefano Zampini ierr = MPI_Comm_size(ccomm,&b->size);CHKERRQ(ierr); 84551e0482f5SStefano Zampini ierr = MPI_Comm_rank(ccomm,&b->rank);CHKERRQ(ierr); 84561e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->A);CHKERRQ(ierr); 84571e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->B);CHKERRQ(ierr); 84581e0482f5SStefano Zampini b->A = a->A; 84591e0482f5SStefano Zampini b->B = a->B; 84601e0482f5SStefano Zampini 84611e0482f5SStefano Zampini b->donotstash = a->donotstash; 84621e0482f5SStefano Zampini b->roworiented = a->roworiented; 84631e0482f5SStefano Zampini b->rowindices = 0; 84641e0482f5SStefano Zampini b->rowvalues = 0; 84651e0482f5SStefano Zampini b->getrowactive = PETSC_FALSE; 84661e0482f5SStefano Zampini 84671e0482f5SStefano Zampini (*B)->rmap = rmap; 84681e0482f5SStefano Zampini (*B)->factortype = A->factortype; 84691e0482f5SStefano Zampini (*B)->assembled = PETSC_TRUE; 84701e0482f5SStefano Zampini (*B)->insertmode = NOT_SET_VALUES; 84711e0482f5SStefano Zampini (*B)->preallocated = PETSC_TRUE; 84721e0482f5SStefano Zampini 84731e0482f5SStefano Zampini if (a->colmap) { 84741e0482f5SStefano Zampini #if defined(PETSC_USE_CTABLE) 84751e0482f5SStefano Zampini ierr = PetscTableCreateCopy(a->colmap,&b->colmap);CHKERRQ(ierr); 84761e0482f5SStefano Zampini #else 84771e0482f5SStefano Zampini ierr = PetscMalloc1(At->cmap->N,&b->colmap);CHKERRQ(ierr); 84781e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)*B,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 84791e0482f5SStefano Zampini ierr = PetscMemcpy(b->colmap,a->colmap,At->cmap->N*sizeof(PetscInt));CHKERRQ(ierr); 84801e0482f5SStefano Zampini #endif 84811e0482f5SStefano Zampini } else b->colmap = 0; 84821e0482f5SStefano Zampini if (a->garray) { 84831e0482f5SStefano Zampini PetscInt len; 84841e0482f5SStefano Zampini len = a->B->cmap->n; 84851e0482f5SStefano Zampini ierr = PetscMalloc1(len+1,&b->garray);CHKERRQ(ierr); 84861e0482f5SStefano Zampini ierr = PetscLogObjectMemory((PetscObject)(*B),len*sizeof(PetscInt));CHKERRQ(ierr); 84871e0482f5SStefano Zampini if (len) { ierr = PetscMemcpy(b->garray,a->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 84881e0482f5SStefano Zampini } else b->garray = 0; 84891e0482f5SStefano Zampini 84901e0482f5SStefano Zampini ierr = PetscObjectReference((PetscObject)a->lvec);CHKERRQ(ierr); 84911e0482f5SStefano Zampini b->lvec = a->lvec; 84921e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->lvec);CHKERRQ(ierr); 84931e0482f5SStefano Zampini 84941e0482f5SStefano Zampini /* cannot use VecScatterCopy */ 84951e0482f5SStefano Zampini ierr = VecGetLocalSize(b->lvec,&lsize);CHKERRQ(ierr); 84961e0482f5SStefano Zampini ierr = ISCreateGeneral(ccomm,lsize,b->garray,PETSC_USE_POINTER,&from);CHKERRQ(ierr); 84971e0482f5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,lsize,0,1,&to);CHKERRQ(ierr); 84981e0482f5SStefano Zampini ierr = MatCreateVecs(*B,&gvec,NULL);CHKERRQ(ierr); 84991e0482f5SStefano Zampini ierr = VecScatterCreate(gvec,from,b->lvec,to,&b->Mvctx);CHKERRQ(ierr); 85001e0482f5SStefano Zampini ierr = PetscLogObjectParent((PetscObject)*B,(PetscObject)b->Mvctx);CHKERRQ(ierr); 85011e0482f5SStefano Zampini ierr = ISDestroy(&from);CHKERRQ(ierr); 85021e0482f5SStefano Zampini ierr = ISDestroy(&to);CHKERRQ(ierr); 85031e0482f5SStefano Zampini ierr = VecDestroy(&gvec);CHKERRQ(ierr); 85041e0482f5SStefano Zampini } 85051e0482f5SStefano Zampini ierr = MatDestroy(&At);CHKERRQ(ierr); 85061e0482f5SStefano Zampini PetscFunctionReturn(0); 85071e0482f5SStefano Zampini } 8508