1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini 8674ae819SStefano Zampini #undef __FUNCT__ 9c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 10c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 11c263805aSStefano Zampini { 12c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13c263805aSStefano Zampini const PetscInt *idxs; 14c263805aSStefano Zampini PetscInt nz = 0; 15c263805aSStefano Zampini PetscErrorCode ierr; 16c263805aSStefano Zampini 17c263805aSStefano Zampini PetscFunctionBegin; 18c263805aSStefano Zampini if (!pcbddc->zerodiag) { 19c263805aSStefano Zampini PetscFunctionReturn(0); 20c263805aSStefano Zampini } else { 21c263805aSStefano Zampini ierr = ISGetLocalSize(pcbddc->zerodiag,&nz);CHKERRQ(ierr); 22c263805aSStefano Zampini if (!nz) PetscFunctionReturn(0); 23c263805aSStefano Zampini } 24c263805aSStefano Zampini ierr = ISGetIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 25c263805aSStefano Zampini /* TODO: add error checking 26c263805aSStefano Zampini - avoid nested pop (or push) calls. 27c263805aSStefano Zampini - cannot push before pop. 28c263805aSStefano Zampini */ 29c263805aSStefano Zampini if (pop) { 30c263805aSStefano Zampini const PetscInt *cB0_cols; 31c263805aSStefano Zampini PetscInt cB0_ncol; 32c263805aSStefano Zampini const PetscScalar *cB0_vals; 33c263805aSStefano Zampini 34c263805aSStefano Zampini /* extract B_0 */ 35c263805aSStefano Zampini ierr = MatGetRow(pcbddc->local_mat,idxs[nz-1],&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 36c263805aSStefano Zampini pcbddc->B0_ncol = cB0_ncol; 37c263805aSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 38c263805aSStefano Zampini ierr = PetscMalloc2(cB0_ncol,&pcbddc->B0_cols,cB0_ncol,&pcbddc->B0_vals);CHKERRQ(ierr); 39c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_cols,cB0_cols,cB0_ncol*sizeof(PetscInt));CHKERRQ(ierr); 40c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_vals,cB0_vals,cB0_ncol*sizeof(PetscScalar));CHKERRQ(ierr); 41c263805aSStefano Zampini ierr = MatRestoreRow(pcbddc->local_mat,idxs[nz-1],&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 42c263805aSStefano Zampini /* remove rows and cols from local problem */ 43c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 44c263805aSStefano Zampini ierr = MatZeroRowsColumns(pcbddc->local_mat,1,idxs+nz-1,1.,NULL,NULL);CHKERRQ(ierr); 45c263805aSStefano Zampini } else { /* push */ 46c263805aSStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,&idxs[nz-1],pcbddc->B0_ncol,pcbddc->B0_cols,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 47*7b026bddSStefano Zampini ierr = MatSetValues(pcbddc->local_mat,pcbddc->B0_ncol,pcbddc->B0_cols,1,&idxs[nz-1],pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 48c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 49c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 50c263805aSStefano Zampini } 51c263805aSStefano Zampini PetscFunctionReturn(0); 52c263805aSStefano Zampini } 53c263805aSStefano Zampini 54c263805aSStefano Zampini #undef __FUNCT__ 55b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 5608122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 57b1b3d7a2SStefano Zampini { 58b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 6008122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 6108122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 6208122e43SStefano Zampini PetscScalar *work,lwork; 6308122e43SStefano Zampini PetscScalar *St,*S,*eigv; 6408122e43SStefano Zampini PetscScalar *Sarray,*Starray; 6508122e43SStefano Zampini PetscReal *eigs,thresh; 661b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 67f6f667cfSStefano Zampini PetscBool allocated_S_St; 6808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 6908122e43SStefano Zampini PetscReal *rwork; 7008122e43SStefano Zampini #endif 71b1b3d7a2SStefano Zampini PetscErrorCode ierr; 72b1b3d7a2SStefano Zampini 73b1b3d7a2SStefano Zampini PetscFunctionBegin; 7408122e43SStefano Zampini if (!sub_schurs->use_mumps) { 7508122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 7608122e43SStefano Zampini } 7708122e43SStefano Zampini 7806a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 7906a4e24aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 8006a4e24aSStefano Zampini } 8106a4e24aSStefano Zampini 82fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 83fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 84fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 85fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 86fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 87fd14bc51SStefano Zampini } 88fd14bc51SStefano Zampini 89e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 90e496cd5dSStefano 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); 91e496cd5dSStefano Zampini } 92e496cd5dSStefano Zampini 9308122e43SStefano Zampini /* max size of subsets */ 9408122e43SStefano Zampini mss = 0; 9508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 9608122e43SStefano Zampini PetscInt subset_size; 97862806e4SStefano Zampini 9808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 9908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 10008122e43SStefano Zampini } 10108122e43SStefano Zampini 10208122e43SStefano Zampini /* min/max and threshold */ 10308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 104f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 10508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 106f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 107f6f667cfSStefano Zampini if (nmin) { 108f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 109f6f667cfSStefano Zampini } 11008122e43SStefano Zampini 11108122e43SStefano Zampini /* allocate lapack workspace */ 11208122e43SStefano Zampini cum = cum2 = 0; 11308122e43SStefano Zampini maxneigs = 0; 11408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 11508122e43SStefano Zampini PetscInt n,subset_size; 116f6f667cfSStefano Zampini 11708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 11808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 1199162d606SStefano Zampini cum += subset_size; 1209162d606SStefano Zampini cum2 += subset_size*n; 12108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 12208122e43SStefano Zampini } 12308122e43SStefano Zampini if (mss) { 1249ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 12508122e43SStefano Zampini PetscBLASInt B_itype = 1; 12608122e43SStefano Zampini PetscBLASInt B_N = mss; 1274c6709b3SStefano Zampini PetscReal zero = 0.0; 1284c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 12908122e43SStefano Zampini 13008122e43SStefano Zampini B_lwork = -1; 13108122e43SStefano Zampini S = NULL; 13208122e43SStefano Zampini St = NULL; 133a58a30b4SStefano Zampini eigs = NULL; 134a58a30b4SStefano Zampini eigv = NULL; 135a58a30b4SStefano Zampini B_iwork = NULL; 136a58a30b4SStefano Zampini B_ifail = NULL; 137d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 138d1710679SStefano Zampini rwork = NULL; 139d1710679SStefano Zampini #endif 1408bec7fa6SStefano Zampini thresh = 1.0; 14108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 14208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 14308122e43SStefano 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)); 14408122e43SStefano Zampini #else 14508122e43SStefano 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)); 14608122e43SStefano Zampini #endif 14708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 14808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 14908122e43SStefano Zampini } else { 15008122e43SStefano Zampini /* TODO */ 15108122e43SStefano Zampini } 15208122e43SStefano Zampini } else { 15308122e43SStefano Zampini lwork = 0; 15408122e43SStefano Zampini } 15508122e43SStefano Zampini 15608122e43SStefano Zampini nv = 0; 157d62866d3SStefano 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) */ 158d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 15908122e43SStefano Zampini } 1604c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 161f6f667cfSStefano Zampini if (allocated_S_St) { 162f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 163f6f667cfSStefano Zampini } 164f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 16508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 16608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 16708122e43SStefano Zampini #endif 1689162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1699162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1709162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 17108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1729162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 17308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 17408122e43SStefano Zampini 17508122e43SStefano Zampini maxneigs = 0; 17608122e43SStefano Zampini cum = cum2 = cumarray = 0; 1779162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1789162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 179d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 18008122e43SStefano Zampini const PetscInt *idxs; 18108122e43SStefano Zampini 182d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 18308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 18408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 18508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 18608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1879162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1889162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 18908122e43SStefano Zampini } 19008122e43SStefano Zampini cum2 = cum; 191d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 19208122e43SStefano Zampini } 19308122e43SStefano Zampini 19408122e43SStefano Zampini if (mss) { /* multilevel */ 19508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 19608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 19708122e43SStefano Zampini } 19808122e43SStefano Zampini 19908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 20008122e43SStefano Zampini 20108122e43SStefano Zampini const PetscInt *idxs; 202f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 203862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 20408122e43SStefano Zampini PetscBLASInt B_N; 205aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 20608122e43SStefano Zampini 207862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 208f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 209f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 2109ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 211aff50787SStefano Zampini PetscInt j,k; 212aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 213aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 214aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 21508122e43SStefano Zampini } 21608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 217aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 218aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 219aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 220aff50787SStefano Zampini } 22108122e43SStefano Zampini } 22208122e43SStefano Zampini } else { 22308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 22408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 22508122e43SStefano Zampini } 2268bec7fa6SStefano Zampini } else { 227f6f667cfSStefano Zampini S = Sarray + cumarray; 228f6f667cfSStefano Zampini St = Starray + cumarray; 2298bec7fa6SStefano Zampini } 23008122e43SStefano Zampini 231f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 232aff50787SStefano Zampini /* see if we can save some work */ 233aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 234aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 235aff50787SStefano Zampini } 236aff50787SStefano Zampini 237aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 238aff50787SStefano Zampini B_neigs = 0; 239aff50787SStefano Zampini } else { 240aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 241f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 242f6f667cfSStefano Zampini 2439ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 24408122e43SStefano Zampini PetscBLASInt B_itype = 1; 245f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 2464c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 2479552c7c7SStefano Zampini PetscInt nmin_s; 24808122e43SStefano Zampini 249f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 250fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2518bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 252fd14bc51SStefano Zampini } 25308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 255f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 25608122e43SStefano Zampini #else 257f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 25808122e43SStefano Zampini #endif 25908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 26008122e43SStefano Zampini if (B_ierr) { 26108122e43SStefano Zampini if (B_ierr < 0 ) { 26208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 26308122e43SStefano Zampini } else if (B_ierr <= B_N) { 26408122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 26508122e43SStefano Zampini } else { 2669552c7c7SStefano Zampini 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); 26708122e43SStefano Zampini } 26808122e43SStefano Zampini } 26908122e43SStefano Zampini 27008122e43SStefano Zampini if (B_neigs > nmax) { 271fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 272fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 273fd14bc51SStefano Zampini } 274f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 27508122e43SStefano Zampini B_neigs = nmax; 27608122e43SStefano Zampini } 27708122e43SStefano Zampini 2789552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2799552c7c7SStefano Zampini if (B_neigs < nmin_s) { 28008122e43SStefano Zampini PetscBLASInt B_neigs2; 28108122e43SStefano Zampini 282f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 283f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 284fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 285fd14bc51SStefano 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); 286fd14bc51SStefano Zampini } 2879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 28808122e43SStefano Zampini PetscInt j; 28908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 29008122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 29108122e43SStefano Zampini } 29208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 29308122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 29408122e43SStefano Zampini } 29508122e43SStefano Zampini } else { 29608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 29708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 29808122e43SStefano Zampini } 29908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 30008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 301f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&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)); 30208122e43SStefano Zampini #else 303f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&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)); 30408122e43SStefano Zampini #endif 30508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 30608122e43SStefano Zampini B_neigs += B_neigs2; 30708122e43SStefano Zampini } 30808122e43SStefano Zampini if (B_ierr) { 30908122e43SStefano Zampini if (B_ierr < 0 ) { 31008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 31108122e43SStefano Zampini } else if (B_ierr <= B_N) { 31208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 31308122e43SStefano Zampini } else { 3149552c7c7SStefano Zampini 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); 31508122e43SStefano Zampini } 31608122e43SStefano Zampini } 317fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 318ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 31908122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 32008122e43SStefano Zampini if (eigs[j] == 0.0) { 321ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 32208122e43SStefano Zampini } else { 323ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 324fd14bc51SStefano Zampini } 32508122e43SStefano Zampini } 32608122e43SStefano Zampini } 32708122e43SStefano Zampini } else { 32808122e43SStefano Zampini /* TODO */ 32908122e43SStefano Zampini } 330aff50787SStefano Zampini } 3318bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 3328bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 3339162d606SStefano Zampini if (B_neigs) { 3349162d606SStefano 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); 335fd14bc51SStefano Zampini 336fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 3379552c7c7SStefano Zampini PetscInt ii; 3389552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 339ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 3409552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 341ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 342ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 343ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 344ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 345ac47001eSStefano Zampini #else 346ac47001eSStefano 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); 347ac47001eSStefano Zampini #endif 3489552c7c7SStefano Zampini } 3499552c7c7SStefano Zampini } 350fd14bc51SStefano Zampini } 35108122e43SStefano Zampini #if 0 3529162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 35308122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 35408122e43SStefano Zampini PetscScalar norm; 35508122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3569162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3579162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 35808122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 35908122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36008122e43SStefano Zampini } else { 36108122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 36208122e43SStefano Zampini } 3639162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 364b1b3d7a2SStefano Zampini } 365b1b3d7a2SStefano Zampini #endif 3669162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3679162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3689162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3699162d606SStefano Zampini cum++; 37008122e43SStefano Zampini } 37108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 37208122e43SStefano Zampini /* shift for next computation */ 37308122e43SStefano Zampini cumarray += subset_size*subset_size; 37408122e43SStefano Zampini } 375fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 376fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 377fd14bc51SStefano Zampini } 37808122e43SStefano Zampini 37908122e43SStefano Zampini if (mss) { 38008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 38108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 382f6f667cfSStefano Zampini /* destroy matrices (junk) */ 383f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 384f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 38508122e43SStefano Zampini } 386f6f667cfSStefano Zampini if (allocated_S_St) { 387f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 388f6f667cfSStefano Zampini } 389f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 39008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 39108122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 39208122e43SStefano Zampini #endif 39308122e43SStefano Zampini if (pcbddc->dbg_flag) { 3941b968477SStefano Zampini PetscInt maxneigs_r; 39508122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3969b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 39708122e43SStefano Zampini } 39808122e43SStefano Zampini PetscFunctionReturn(0); 39908122e43SStefano Zampini } 400b1b3d7a2SStefano Zampini 401674ae819SStefano Zampini #undef __FUNCT__ 402c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 403c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 404c8587f34SStefano Zampini { 405c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4068629588bSStefano Zampini PetscScalar *coarse_submat_vals; 407c8587f34SStefano Zampini PetscErrorCode ierr; 408c8587f34SStefano Zampini 409c8587f34SStefano Zampini PetscFunctionBegin; 410f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 4115e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 412c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 413c8587f34SStefano Zampini 414684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 4150fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 416684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 417c8587f34SStefano Zampini 418c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 419b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 420c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 421c8587f34SStefano Zampini } 422c8587f34SStefano Zampini 4238629588bSStefano Zampini /* 4248629588bSStefano Zampini Setup local correction and local part of coarse basis. 4258629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 4268629588bSStefano Zampini */ 42747f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 4288629588bSStefano Zampini 4298629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 4308629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 4318629588bSStefano Zampini 4328629588bSStefano Zampini /* free */ 4338629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 434c8587f34SStefano Zampini PetscFunctionReturn(0); 435c8587f34SStefano Zampini } 436c8587f34SStefano Zampini 437c8587f34SStefano Zampini #undef __FUNCT__ 438674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 439674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 440674ae819SStefano Zampini { 441674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 442674ae819SStefano Zampini PetscErrorCode ierr; 443674ae819SStefano Zampini 444674ae819SStefano Zampini PetscFunctionBegin; 445674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 446674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 447674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 448674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 449785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 450674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 451f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 452f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 453785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 45463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 45563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 456674ae819SStefano Zampini PetscFunctionReturn(0); 457674ae819SStefano Zampini } 458674ae819SStefano Zampini 459674ae819SStefano Zampini #undef __FUNCT__ 460674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 461674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 462674ae819SStefano Zampini { 463674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 464674ae819SStefano Zampini PetscErrorCode ierr; 465674ae819SStefano Zampini 466674ae819SStefano Zampini PetscFunctionBegin; 467b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 468674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 469674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 470674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 471b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 472674ae819SStefano Zampini PetscFunctionReturn(0); 473674ae819SStefano Zampini } 474674ae819SStefano Zampini 475674ae819SStefano Zampini #undef __FUNCT__ 476674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 477674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 478674ae819SStefano Zampini { 479674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 48006656605SStefano Zampini PetscScalar *array; 481674ae819SStefano Zampini PetscErrorCode ierr; 482674ae819SStefano Zampini 483674ae819SStefano Zampini PetscFunctionBegin; 484674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 48558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 48606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 48706656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 48858da7f69SStefano Zampini } 489674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 490674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 49115aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 49215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 493674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 494674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 495674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 49606656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 497674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 498674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4998ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 500674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 501674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 502674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 503f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 504f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 505f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 506f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 507727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5080e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 509f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 51070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5116e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 512f5fd1fbfSStefano Zampini ierr = ISDestroy(&pcbddc->zerodiag);CHKERRQ(ierr); 51381d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 51481d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 515674ae819SStefano Zampini PetscFunctionReturn(0); 516674ae819SStefano Zampini } 517674ae819SStefano Zampini 518674ae819SStefano Zampini #undef __FUNCT__ 519f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 520f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 5216bfb1811SStefano Zampini { 5226bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5236bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 5246bfb1811SStefano Zampini VecType impVecType; 525e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 5266bfb1811SStefano Zampini PetscErrorCode ierr; 5276bfb1811SStefano Zampini 5286bfb1811SStefano Zampini PetscFunctionBegin; 529f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 530f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 531f4ddd8eeSStefano Zampini } 532e7b262bdSStefano Zampini /* get sizes */ 533b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 534b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 5356bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 536e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 537e7b262bdSStefano Zampini /* R nodes */ 538e7b262bdSStefano Zampini old_size = -1; 539e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 540e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 541e7b262bdSStefano Zampini } 542e7b262bdSStefano Zampini if (n_R != old_size) { 543e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 544e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5456bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5466bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5476bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5486bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 549e7b262bdSStefano Zampini } 550e7b262bdSStefano Zampini /* local primal dofs */ 551e7b262bdSStefano Zampini old_size = -1; 552e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 553e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 554e7b262bdSStefano Zampini } 555e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 556e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 55783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 558e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5596bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 560e7b262bdSStefano Zampini } 561e7b262bdSStefano Zampini /* local explicit constraints */ 562e7b262bdSStefano Zampini old_size = -1; 563e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 564e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 565e7b262bdSStefano Zampini } 566e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 567e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 56883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 56983b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 57083b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 57183b7ccabSStefano Zampini } 5726bfb1811SStefano Zampini PetscFunctionReturn(0); 5736bfb1811SStefano Zampini } 5746bfb1811SStefano Zampini 5756bfb1811SStefano Zampini #undef __FUNCT__ 57647f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 57747f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 57888ebb749SStefano Zampini { 57925084f0cSStefano Zampini PetscErrorCode ierr; 58025084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 58188ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 58288ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 583d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 58425084f0cSStefano Zampini /* submatrices of local problem */ 58580677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 58606656605SStefano Zampini /* submatrices of local coarse problem */ 58706656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 58825084f0cSStefano Zampini /* working matrices */ 58906656605SStefano Zampini Mat C_CR; 59025084f0cSStefano Zampini /* additional working stuff */ 59106656605SStefano Zampini PC pc_R; 59206656605SStefano Zampini Mat F; 59306656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 59406656605SStefano Zampini 59525084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 59606656605SStefano Zampini PetscScalar *work; 59706656605SStefano Zampini PetscInt *idx_V_B; 59806656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 59906656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 600b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 60145a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 60288ebb749SStefano Zampini MatType impMatType; 60325084f0cSStefano Zampini /* some shortcuts to scalars */ 60406656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 60588ebb749SStefano Zampini 60688ebb749SStefano Zampini PetscFunctionBegin; 607b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 60888ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 60988ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 610b371cd4fSStefano Zampini n_B = pcis->n_B; 611b371cd4fSStefano Zampini n_D = pcis->n - n_B; 61288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 61388ebb749SStefano Zampini 61488ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 61588ebb749SStefano Zampini impMatType = MATSEQDENSE; 61688ebb749SStefano Zampini 61788ebb749SStefano Zampini /* vertices in boundary numbering */ 618785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 6190e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 62088ebb749SStefano Zampini if (i != n_vertices) { 62122d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 62288ebb749SStefano Zampini } 62388ebb749SStefano Zampini 62406656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 62506656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 62606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 62706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 62806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 62906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 63006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 63106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 63206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 63306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 63406656605SStefano Zampini 63506656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 63606656605SStefano Zampini /* allocate workspace */ 63706656605SStefano Zampini n = 0; 63806656605SStefano Zampini if (n_constraints) { 63906656605SStefano Zampini n += n_R*n_constraints; 64006656605SStefano Zampini } 64106656605SStefano Zampini if (n_vertices) { 64206656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 64380677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 64406656605SStefano Zampini } 6453301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 64606656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 64706656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 64806656605SStefano Zampini } 64906656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 65006656605SStefano Zampini 65106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 65206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 65306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 65406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 65506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 65606656605SStefano Zampini if (isLU || isILU || isCHOL) { 65706656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 658d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 659d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 660d62866d3SStefano Zampini MatFactorType type; 661d62866d3SStefano Zampini 6626816873aSStefano Zampini F = reuse_mumps->F; 6636816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 664d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 66506656605SStefano Zampini } else { 66606656605SStefano Zampini F = NULL; 66706656605SStefano Zampini } 66806656605SStefano Zampini 66988ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 67088ebb749SStefano Zampini if (n_constraints) { 67106656605SStefano Zampini Mat M1,M2,M3; 67280677318SStefano Zampini Mat auxmat; 67306656605SStefano Zampini IS is_aux; 67480677318SStefano Zampini PetscScalar *array,*array2; 67506656605SStefano Zampini 676f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 67780677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 67888ebb749SStefano Zampini 67925084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 68025084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6818ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 68280677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 68388ebb749SStefano Zampini 68480677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 68580677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 68606656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 68788ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 68806656605SStefano Zampini const PetscScalar *row_cmat_values; 68906656605SStefano Zampini const PetscInt *row_cmat_indices; 69006656605SStefano Zampini PetscInt size_of_constraint,j; 69188ebb749SStefano Zampini 69206656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 69306656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 69406656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 69506656605SStefano Zampini } 69606656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 69706656605SStefano Zampini } 69880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 69906656605SStefano Zampini if (F) { 70006656605SStefano Zampini Mat B; 70106656605SStefano Zampini 70206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 70380677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 70406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 70506656605SStefano Zampini } else { 70680677318SStefano Zampini PetscScalar *marr; 70780677318SStefano Zampini 70880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 70906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 71006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 71180677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 71206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 71306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 71406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 71506656605SStefano Zampini } 71680677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 71706656605SStefano Zampini } 71880677318SStefano Zampini if (!pcbddc->switch_static) { 71980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 72080677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 72180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 72280677318SStefano Zampini for (i=0;i<n_constraints;i++) { 72380677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 72480677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 72580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 72680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 72780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 72880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 72980677318SStefano Zampini } 73080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 73180677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 73280677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 73380677318SStefano Zampini } else { 73480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 73580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 73625084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 73780677318SStefano Zampini } 73880677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 73980677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 74080677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 74106656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 74206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 74380677318SStefano Zampini if (isCHOL) { 74480677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 74580677318SStefano Zampini } else { 74625084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 74780677318SStefano Zampini } 74880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 74906656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 75025084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 75125084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 75225084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 75380677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 75480677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 75580677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 75606656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 75706656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 758f4ddd8eeSStefano Zampini } 75988ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 76088ebb749SStefano Zampini if (n_vertices) { 76106656605SStefano Zampini IS is_aux; 7623a50541eSStefano Zampini 7636816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7646816873aSStefano Zampini IS tis; 7656816873aSStefano Zampini 7666816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7676816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7686816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7696816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7706816873aSStefano Zampini } else { 7713a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7726816873aSStefano Zampini } 7739577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7749577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 77504708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 77625084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 77788ebb749SStefano Zampini } 77888ebb749SStefano Zampini 77988ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 780f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 78106656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 78206656605SStefano Zampini if (pcbddc->coarse_phi_D) { 78306656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 78406656605SStefano Zampini } 785f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 78606656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 78706656605SStefano Zampini PetscScalar *marray; 78806656605SStefano Zampini 78906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 79006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 791f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 792f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 793f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 794f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 795f4ddd8eeSStefano Zampini } 796f4ddd8eeSStefano Zampini } 79706656605SStefano Zampini 798f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 79906656605SStefano Zampini PetscScalar *marray; 80088ebb749SStefano Zampini 80106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8028eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 80306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 80488ebb749SStefano Zampini } 8053301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 80606656605SStefano Zampini n *= 2; 80788ebb749SStefano Zampini } 80806656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 80906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 81006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8118eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 81206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 81306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 81488ebb749SStefano Zampini } 8153301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 81606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 8178eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 81806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 81906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 82088ebb749SStefano Zampini } 82188ebb749SStefano Zampini } else { 822c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 823c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8241b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 825c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 826c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 827c0553b1fSStefano Zampini } 82888ebb749SStefano Zampini } 82906656605SStefano Zampini } 83006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 83106656605SStefano Zampini /* vertices */ 83206656605SStefano Zampini if (n_vertices) { 83316f15bc4SStefano Zampini 83404708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 83504708bb6SStefano Zampini 83616f15bc4SStefano Zampini if (n_R) { 83706656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 83806656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 83916f15bc4SStefano Zampini PetscScalar *x,*y; 84004708bb6SStefano Zampini PetscBool isseqaij; 84106656605SStefano Zampini 84221eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 84404708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 8456816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 84606656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 84706656605SStefano Zampini } else { 84806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 84906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 85206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 85406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 85506656605SStefano Zampini } 85606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 85706656605SStefano Zampini } 85880677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 85906656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 86006656605SStefano Zampini if (n_constraints) { 86106656605SStefano Zampini Mat B; 86280677318SStefano Zampini 863b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 86480677318SStefano Zampini for (i=0;i<n_vertices;i++) { 86580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 86680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 86780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 87080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87180677318SStefano Zampini } 87280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 87380677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 87480677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 87506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 87680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 87706656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 87806656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 87906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 88006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 88106656605SStefano Zampini } 88204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 88304708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 88404708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 88504708bb6SStefano Zampini } 88606656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 88780677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 88806656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 88906656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 89006656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 89106656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 89206656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 89406656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89506656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 89616f15bc4SStefano Zampini } else { 89716f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89816f15bc4SStefano Zampini } 89921eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 90006656605SStefano Zampini /* coarse basis functions */ 90106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 90216f15bc4SStefano Zampini PetscScalar *y; 90316f15bc4SStefano Zampini 90406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 91006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91206656605SStefano Zampini 91306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 91406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 92006656605SStefano Zampini } 92106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92206656605SStefano Zampini } 92304708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 92404708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 92506656605SStefano Zampini } 92606656605SStefano Zampini 92706656605SStefano Zampini if (n_constraints) { 92806656605SStefano Zampini Mat B; 92906656605SStefano Zampini 93006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 93106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 93280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 93306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 93406656605SStefano Zampini if (n_vertices) { 93580677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 93680677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 93780677318SStefano Zampini } else { 93880677318SStefano Zampini Mat S_VCt; 93980677318SStefano Zampini 94080677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 94180677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 94280677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 94380677318SStefano Zampini } 94406656605SStefano Zampini } 94506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94606656605SStefano Zampini /* coarse basis functions */ 94706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 94806656605SStefano Zampini PetscScalar *y; 94906656605SStefano Zampini 95006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 95206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 95306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 95706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 95806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 96006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 96306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 96406656605SStefano Zampini } 96506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96606656605SStefano Zampini } 96706656605SStefano Zampini } 96880677318SStefano Zampini if (n_constraints) { 96980677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 97080677318SStefano Zampini } 97106656605SStefano Zampini 97206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9733301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 97406656605SStefano Zampini 97506656605SStefano Zampini if (n_constraints) { 97616f15bc4SStefano Zampini Mat S_CCT,B_C; 97706656605SStefano Zampini 97880677318SStefano Zampini /* this is a lazy thing */ 97980677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 98006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 98316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 98406656605SStefano Zampini if (n_vertices) { 98516f15bc4SStefano Zampini Mat B_V,S_VCT; 98606656605SStefano Zampini 98706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 98806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 99116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 99206656605SStefano Zampini } 99306656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 99480677318SStefano Zampini } else { /* if there are no constraints, reset work */ 99580677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 99606656605SStefano Zampini } 99716f15bc4SStefano Zampini if (n_vertices && n_R) { 99806656605SStefano Zampini Mat A_VRT; 99980677318SStefano Zampini PetscScalar *marray; 100006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 100106656605SStefano Zampini 100280677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 100380677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 100480677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 100506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 100680677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 100780677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 100816f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 100906656605SStefano Zampini } 101006656605SStefano Zampini 101106656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 101206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 101306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 101406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 101506656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 101606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 101706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 101806656605SStefano Zampini } 101906656605SStefano Zampini } else { 102006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 102106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 102206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 102306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 102406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 102506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 102606656605SStefano Zampini } 102706656605SStefano Zampini } 102806656605SStefano Zampini /* coarse basis functions */ 102906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 103006656605SStefano Zampini PetscScalar *y; 103106656605SStefano Zampini 103206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 103306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 103406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 103506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 103606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 103706656605SStefano Zampini if (i<n_vertices) { 103806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 103906656605SStefano Zampini } 104006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 104106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 104206656605SStefano Zampini 104306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 104406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 104506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 104606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 104706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 104806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 104906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 105006656605SStefano Zampini } 105106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 105206656605SStefano Zampini } 105306656605SStefano Zampini } 1054d62866d3SStefano Zampini /* free memory */ 105588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 105606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 105706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 105806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 105906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1060d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1061d62866d3SStefano Zampini if (n_vertices) { 1062d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1063d62866d3SStefano Zampini } 1064d62866d3SStefano Zampini if (n_constraints) { 1065d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1066d62866d3SStefano Zampini } 106788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 106888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 106988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 107025084f0cSStefano Zampini if (pcbddc->dbg_flag) { 107188ebb749SStefano Zampini Mat coarse_sub_mat; 107225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 107388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 107488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 107588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10768bec7fa6SStefano Zampini Mat C_B,CPHI; 10778bec7fa6SStefano Zampini IS is_dummy; 10788bec7fa6SStefano Zampini Vec mones; 107988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 108088ebb749SStefano Zampini PetscReal real_value; 108188ebb749SStefano Zampini 108288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 108388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 108488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 108588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 108688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 108788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1088c0553b1fSStefano Zampini if (unsymmetric_check) { 108988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 109088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 109188ebb749SStefano Zampini } 109288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 109388ebb749SStefano Zampini 109425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10953301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 109625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1097c0553b1fSStefano Zampini if (unsymmetric_check) { 109888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 109988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 110088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 110188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 110288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 110388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 110488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 110888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 110988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 111088ebb749SStefano Zampini } else { 111188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 111288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 111388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 111488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 111588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 111688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 111788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 111888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 111988ebb749SStefano Zampini } 112088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 112188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 112288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 112388ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 112481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 11258bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 11260fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 112706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 11288bec7fa6SStefano Zampini 11298bec7fa6SStefano Zampini /* check constraints */ 11308bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 11318bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 11328bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 11338bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 11348bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 11358bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 11368bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1137bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1138c0553b1fSStefano Zampini if (unsymmetric_check) { 1139bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1140bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1141bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1142bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1143bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 114488ebb749SStefano Zampini } 11458bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 11468bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 11478bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 11488bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11498bec7fa6SStefano Zampini 115025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 115188ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 115288ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 115388ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 115488ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 115588ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 115688ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 115788ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 115888ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 115988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 116088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1161c0553b1fSStefano Zampini if (unsymmetric_check) { 116288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 116388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 116488ebb749SStefano Zampini } 116588ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 116688ebb749SStefano Zampini } 11678629588bSStefano Zampini /* get back data */ 11688629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 116988ebb749SStefano Zampini PetscFunctionReturn(0); 117088ebb749SStefano Zampini } 117188ebb749SStefano Zampini 117288ebb749SStefano Zampini #undef __FUNCT__ 1173d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1174d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1175aa0d41d4SStefano Zampini { 1176d65f70fdSStefano Zampini Mat *work_mat; 1177d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1178d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1179d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1180aa0d41d4SStefano Zampini PetscErrorCode ierr; 1181aa0d41d4SStefano Zampini 1182aa0d41d4SStefano Zampini PetscFunctionBegin; 1183d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1184d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1185d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1187aa0d41d4SStefano Zampini 1188d65f70fdSStefano Zampini if (!rsorted) { 1189906d46d4SStefano Zampini const PetscInt *idxs; 1190906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1191aa0d41d4SStefano Zampini 1192d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1193d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1194d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1195d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1196aa0d41d4SStefano Zampini } 1197d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1198d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1200d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1201aa0d41d4SStefano Zampini } 1202d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini } else { 1205d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1206d65f70fdSStefano Zampini isrow_s = isrow; 1207aa0d41d4SStefano Zampini } 1208906d46d4SStefano Zampini 1209d65f70fdSStefano Zampini if (!csorted) { 1210d65f70fdSStefano Zampini if (isrow == iscol) { 1211d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini iscol_s = isrow_s; 1213d65f70fdSStefano Zampini } else { 1214d65f70fdSStefano Zampini const PetscInt *idxs; 1215d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1216906d46d4SStefano Zampini 1217d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1219d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1220d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1221d65f70fdSStefano Zampini } 1222d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1223d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1225d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1226d65f70fdSStefano Zampini } 1227d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1228d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1229d65f70fdSStefano Zampini } 1230d65f70fdSStefano Zampini } else { 1231d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1232d65f70fdSStefano Zampini iscol_s = iscol; 1233d65f70fdSStefano Zampini } 1234d65f70fdSStefano Zampini 1235d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1236d65f70fdSStefano Zampini 1237d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1238906d46d4SStefano Zampini Mat new_mat; 1239d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1240906d46d4SStefano Zampini 1241d65f70fdSStefano Zampini if (!rsorted) { 1242d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1243d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1244d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1245d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1246906d46d4SStefano Zampini } 1247d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1248d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1249d65f70fdSStefano Zampini } else { 1250d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1251906d46d4SStefano Zampini } 1252d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1253d65f70fdSStefano Zampini 1254d65f70fdSStefano Zampini if (!csorted) { 1255d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1256d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1257d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1258d65f70fdSStefano Zampini } else { 1259d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1260d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1261d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1262d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1263d65f70fdSStefano Zampini } 1264d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1265d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1266d65f70fdSStefano Zampini } 1267d65f70fdSStefano Zampini } else { 1268d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1269d65f70fdSStefano Zampini } 1270d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1271d65f70fdSStefano Zampini 1272d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1273d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1274d65f70fdSStefano Zampini work_mat[0] = new_mat; 1275d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1276d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1277d65f70fdSStefano Zampini } 1278d65f70fdSStefano Zampini 1279d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1280d65f70fdSStefano Zampini *B = work_mat[0]; 1281d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1282d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1283d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1284d65f70fdSStefano Zampini PetscFunctionReturn(0); 1285d65f70fdSStefano Zampini } 1286d65f70fdSStefano Zampini 1287d65f70fdSStefano Zampini #undef __FUNCT__ 12885e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12895e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1290aa0d41d4SStefano Zampini { 1291aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12923bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 12935e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1294d65f70fdSStefano Zampini Mat new_mat; 12955e8657edSStefano Zampini IS is_local,is_global; 1296d65f70fdSStefano Zampini PetscInt local_size; 1297d65f70fdSStefano Zampini PetscBool isseqaij; 1298aa0d41d4SStefano Zampini PetscErrorCode ierr; 1299aa0d41d4SStefano Zampini 1300aa0d41d4SStefano Zampini PetscFunctionBegin; 1301aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 13025e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 13035e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 13043bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1305aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1306d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1307aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1308906d46d4SStefano Zampini 1309906d46d4SStefano Zampini /* check */ 1310906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1311906d46d4SStefano Zampini Vec x,x_change; 1312906d46d4SStefano Zampini PetscReal error; 1313906d46d4SStefano Zampini 13145e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1315906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 13165e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1317e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1318e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1319d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1320e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1321e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1322906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1323906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1324906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1325906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1326906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1327906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1328906d46d4SStefano Zampini } 1329906d46d4SStefano Zampini 133022d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 13319b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 133222d5777bSStefano Zampini if (isseqaij) { 1333d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1334aa0d41d4SStefano Zampini } else { 1335aa0d41d4SStefano Zampini Mat work_mat; 1336aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1337d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1338aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1339aa0d41d4SStefano Zampini } 13403301b35fSStefano Zampini if (matis->A->symmetric_set) { 13413301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1342e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 13433301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1344e496cd5dSStefano Zampini #endif 13453301b35fSStefano Zampini } 134645a1bb75SStefano Zampini /* 134745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1348d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 134945a1bb75SStefano Zampini */ 1350d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1351aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1352aa0d41d4SStefano Zampini } 1353aa0d41d4SStefano Zampini 1354aa0d41d4SStefano Zampini #undef __FUNCT__ 1355a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13568ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1357a64d13efSStefano Zampini { 1358a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1359a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1360d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 136153892102SStefano Zampini PetscInt *idx_R_local=NULL; 13623a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13633a50541eSStefano Zampini PetscInt vbs,bs; 13646816873aSStefano Zampini PetscBT bitmask=NULL; 1365a64d13efSStefano Zampini PetscErrorCode ierr; 1366a64d13efSStefano Zampini 1367a64d13efSStefano Zampini PetscFunctionBegin; 1368b23d619eSStefano Zampini /* 1369b23d619eSStefano Zampini No need to setup local scatters if 1370b23d619eSStefano Zampini - primal space is unchanged 1371b23d619eSStefano Zampini AND 1372b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1373b23d619eSStefano Zampini AND 1374b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1375b23d619eSStefano Zampini */ 1376b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1377f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1378f4ddd8eeSStefano Zampini } 1379f4ddd8eeSStefano Zampini /* destroy old objects */ 1380f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1381f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1382f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1383a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1384b371cd4fSStefano Zampini n_B = pcis->n_B; 1385b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1386b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13873a50541eSStefano Zampini 1388a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13896816873aSStefano Zampini 139053892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13916816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1392854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1393a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1394a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 13950e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1396a64d13efSStefano Zampini } 1397a64d13efSStefano Zampini 1398a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13994641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 14006816873aSStefano Zampini idx_R_local[n_R++] = i; 1401a64d13efSStefano Zampini } 1402a64d13efSStefano Zampini } 140353892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 14046816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14056816873aSStefano Zampini 140653892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 140753892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 14086816873aSStefano Zampini } 14093a50541eSStefano Zampini 14103a50541eSStefano Zampini /* Block code */ 14113a50541eSStefano Zampini vbs = 1; 14123a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 14133a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 14143a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 14153a50541eSStefano Zampini PetscInt *vary; 1416d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1417785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 14183a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1419d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1420d3df7717SStefano 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 */ 14210e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1422d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 14233a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 14243a50541eSStefano Zampini is_blocked = PETSC_FALSE; 14253a50541eSStefano Zampini break; 14263a50541eSStefano Zampini } 14273a50541eSStefano Zampini } 1428d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1429d3df7717SStefano Zampini } else { 1430d3df7717SStefano Zampini /* Verify directly the R set */ 1431d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1432d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1433d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1434d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1435d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1436d3df7717SStefano Zampini break; 1437d3df7717SStefano Zampini } 1438d3df7717SStefano Zampini } 1439d3df7717SStefano Zampini } 1440d3df7717SStefano Zampini } 14413a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 14423a50541eSStefano Zampini vbs = bs; 14433a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 14443a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 14453a50541eSStefano Zampini } 14463a50541eSStefano Zampini } 14473a50541eSStefano Zampini } 14483a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 144953892102SStefano Zampini if (sub_schurs->reuse_mumps) { 145053892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 145153892102SStefano Zampini 145253892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 145353892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 145453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 145553892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 145653892102SStefano Zampini } else { 14573a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 145853892102SStefano Zampini } 1459a64d13efSStefano Zampini 1460a64d13efSStefano Zampini /* print some info if requested */ 1461a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1462a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1463a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14640fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1465a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1466a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14673a50541eSStefano 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->local_primal_size);CHKERRQ(ierr); 1468a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1469a64d13efSStefano Zampini } 1470a64d13efSStefano Zampini 1471a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14726816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14736816873aSStefano Zampini IS is_aux1,is_aux2; 14746816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14756816873aSStefano Zampini 14763a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1477854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1478854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1479a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14804641a718SStefano Zampini for (i=0; i<n_D; i++) { 14814641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14824641a718SStefano Zampini } 1483a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1484a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14854641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14864641a718SStefano Zampini aux_array1[j++] = i; 1487a64d13efSStefano Zampini } 1488a64d13efSStefano Zampini } 1489a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1490a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1491a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14924641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14934641a718SStefano Zampini aux_array2[j++] = i; 1494a64d13efSStefano Zampini } 1495a64d13efSStefano Zampini } 1496a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1497a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1498a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1499a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1500a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1501a64d13efSStefano Zampini 15028eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1503785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1504a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 15054641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 15064641a718SStefano Zampini aux_array1[j++] = i; 1507a64d13efSStefano Zampini } 1508a64d13efSStefano Zampini } 1509a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1510a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1511a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1512a64d13efSStefano Zampini } 15134641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 15143a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1515d62866d3SStefano Zampini } else { 151653892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15176816873aSStefano Zampini IS tis; 15186816873aSStefano Zampini PetscInt schur_size; 15196816873aSStefano Zampini 152053892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 15216816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 152253892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 15236816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 15246816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 15256816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 15266816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 15276816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1528d62866d3SStefano Zampini } 1529d62866d3SStefano Zampini } 1530a64d13efSStefano Zampini PetscFunctionReturn(0); 1531a64d13efSStefano Zampini } 1532a64d13efSStefano Zampini 1533304d26faSStefano Zampini 1534304d26faSStefano Zampini #undef __FUNCT__ 1535304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1536684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1537304d26faSStefano Zampini { 1538304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1539304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1540304d26faSStefano Zampini PC pc_temp; 1541304d26faSStefano Zampini Mat A_RR; 1542f4ddd8eeSStefano Zampini MatReuse reuse; 1543304d26faSStefano Zampini PetscScalar m_one = -1.0; 1544304d26faSStefano Zampini PetscReal value; 154504708bb6SStefano Zampini PetscInt n_D,n_R; 15469577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1547304d26faSStefano Zampini PetscErrorCode ierr; 1548e604994aSStefano Zampini /* prefixes stuff */ 1549312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1550e604994aSStefano Zampini size_t len; 1551304d26faSStefano Zampini 1552304d26faSStefano Zampini PetscFunctionBegin; 1553304d26faSStefano Zampini 1554e604994aSStefano Zampini /* compute prefixes */ 1555e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1556e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1557e604994aSStefano Zampini if (!pcbddc->current_level) { 1558e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1559e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1560e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1561e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1562e604994aSStefano Zampini } else { 1563e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1564312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1565e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1566e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1567312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1568312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 156934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 157034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1571e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1572e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1573e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1574e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1575e604994aSStefano Zampini } 1576e604994aSStefano Zampini 1577304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1578684f6988SStefano Zampini if (dirichlet) { 1579d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15803301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15813301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1582964fefecSStefano Zampini } 1583ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1584964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1585304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1586304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1587304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1588304d26faSStefano Zampini /* default */ 1589304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1590e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15919577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1592304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15939577ea80SStefano Zampini if (issbaij) { 15949577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15959577ea80SStefano Zampini } else { 1596304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15979577ea80SStefano Zampini } 1598304d26faSStefano Zampini /* Allow user's customization */ 1599304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1600304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1601304d26faSStefano Zampini } 1602d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1603d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1604d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1605d62866d3SStefano Zampini 1606d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1607d5574798SStefano Zampini } 1608304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1609304d26faSStefano Zampini if (!n_D) { 1610304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1611304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1612304d26faSStefano Zampini } 1613304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1614304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1615304d26faSStefano Zampini /* set ksp_D into pcis data */ 1616304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1617304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1618304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1619684f6988SStefano Zampini } 1620304d26faSStefano Zampini 1621304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1622684f6988SStefano Zampini A_RR = 0; 1623684f6988SStefano Zampini if (neumann) { 1624d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 162504708bb6SStefano Zampini PetscInt ibs,mbs; 162604708bb6SStefano Zampini PetscBool issbaij; 162704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1628f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 16298ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1630f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1631f4ddd8eeSStefano Zampini PetscInt nn_R; 163281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1633f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1634f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1635f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1636f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1637f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1638f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1639f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1640727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1641f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1642f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1643f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1644f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1645f4ddd8eeSStefano Zampini } 1646f4ddd8eeSStefano Zampini } 1647f4ddd8eeSStefano Zampini /* last check */ 1648d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1649f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1650f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1651f4ddd8eeSStefano Zampini } 1652f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1653f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1654f4ddd8eeSStefano Zampini } 1655f4ddd8eeSStefano Zampini /* extract A_RR */ 1656af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1657af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 165804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 165904708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 166004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 166104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 166204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1663af732b37SStefano Zampini } else { 166404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16656816873aSStefano Zampini } 166604708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 166704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 166804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 166904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 167004708bb6SStefano Zampini } else { 167104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 167204708bb6SStefano Zampini } 167304708bb6SStefano Zampini } 167404708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1675f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16763301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16773301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16786816873aSStefano Zampini } 16796816873aSStefano Zampini } else { 16806816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16816816873aSStefano Zampini 16826816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16836816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16846816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1685af732b37SStefano Zampini } 1686f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1687304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1688304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1689304d26faSStefano Zampini /* default */ 1690304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1691e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1692304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16939577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16949577ea80SStefano Zampini if (issbaij) { 16959577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16969577ea80SStefano Zampini } else { 1697304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16989577ea80SStefano Zampini } 1699304d26faSStefano Zampini /* Allow user's customization */ 1700304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1701304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1702304d26faSStefano Zampini } 1703d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1704304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1705304d26faSStefano Zampini if (!n_R) { 1706304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1707304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1708304d26faSStefano Zampini } 1709d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1710d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1711d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1712d62866d3SStefano Zampini 1713d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1714d62866d3SStefano Zampini } 1715304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1716304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1717684f6988SStefano Zampini } 17186816873aSStefano Zampini /* free Neumann problem's matrix */ 17196816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1720304d26faSStefano Zampini 1721304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 17220fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1723684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1724684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1725684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1726684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1727684f6988SStefano Zampini } 1728684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 17290fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 17300fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 17310fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 17320fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 17330fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1734304d26faSStefano Zampini /* need to be adapted? */ 1735b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1736b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1737b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1738304d26faSStefano Zampini /* print info */ 1739304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1740e604994aSStefano 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); 1741304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1742304d26faSStefano Zampini } 1743b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1744298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1745304d26faSStefano Zampini } 1746684f6988SStefano Zampini } 1747684f6988SStefano Zampini if (neumann) { /* Neumann */ 17486816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 17490fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 17500fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 17510fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 17520fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 17530fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1754304d26faSStefano Zampini /* need to be adapted? */ 1755b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1756b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1757304d26faSStefano Zampini /* print info */ 1758304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1759e604994aSStefano 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); 1760304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1761304d26faSStefano Zampini } 1762b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1763298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1764304d26faSStefano Zampini } 17650fccc4e9SStefano Zampini } 1766684f6988SStefano Zampini } 1767304d26faSStefano Zampini PetscFunctionReturn(0); 1768304d26faSStefano Zampini } 1769304d26faSStefano Zampini 1770304d26faSStefano Zampini #undef __FUNCT__ 1771ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 177280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1773674ae819SStefano Zampini { 1774674ae819SStefano Zampini PetscErrorCode ierr; 1775674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1776be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1777674ae819SStefano Zampini 1778674ae819SStefano Zampini PetscFunctionBegin; 1779be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 178080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 178120c7b377SStefano Zampini } 178280677318SStefano Zampini if (!pcbddc->switch_static) { 178380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 178480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 178620c7b377SStefano Zampini } 1787be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 178880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 179020c7b377SStefano Zampini } else { 1791be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1792be83ff47SStefano Zampini 179353892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179453892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179520c7b377SStefano Zampini } 1796be83ff47SStefano Zampini } else { 179780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 179880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 179980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 180080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 180180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 180280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 180380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 180480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 180580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1806674ae819SStefano Zampini } 1807674ae819SStefano Zampini } 1808be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 180980677318SStefano Zampini if (applytranspose) { 181080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 181180677318SStefano Zampini } else { 181280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 181380677318SStefano Zampini } 181453892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1815be83ff47SStefano Zampini } else { 1816be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1817be83ff47SStefano Zampini 1818be83ff47SStefano Zampini if (applytranspose) { 181953892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1820be83ff47SStefano Zampini } else { 182153892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1822be83ff47SStefano Zampini } 182353892102SStefano Zampini #endif 1824be83ff47SStefano Zampini } 182580677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 182680677318SStefano Zampini if (!pcbddc->switch_static) { 1827be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 182880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1830be83ff47SStefano Zampini } else { 1831be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1832be83ff47SStefano Zampini 183353892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 183453892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1835be83ff47SStefano Zampini } 183680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 183780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 183880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 183980677318SStefano Zampini } 184080677318SStefano Zampini } else { 184180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 184280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 184380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 184480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 184580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 184680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 184780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 184880677318SStefano Zampini } 184980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 185080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 185180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 185280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1853674ae819SStefano Zampini } 1854674ae819SStefano Zampini PetscFunctionReturn(0); 1855674ae819SStefano Zampini } 1856674ae819SStefano Zampini 1857dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1858674ae819SStefano Zampini #undef __FUNCT__ 1859674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1860dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1861674ae819SStefano Zampini { 1862674ae819SStefano Zampini PetscErrorCode ierr; 1863674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1864674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1865674ae819SStefano Zampini const PetscScalar zero = 0.0; 1866674ae819SStefano Zampini 1867674ae819SStefano Zampini PetscFunctionBegin; 1868dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1869dc359a40SStefano Zampini if (applytranspose) { 1870674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18718eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1872dc359a40SStefano Zampini } else { 1873674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1874674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 187515aaf578SStefano Zampini } 187612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 187712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 187812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 187912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 188012edc857SStefano Zampini 18819f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 188212edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 188312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1884964fefecSStefano Zampini Vec rhs,sol; 1885964fefecSStefano Zampini 1886964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1887964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 188812edc857SStefano Zampini if (applytranspose) { 1889964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 189012edc857SStefano Zampini } else { 1891964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 189212edc857SStefano Zampini } 189312edc857SStefano Zampini } 1894674ae819SStefano Zampini 1895674ae819SStefano Zampini /* Local solution on R nodes */ 189680677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 189780677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18989f00e9b4SStefano Zampini } 1899674ae819SStefano Zampini 19009f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 19019f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 190212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1903674ae819SStefano Zampini 1904674ae819SStefano Zampini /* Sum contributions from two levels */ 1905dc359a40SStefano Zampini if (applytranspose) { 1906dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1907dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1908dc359a40SStefano Zampini } else { 1909674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 19108eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1911dc359a40SStefano Zampini } 1912674ae819SStefano Zampini PetscFunctionReturn(0); 1913674ae819SStefano Zampini } 1914674ae819SStefano Zampini 1915674ae819SStefano Zampini #undef __FUNCT__ 1916674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 191712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1918674ae819SStefano Zampini { 1919674ae819SStefano Zampini PetscErrorCode ierr; 1920674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 192158da7f69SStefano Zampini PetscScalar *array; 192212edc857SStefano Zampini Vec from,to; 1923674ae819SStefano Zampini 1924674ae819SStefano Zampini PetscFunctionBegin; 192512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 192612edc857SStefano Zampini from = pcbddc->coarse_vec; 192712edc857SStefano Zampini to = pcbddc->vec1_P; 192812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 192912edc857SStefano Zampini Vec tvec; 193058da7f69SStefano Zampini 193158da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 193258da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 193312edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 193458da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 193558da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 193658da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 193712edc857SStefano Zampini } 193812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 193912edc857SStefano Zampini from = pcbddc->vec1_P; 194012edc857SStefano Zampini to = pcbddc->coarse_vec; 194112edc857SStefano Zampini } 194212edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1943674ae819SStefano Zampini PetscFunctionReturn(0); 1944674ae819SStefano Zampini } 1945674ae819SStefano Zampini 1946674ae819SStefano Zampini #undef __FUNCT__ 1947674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 194812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1949674ae819SStefano Zampini { 1950674ae819SStefano Zampini PetscErrorCode ierr; 1951674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 195258da7f69SStefano Zampini PetscScalar *array; 195312edc857SStefano Zampini Vec from,to; 1954674ae819SStefano Zampini 1955674ae819SStefano Zampini PetscFunctionBegin; 195612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 195712edc857SStefano Zampini from = pcbddc->coarse_vec; 195812edc857SStefano Zampini to = pcbddc->vec1_P; 195912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 196012edc857SStefano Zampini from = pcbddc->vec1_P; 196112edc857SStefano Zampini to = pcbddc->coarse_vec; 196212edc857SStefano Zampini } 196312edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 196412edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 196512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 196612edc857SStefano Zampini Vec tvec; 196758da7f69SStefano Zampini 196812edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 196958da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 197058da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 197158da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 197258da7f69SStefano Zampini } 197358da7f69SStefano Zampini } else { 197458da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 197558da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 197612edc857SStefano Zampini } 197712edc857SStefano Zampini } 1978674ae819SStefano Zampini PetscFunctionReturn(0); 1979674ae819SStefano Zampini } 1980674ae819SStefano Zampini 1981984c4197SStefano Zampini /* uncomment for testing purposes */ 1982984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1983674ae819SStefano Zampini #undef __FUNCT__ 1984674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1985674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1986674ae819SStefano Zampini { 1987674ae819SStefano Zampini PetscErrorCode ierr; 1988674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1989674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1990674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1991984c4197SStefano Zampini /* one and zero */ 1992984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1993984c4197SStefano Zampini /* space to store constraints and their local indices */ 19949162d606SStefano Zampini PetscScalar *constraints_data; 19959162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19969162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19979162d606SStefano Zampini PetscInt *constraints_n; 1998984c4197SStefano Zampini /* iterators */ 1999b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2000984c4197SStefano Zampini /* BLAS integers */ 2001e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2002e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2003c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2004727cdba6SStefano Zampini /* reuse */ 20050e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 20060e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2007984c4197SStefano Zampini /* change of basis */ 2008b3d85658SStefano Zampini PetscBool qr_needed; 20099162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2010984c4197SStefano Zampini /* auxiliary stuff */ 201164efe560SStefano Zampini PetscInt *nnz,*is_indices; 20128a0068c3SStefano Zampini PetscInt ncc; 2013984c4197SStefano Zampini /* some quantities */ 201445a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2015a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2016984c4197SStefano Zampini 2017674ae819SStefano Zampini PetscFunctionBegin; 20188e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 20198e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 20208e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2021088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2022088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 20230e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 20240e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 20250e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 20260e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 20270e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2028088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2029cf5a6209SStefano Zampini 2030cf5a6209SStefano Zampini /* print some info */ 2031cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2032cf5a6209SStefano Zampini IS vertices; 2033cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2034cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2035cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2036cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2037cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2038cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2039cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2040fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2041fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2042cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2043cf5a6209SStefano Zampini } 2044cf5a6209SStefano Zampini 2045cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 20469162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2047cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2048cf5a6209SStefano Zampini const Vec *nearnullvecs; 2049cf5a6209SStefano Zampini Vec *localnearnullsp; 2050cf5a6209SStefano Zampini PetscScalar *array; 2051cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2052cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2053674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2054b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2055674ae819SStefano Zampini PetscScalar *work; 2056674ae819SStefano Zampini PetscReal *singular_vals; 2057674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2058674ae819SStefano Zampini PetscReal *rwork; 2059674ae819SStefano Zampini #endif 2060674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2061674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2062674ae819SStefano Zampini #else 2063964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2064964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2065674ae819SStefano Zampini #endif 2066674ae819SStefano Zampini 2067674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2068d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2069d06fc5fdSStefano Zampini /* free unneeded index sets */ 2070d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2071d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2072674ae819SStefano Zampini } 2073d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2074d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2075d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2076d06fc5fdSStefano Zampini } 2077d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2078d06fc5fdSStefano Zampini n_ISForEdges = 0; 2079d06fc5fdSStefano Zampini } 2080d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2081d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2082d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2083d06fc5fdSStefano Zampini } 2084d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2085d06fc5fdSStefano Zampini n_ISForFaces = 0; 2086d06fc5fdSStefano Zampini } 208770022509SStefano Zampini 208870022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 208970022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 209070022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 209170022509SStefano Zampini if (pcbddc->NullSpace) { 209270022509SStefano Zampini PetscBool tbool[2],gbool[2]; 209370022509SStefano Zampini 209470022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2095b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2096d06fc5fdSStefano Zampini if (!ISForEdges) { 2097d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2098d06fc5fdSStefano Zampini } 2099b8ffe317SStefano Zampini } 2100d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2101d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2102d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2103d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2104d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 210598a51de6SStefano Zampini } 210670022509SStefano Zampini #endif 210708122e43SStefano Zampini 2108674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2109674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2110674ae819SStefano Zampini if (nearnullsp) { 2111674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2112f4ddd8eeSStefano Zampini /* remove any stored info */ 2113f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2114f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2115f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2116f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2117f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2118473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2119f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2120f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2121f4ddd8eeSStefano Zampini } 2122984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2123984c4197SStefano Zampini nnsp_size = 0; 2124674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2125674ae819SStefano Zampini } 2126984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2127984c4197SStefano Zampini max_constraints = nnsp_size; 2128984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2129984c4197SStefano Zampini 2130674ae819SStefano Zampini /* 2131674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 21329162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 21339162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 21349162d606SStefano Zampini There can be multiple constraints per connected component 2135674ae819SStefano Zampini */ 2136674ae819SStefano Zampini n_vertices = 0; 2137674ae819SStefano Zampini if (ISForVertices) { 2138674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2139674ae819SStefano Zampini } 21409162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 21419162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 21429162d606SStefano Zampini 21439162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 21449162d606SStefano Zampini total_counts *= max_constraints; 2145674ae819SStefano Zampini total_counts += n_vertices; 21464641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 21479162d606SStefano Zampini 2148674ae819SStefano Zampini total_counts = 0; 2149674ae819SStefano Zampini max_size_of_constraint = 0; 2150674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 21519162d606SStefano Zampini IS used_is; 2152674ae819SStefano Zampini if (i<n_ISForEdges) { 21539162d606SStefano Zampini used_is = ISForEdges[i]; 2154674ae819SStefano Zampini } else { 21559162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2156674ae819SStefano Zampini } 21579162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2158674ae819SStefano Zampini total_counts += j; 2159674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2160674ae819SStefano Zampini } 21619162d606SStefano 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); 21629162d606SStefano Zampini 2163984c4197SStefano Zampini /* get local part of global near null space vectors */ 2164785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2165984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2166984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2167e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2168e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2169984c4197SStefano Zampini } 2170674ae819SStefano Zampini 2171242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2172242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2173a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2174242a89d7SStefano Zampini 2175984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2176a773dcb8SStefano Zampini if (!skip_lapack) { 2177674ae819SStefano Zampini PetscScalar temp_work; 2178911cabfeSStefano Zampini 2179674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2180984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2181785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2182785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2183785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2184674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2185785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2186674ae819SStefano Zampini #endif 2187674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2188c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2189c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2190674ae819SStefano Zampini lwork = -1; 2191674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2192674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2193c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2194674ae819SStefano Zampini #else 2195c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2196674ae819SStefano Zampini #endif 2197674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2198984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2199674ae819SStefano Zampini #else /* on missing GESVD */ 2200674ae819SStefano Zampini /* SVD */ 2201674ae819SStefano Zampini PetscInt max_n,min_n; 2202674ae819SStefano Zampini max_n = max_size_of_constraint; 2203984c4197SStefano Zampini min_n = max_constraints; 2204984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2205674ae819SStefano Zampini min_n = max_size_of_constraint; 2206984c4197SStefano Zampini max_n = max_constraints; 2207674ae819SStefano Zampini } 2208785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2209674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2210785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2211674ae819SStefano Zampini #endif 2212674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2213674ae819SStefano Zampini lwork = -1; 2214e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2215e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2216b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2217674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2218674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 22199162d606SStefano 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)); 2220674ae819SStefano Zampini #else 22219162d606SStefano 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)); 2222674ae819SStefano Zampini #endif 2223674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2224984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2225984c4197SStefano Zampini #endif /* on missing GESVD */ 2226674ae819SStefano Zampini /* Allocate optimal workspace */ 2227674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2228854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2229674ae819SStefano Zampini } 2230674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2231674ae819SStefano Zampini total_counts = 0; 22329162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 22339162d606SStefano Zampini constraints_data_ptr[0] = 0; 2234674ae819SStefano Zampini /* vertices */ 22359162d606SStefano Zampini if (n_vertices) { 2236674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 22379162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 22389162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2239674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 22409162d606SStefano Zampini constraints_n[total_counts] = 1; 22419162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22429162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22439162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2244674ae819SStefano Zampini total_counts++; 2245674ae819SStefano Zampini } 2246674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2247984c4197SStefano Zampini PetscBool used_vertex; 2248674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2249674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2250674ae819SStefano Zampini k = 0; 2251674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2252984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2253984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 22549162d606SStefano Zampini constraints_n[total_counts] = 1; 22559162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 22569162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22579162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22589162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2259674ae819SStefano Zampini total_counts++; 2260674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2261674ae819SStefano Zampini } 2262984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2263674ae819SStefano Zampini k++; 2264674ae819SStefano Zampini } 2265674ae819SStefano Zampini } 2266674ae819SStefano Zampini } 2267674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2268674ae819SStefano Zampini n_vertices = total_counts; 2269674ae819SStefano Zampini } 2270984c4197SStefano Zampini 2271674ae819SStefano Zampini /* edges and faces */ 22729162d606SStefano Zampini total_counts_cc = total_counts; 2273911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22749162d606SStefano Zampini IS used_is; 22759162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22769162d606SStefano Zampini 2277911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22789162d606SStefano Zampini used_is = ISForEdges[ncc]; 2279984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2280674ae819SStefano Zampini } else { 22819162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2282984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2283674ae819SStefano Zampini } 2284674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22859162d606SStefano Zampini 22869162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22879162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2288984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2289984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2290674ae819SStefano Zampini if (nnsp_has_cnst) { 22915b08dc53SStefano Zampini PetscScalar quad_value; 22929162d606SStefano Zampini 22939162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22949162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22959162d606SStefano Zampini 2296a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2297674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2298a773dcb8SStefano Zampini } else { 2299a773dcb8SStefano Zampini quad_value = 1.0; 2300a773dcb8SStefano Zampini } 2301674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 23029162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2303674ae819SStefano Zampini } 23049162d606SStefano Zampini temp_constraints++; 2305674ae819SStefano Zampini total_counts++; 2306674ae819SStefano Zampini } 2307674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2308984c4197SStefano Zampini PetscReal real_value; 23099162d606SStefano Zampini PetscScalar *ptr_to_data; 23109162d606SStefano Zampini 2311984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 23129162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2313674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 23149162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2315674ae819SStefano Zampini } 2316984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2317984c4197SStefano Zampini /* check if array is null on the connected component */ 2318e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 23199162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 23205b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2321674ae819SStefano Zampini temp_constraints++; 2322674ae819SStefano Zampini total_counts++; 23239162d606SStefano Zampini if (!idxs_copied) { 23249162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 23259162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2326674ae819SStefano Zampini } 2327674ae819SStefano Zampini } 23289162d606SStefano Zampini } 23299162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 233045a1bb75SStefano Zampini valid_constraints = temp_constraints; 2331eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2332a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 23339162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 23349162d606SStefano Zampini 23359162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2336a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 23379162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2338a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 23399162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2340a773dcb8SStefano Zampini } else { /* perform SVD */ 2341984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 23429162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2343674ae819SStefano Zampini 2344674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2345984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2346984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2347984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2348984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2349984c4197SStefano Zampini from that computed using LAPACKgesvd 2350984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2351984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2352984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2353674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2354e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2355984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2356674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2357674ae819SStefano Zampini for (k=0;k<j+1;k++) { 23589162d606SStefano 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)); 2359674ae819SStefano Zampini } 2360674ae819SStefano Zampini } 2361e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2362e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2363e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2364674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2365c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2366674ae819SStefano Zampini #else 2367c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2368674ae819SStefano Zampini #endif 2369674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2370984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2371984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2372674ae819SStefano Zampini j = 0; 2373984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2374674ae819SStefano Zampini total_counts = total_counts-j; 237545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2376e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2377c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2378c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2379c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2380c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2381c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2382c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2383674ae819SStefano Zampini if (j<temp_constraints) { 2384984c4197SStefano Zampini PetscInt ii; 2385984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2386674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23879162d606SStefano 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)); 2388674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2389984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2390674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23919162d606SStefano 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]; 2392674ae819SStefano Zampini } 2393674ae819SStefano Zampini } 2394674ae819SStefano Zampini } 2395674ae819SStefano Zampini #else /* on missing GESVD */ 2396e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2397e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2398b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2399674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2400674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24019162d606SStefano 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)); 2402674ae819SStefano Zampini #else 24039162d606SStefano 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)); 2404674ae819SStefano Zampini #endif 2405984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2406674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2407984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2408e310c8b4SStefano Zampini k = temp_constraints; 2409e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2410674ae819SStefano Zampini j = 0; 2411e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 241245a1bb75SStefano Zampini valid_constraints = k-j; 2413911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2414984c4197SStefano Zampini #endif /* on missing GESVD */ 2415674ae819SStefano Zampini } 2416a773dcb8SStefano Zampini } 24179162d606SStefano Zampini /* update pointers information */ 24189162d606SStefano Zampini if (valid_constraints) { 24199162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 24209162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 24219162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 24229162d606SStefano Zampini /* set change_of_basis flag */ 242345a1bb75SStefano Zampini if (boolforchange) { 2424b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 24259162d606SStefano Zampini } 2426b3d85658SStefano Zampini total_counts_cc++; 242745a1bb75SStefano Zampini } 242845a1bb75SStefano Zampini } 2429984c4197SStefano Zampini /* free workspace */ 24308f1c130eSStefano Zampini if (!skip_lapack) { 2431984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2432984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2433984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2434984c4197SStefano Zampini #endif 2435984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2436984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2437984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2438984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2439984c4197SStefano Zampini #endif 2440984c4197SStefano Zampini } 2441984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2442984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2443984c4197SStefano Zampini } 2444984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2445cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2446cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2447cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2448cf5a6209SStefano Zampini } 2449cf5a6209SStefano Zampini if (n_ISForFaces) { 2450cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2451cf5a6209SStefano Zampini } 2452cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2453cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2454cf5a6209SStefano Zampini } 2455cf5a6209SStefano Zampini if (n_ISForEdges) { 2456cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2457cf5a6209SStefano Zampini } 2458cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 245908122e43SStefano Zampini } else { 246008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2461984c4197SStefano Zampini 246208122e43SStefano Zampini total_counts = 0; 246308122e43SStefano Zampini n_vertices = 0; 2464d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2465d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 246608122e43SStefano Zampini } 246708122e43SStefano Zampini max_constraints = 0; 24689162d606SStefano Zampini total_counts_cc = 0; 246908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 247008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24719162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 247208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 247308122e43SStefano Zampini } 24749162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24759162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24769162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24779162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 247874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24799162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24809162d606SStefano Zampini total_counts_cc = 0; 24819162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24829162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24839162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 248408122e43SStefano Zampini } 248508122e43SStefano Zampini } 24869162d606SStefano Zampini #if 0 24879162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24889162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24899162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24909162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24919162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24929162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24939162d606SStefano Zampini } 24949162d606SStefano Zampini printf("\n"); 24959162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24969162d606SStefano Zampini } 24971b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24988bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24991b968477SStefano Zampini } 25001b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 25018bec7fa6SStefano 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]); 25021b968477SStefano Zampini } 250308122e43SStefano Zampini #endif 250408122e43SStefano Zampini 25058bec7fa6SStefano Zampini max_size_of_constraint = 0; 25069162d606SStefano 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]); 25079162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 250808122e43SStefano Zampini /* Change of basis */ 2509b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 251008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 251108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 251208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2513b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 251408122e43SStefano Zampini } 251508122e43SStefano Zampini } 251608122e43SStefano Zampini } 251708122e43SStefano Zampini } 2518984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 25190e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 252008122e43SStefano Zampini 25219162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 25229162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 25239162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 25249162d606SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",constraints_idxs_ptr[total_counts_cc],i); 252508122e43SStefano Zampini } 2526674ae819SStefano Zampini 2527674ae819SStefano Zampini /* Create constraint matrix */ 2528674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 252916f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2530984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2531984c4197SStefano Zampini 2532984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2533a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2534a717540cSStefano Zampini qr_needed = PETSC_FALSE; 253574d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2536984c4197SStefano Zampini total_primal_vertices=0; 2537b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 25389162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 25399162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2540984c4197SStefano Zampini if (size_of_constraint == 1) { 25419162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2542b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 254364efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 25449162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25459162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2546a717540cSStefano Zampini } 2547b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 254874d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2549a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2550a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2551a717540cSStefano Zampini } 2552fa434743SStefano Zampini } else { 2553b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2554fa434743SStefano Zampini } 2555a717540cSStefano Zampini } 2556b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2557b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2558674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 255970022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2560b3d85658SStefano Zampini 2561b3d85658SStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 25620e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25630e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2564984c4197SStefano Zampini 2565984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 256674d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2567785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2568984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 256974d5cdf7SStefano Zampini 2570984c4197SStefano Zampini j = total_primal_vertices; 257174d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2572b3d85658SStefano Zampini cum = total_primal_vertices; 25739162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25744641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2575b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2576b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2577b3d85658SStefano Zampini cum++; 25789162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 257974d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 258074d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 258174d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 258274d5cdf7SStefano Zampini } 25839162d606SStefano Zampini j += constraints_n[i]; 2584674ae819SStefano Zampini } 2585674ae819SStefano Zampini } 2586674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2587674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2588088faed8SStefano Zampini 2589674ae819SStefano Zampini /* set values in constraint matrix */ 2590984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25910e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2592674ae819SStefano Zampini } 2593984c4197SStefano Zampini total_counts = total_primal_vertices; 25949162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25954641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25969162d606SStefano Zampini PetscInt *cols; 25979162d606SStefano Zampini 25989162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25999162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 26009162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 26019162d606SStefano Zampini PetscInt row = total_counts+k; 26029162d606SStefano Zampini PetscScalar *vals; 26039162d606SStefano Zampini 26049162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 26059162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 26069162d606SStefano Zampini } 26079162d606SStefano Zampini total_counts += constraints_n[i]; 2608674ae819SStefano Zampini } 2609674ae819SStefano Zampini } 2610674ae819SStefano Zampini /* assembling */ 2611674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2612674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2613088faed8SStefano Zampini 2614984c4197SStefano Zampini /* 261545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2616984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2617984c4197SStefano Zampini */ 2618674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2619674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2620026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2621984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2622984c4197SStefano Zampini /* working stuff for GEQRF */ 262381d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2624984c4197SStefano Zampini PetscBLASInt lqr_work; 2625984c4197SStefano Zampini /* working stuff for UNGQR */ 2626984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2627984c4197SStefano Zampini PetscBLASInt lgqr_work; 2628984c4197SStefano Zampini /* working stuff for TRTRS */ 2629984c4197SStefano Zampini PetscScalar *trs_rhs; 26303f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2631984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2632984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2633984c4197SStefano Zampini PetscScalar *start_vals; 2634984c4197SStefano Zampini /* working stuff for values insertion */ 26354641a718SStefano Zampini PetscBT is_primal; 263664efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2637906d46d4SStefano Zampini /* matrix sizes */ 2638906d46d4SStefano Zampini PetscInt global_size,local_size; 2639906d46d4SStefano Zampini /* temporary change of basis */ 2640906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2641cf5a6209SStefano Zampini /* extra space for debugging */ 2642cf5a6209SStefano Zampini PetscScalar *dbg_work; 2643984c4197SStefano Zampini 2644906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2645906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 264616f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2647bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2648906d46d4SStefano Zampini /* nonzeros for local mat */ 2649bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2650bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 26519162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2652a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 26539162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2654a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 26559162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2656a717540cSStefano Zampini } else { 26579162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 26589162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2659a717540cSStefano Zampini } 2660a717540cSStefano Zampini } 2661a717540cSStefano Zampini } 2662906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2663bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2664a717540cSStefano Zampini /* Set initial identity in the matrix */ 2665bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2666906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2667a717540cSStefano Zampini } 2668a717540cSStefano Zampini 2669a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2670a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2671a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2672a717540cSStefano Zampini } 2673a717540cSStefano Zampini 2674a717540cSStefano Zampini 2675a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2676a717540cSStefano Zampini /* 2677a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2678a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2679a717540cSStefano Zampini 2680a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2681a717540cSStefano Zampini 2682a6b551f4SStefano 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) 2683a6b551f4SStefano Zampini 2684a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2685a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2686a717540cSStefano Zampini | ... | 2687a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2688a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2689a717540cSStefano Zampini 2690a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2691a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2692a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2693a6b551f4SStefano Zampini 2694a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2695a717540cSStefano Zampini */ 2696a717540cSStefano Zampini if (qr_needed) { 2697984c4197SStefano Zampini /* space to store Q */ 2698854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2699984c4197SStefano Zampini /* first we issue queries for optimal work */ 27003f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 27013f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 27023f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2703984c4197SStefano Zampini lqr_work = -1; 27043f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2705984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2706984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2707785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2708984c4197SStefano Zampini lgqr_work = -1; 27093f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 27103f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 27113f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 27123f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27133f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 27143f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2715984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2716984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2717785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2718984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2719785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2720984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2721785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2722a717540cSStefano Zampini /* allocating workspace for check */ 2723a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2724cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2725a717540cSStefano Zampini } 2726a717540cSStefano Zampini } 2727984c4197SStefano Zampini /* array to store whether a node is primal or not */ 27284641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2729473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 27300e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 273139e2fb2aSStefano Zampini if (i != total_primal_vertices) { 273239e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 27334641a718SStefano Zampini } 273439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 273539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 273639e2fb2aSStefano Zampini } 273739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2738984c4197SStefano Zampini 2739a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 27409162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 27419162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 27424641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2743984c4197SStefano Zampini /* get constraint info */ 27449162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2745984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2746984c4197SStefano Zampini 2747984c4197SStefano Zampini if (pcbddc->dbg_flag) { 27489162d606SStefano 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); 2749674ae819SStefano Zampini } 2750984c4197SStefano Zampini 2751fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2752a717540cSStefano Zampini 2753a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2754a717540cSStefano Zampini if (pcbddc->dbg_flag) { 27559162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2756a717540cSStefano Zampini } 2757984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 27589162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2759984c4197SStefano Zampini 2760984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27613f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27623f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27633f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2764674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27653f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2766984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2767674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2768984c4197SStefano Zampini 2769984c4197SStefano Zampini /* explictly compute R^-T */ 2770984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2771984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27753f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2776984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27773f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2778984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2779984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2780984c4197SStefano Zampini 2781a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27833f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27843f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27853f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2786984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27873f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2788984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2789984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2790984c4197SStefano Zampini 2791984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2792984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2793984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27963f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27973f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27983f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2800984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 28019162d606SStefano 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)); 2802984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 28039162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2804984c4197SStefano Zampini 2805984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 28069162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2807984c4197SStefano Zampini /* insert cols for primal dofs */ 2808984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2809984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 28109162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2811906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2812984c4197SStefano Zampini } 2813984c4197SStefano Zampini /* insert cols for dual dofs */ 2814984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 28159162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2816984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 28179162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2818906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2819984c4197SStefano Zampini j++; 2820674ae819SStefano Zampini } 2821674ae819SStefano Zampini } 2822984c4197SStefano Zampini 2823984c4197SStefano Zampini /* check change of basis */ 2824984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2825984c4197SStefano Zampini PetscInt ii,jj; 2826984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2827c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2828c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2829c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2830c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2831c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2832c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2833984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2834cf5a6209SStefano 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)); 2835984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2836984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2837984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2838cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2839cf5a6209SStefano 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; 2840674ae819SStefano Zampini } 2841674ae819SStefano Zampini } 2842984c4197SStefano Zampini if (!valid_qr) { 284322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2844984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2845984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2846cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2847cf5a6209SStefano 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])); 2848674ae819SStefano Zampini } 2849cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2850cf5a6209SStefano 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])); 2851984c4197SStefano Zampini } 2852984c4197SStefano Zampini } 2853984c4197SStefano Zampini } 2854674ae819SStefano Zampini } else { 285522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2856674ae819SStefano Zampini } 2857674ae819SStefano Zampini } 2858a717540cSStefano Zampini } else { /* simple transformation block */ 2859a717540cSStefano Zampini PetscInt row,col; 2860a6b551f4SStefano Zampini PetscScalar val,norm; 2861a6b551f4SStefano Zampini 2862a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28639162d606SStefano 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)); 2864a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28659162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28669162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2867bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28689162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2869906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28709162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2871a717540cSStefano Zampini } else { 2872a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28739162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2874a717540cSStefano Zampini if (row != col) { 28759162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2876a717540cSStefano Zampini } else { 28779162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2878a717540cSStefano Zampini } 2879906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2880a717540cSStefano Zampini } 2881a717540cSStefano Zampini } 2882a717540cSStefano Zampini } 288398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 288422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2885a717540cSStefano Zampini } 2886674ae819SStefano Zampini } 2887984c4197SStefano Zampini } else { 2888984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28899162d606SStefano 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); 2890674ae819SStefano Zampini } 2891674ae819SStefano Zampini } 2892674ae819SStefano Zampini } 2893a717540cSStefano Zampini 2894a717540cSStefano Zampini /* free workspace */ 2895a717540cSStefano Zampini if (qr_needed) { 2896984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2897cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2898984c4197SStefano Zampini } 2899984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2900984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2901984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2902984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2903984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2904674ae819SStefano Zampini } 2905a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2906906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2907906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2908906d46d4SStefano Zampini 2909906d46d4SStefano Zampini /* assembling of global change of variable */ 2910bbb9e6c6SStefano Zampini { 2911bbb9e6c6SStefano Zampini Mat tmat; 291216f15bc4SStefano Zampini PetscInt bs; 291316f15bc4SStefano Zampini 2914906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2915906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2916bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2917bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2918bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2919bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 292016f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 292116f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2922906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2923bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2924bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2925bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2926bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2927bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2928e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2929e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2930bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2931bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2932906d46d4SStefano Zampini } 2933906d46d4SStefano Zampini /* check */ 2934906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2935906d46d4SStefano Zampini PetscReal error; 2936906d46d4SStefano Zampini Vec x,x_change; 2937906d46d4SStefano Zampini 2938906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2939906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2940906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2941906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2942e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2943e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2944bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2945e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2946e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2947906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2951bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2952906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2953906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2954906d46d4SStefano Zampini } 2955b96c3477SStefano Zampini 2956b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2957b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2958b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2959b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2960ac632422SStefano Zampini Mat S_new,tmat; 2961bbb9e6c6SStefano Zampini IS is_all_N; 2962bbb9e6c6SStefano Zampini 2963bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29646816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2965bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2966ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2967b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2968ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2969ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2970ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2971ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2972ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2973b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2974ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2975ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2976ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2977ac632422SStefano Zampini } 2978b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2979b96c3477SStefano Zampini } 2980b96c3477SStefano Zampini } 2981906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2982906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2983b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2984b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2985b9b85e73SStefano Zampini } 2986906d46d4SStefano Zampini 2987906d46d4SStefano Zampini /* set up change of basis context */ 2988906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2989906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2990906d46d4SStefano Zampini 2991906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2992906d46d4SStefano Zampini PetscInt global_size,local_size; 2993906d46d4SStefano Zampini 2994906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2995906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2996906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2997906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2998906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2999906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3000906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3001906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3002906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3003906d46d4SStefano Zampini } else { 3004906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3005906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3006906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3007906d46d4SStefano Zampini } 3008906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3009906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3010906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3011906d46d4SStefano Zampini } else { 3012906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3013906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3014906d46d4SStefano Zampini } 3015906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3016906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3017906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3018906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3019b9b85e73SStefano Zampini } 3020a717540cSStefano Zampini 3021727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 3022727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3023727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3024aff50787SStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc*sizeof(PetscScalar),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 3025c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 30260e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3027aff50787SStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscScalar),&pcbddc->new_primal_space_local);CHKERRQ(ierr); 3028727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3029727cdba6SStefano Zampini } 30300e6343abSStefano Zampini } 30310e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3032727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3033727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3034727cdba6SStefano Zampini 3035a717540cSStefano Zampini /* flush dbg viewer */ 3036b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3037b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3038b8ffe317SStefano Zampini } 3039a717540cSStefano Zampini 3040e310c8b4SStefano Zampini /* free workspace */ 3041a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 30424641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 304308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 30449162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 30459162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 304608122e43SStefano Zampini } else { 30479162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 30489162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 30499162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 305008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 305108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 30529162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 30539162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 305408122e43SStefano Zampini } 3055674ae819SStefano Zampini PetscFunctionReturn(0); 3056674ae819SStefano Zampini } 3057674ae819SStefano Zampini 3058674ae819SStefano Zampini #undef __FUNCT__ 3059674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3060674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3061674ae819SStefano Zampini { 3062674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3063674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3064674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30657fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3066674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3067674ae819SStefano Zampini 3068674ae819SStefano Zampini PetscFunctionBegin; 30698e61c736SStefano Zampini /* Reset previously computed graph */ 30708e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3071674ae819SStefano Zampini /* Init local Graph struct */ 30727fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30733bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3074674ae819SStefano Zampini 3075575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3076575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3077575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3078575ad6abSStefano Zampini } 30799577ea80SStefano Zampini 3080674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30814d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30824d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30834d379d7bSStefano Zampini PetscInt nvtxs; 3084e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3085674ae819SStefano Zampini 30864d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30872fffb893SStefano Zampini 30882fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30892fffb893SStefano Zampini if (flg_row) { 30904d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3091b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30922fffb893SStefano Zampini } 30932fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30949b28b941SStefano Zampini } else if (pcbddc->current_level && pcis->n_B) { /* just compute subdomain's connected components for coarser levels when the local boundary is not empty */ 30954d379d7bSStefano Zampini IS is_dummy; 30964d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30974d379d7bSStefano Zampini PetscInt j,sum; 30984d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30994d379d7bSStefano Zampini const PetscInt *idxs; 31004d379d7bSStefano Zampini PCBDDCGraph graph; 31014d379d7bSStefano Zampini PetscBT is_on_boundary; 31024d379d7bSStefano Zampini 31034d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 31044d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 31054d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 31064d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 31077fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 31084d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3109e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3110e496cd5dSStefano Zampini if (flg_row) { 31114d379d7bSStefano Zampini graph->xadj = xadj; 31124d379d7bSStefano Zampini graph->adjncy = adjncy; 3113e496cd5dSStefano Zampini } 31144d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 31154d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3116e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 31174d379d7bSStefano Zampini 31184d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 31199b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 31204d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31214d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 31224d379d7bSStefano Zampini } 31234d379d7bSStefano Zampini } 31244d379d7bSStefano Zampini 3125e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 31264d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 31274d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 31284d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 31294d379d7bSStefano Zampini } 31304d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 31314d379d7bSStefano Zampini 3132e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 31334d379d7bSStefano Zampini sum = 0; 31344d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31354d379d7bSStefano Zampini PetscInt sizecc = 0; 31364d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31374d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31384d379d7bSStefano Zampini sizecc++; 31394d379d7bSStefano Zampini } 31404d379d7bSStefano Zampini } 31414d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31424d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31434d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 31444d379d7bSStefano Zampini } 31454d379d7bSStefano Zampini } 31464d379d7bSStefano Zampini sum += sizecc*sizecc; 31474d379d7bSStefano Zampini } 31484d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 31494d379d7bSStefano Zampini sum = 0; 3150e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 31514d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 31524d379d7bSStefano Zampini cxadj[i] = sum; 31534d379d7bSStefano Zampini sum += temp; 31544d379d7bSStefano Zampini } 3155e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 31564d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31574d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31584d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31594d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31604d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31614d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31624d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31634d379d7bSStefano Zampini sizecc++; 31644d379d7bSStefano Zampini } 31654d379d7bSStefano Zampini } 31664d379d7bSStefano Zampini } 31674d379d7bSStefano Zampini } 31684d379d7bSStefano Zampini } 31699b28b941SStefano Zampini if (sum) { 3170e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31714d379d7bSStefano Zampini } else { 31724d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31734d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31744d379d7bSStefano Zampini } 31754d379d7bSStefano Zampini graph->xadj = 0; 31764d379d7bSStefano Zampini graph->adjncy = 0; 31774d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31784d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31794d379d7bSStefano Zampini } 3180674ae819SStefano Zampini } 31819b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31829b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3183674ae819SStefano Zampini } 3184674ae819SStefano Zampini 318563602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3186674ae819SStefano Zampini vertex_size = 1; 318763602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 318863602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 318995ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 319063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3191e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 319263602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3193674ae819SStefano Zampini } 319463602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 319563602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 319663602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3197674ae819SStefano Zampini } 319863602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3199674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 320063602bcaSStefano Zampini } else { 320163602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 320263602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3203854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 320463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 320563602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 320663602bcaSStefano Zampini } 320763602bcaSStefano Zampini } 3208674ae819SStefano Zampini } 3209674ae819SStefano Zampini 3210674ae819SStefano Zampini /* Setup of Graph */ 3211785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3212e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3213785d1243SStefano Zampini } 3214785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3215e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3216785d1243SStefano Zampini } 3217302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3218674ae819SStefano Zampini 3219674ae819SStefano Zampini /* Graph's connected components analysis */ 3220674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3221674ae819SStefano Zampini 3222674ae819SStefano Zampini /* print some info to stdout */ 3223674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3224302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3225674ae819SStefano Zampini } 3226fb180af4SStefano Zampini 3227fb180af4SStefano Zampini /* mark topography has done */ 3228fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3229674ae819SStefano Zampini PetscFunctionReturn(0); 3230674ae819SStefano Zampini } 3231674ae819SStefano Zampini 3232dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3233674ae819SStefano Zampini #undef __FUNCT__ 3234674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3235dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3236674ae819SStefano Zampini { 3237dc456d91SStefano Zampini PetscSF sf; 3238dc456d91SStefano Zampini PetscLayout map; 3239dc456d91SStefano Zampini const PetscInt *idxs; 3240dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3241dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3242dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3243dc456d91SStefano Zampini PetscMPIInt commsize; 3244674ae819SStefano Zampini PetscBool first_found; 3245674ae819SStefano Zampini PetscErrorCode ierr; 3246674ae819SStefano Zampini 3247674ae819SStefano Zampini PetscFunctionBegin; 3248dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3249dc456d91SStefano Zampini if (subset_mult) { 3250dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3251dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3252dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3253674ae819SStefano Zampini } 3254dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3255dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3256dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3257dc456d91SStefano Zampini for (i=0;i<n;i++) { 3258dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3259dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3260674ae819SStefano Zampini } 3261dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3262dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3263dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3264dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3265dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3266dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3267dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3268dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3269dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3270dc456d91SStefano Zampini 3271dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3272dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3273dc456d91SStefano Zampini if (subset_mult) { 3274dc456d91SStefano Zampini const PetscInt* idxs_mult; 3275dc456d91SStefano Zampini 3276dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3277dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3278dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3279674ae819SStefano Zampini } else { 3280dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3281674ae819SStefano Zampini } 3282dc456d91SStefano Zampini /* local size of new subset */ 3283dc456d91SStefano Zampini n_n = 0; 3284dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3285dc456d91SStefano Zampini 3286dc456d91SStefano Zampini /* global indexes in layout */ 3287dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3288dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3289dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3290dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3291dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3292dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3293dc456d91SStefano Zampini 3294dc456d91SStefano Zampini /* reduce from leaves to roots */ 3295dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 329664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 329764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3298dc456d91SStefano Zampini 3299dc456d91SStefano Zampini /* count indexes in local part of layout */ 3300674ae819SStefano Zampini nlocals = 0; 3301674ae819SStefano Zampini first_index = -1; 3302674ae819SStefano Zampini first_found = PETSC_FALSE; 3303dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3304dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3305674ae819SStefano Zampini first_found = PETSC_TRUE; 3306674ae819SStefano Zampini first_index = i; 3307674ae819SStefano Zampini } 3308dc456d91SStefano Zampini nlocals += root_data[i]; 3309674ae819SStefano Zampini } 3310dc456d91SStefano Zampini 3311dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 33125fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3313dc456d91SStefano Zampini start = 0; 331464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 33155fa240b1SStefano Zampini #else 331664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 33175fa240b1SStefano Zampini start = start-nlocals; 33185fa240b1SStefano Zampini #endif 33195fa240b1SStefano Zampini 3320dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3321dc456d91SStefano Zampini *N_n = start + nlocals; 3322dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3323dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3324674ae819SStefano Zampini } 33255fa240b1SStefano Zampini 33265fa240b1SStefano Zampini /* adapt root data with cumulative */ 3327674ae819SStefano Zampini if (first_found) { 3328dc456d91SStefano Zampini PetscInt old_index; 3329dc456d91SStefano Zampini 3330dc456d91SStefano Zampini root_data[first_index] += start; 3331674ae819SStefano Zampini old_index = first_index; 3332dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3333dc456d91SStefano Zampini if (root_data[i]) { 3334dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3335674ae819SStefano Zampini old_index = i; 3336674ae819SStefano Zampini } 3337674ae819SStefano Zampini } 3338674ae819SStefano Zampini } 3339dc456d91SStefano Zampini 3340dc456d91SStefano Zampini /* from roots to leaves */ 3341dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3342dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3343dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3344dc456d91SStefano Zampini 3345dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3346dc456d91SStefano Zampini if (subset_mult) { 3347dc456d91SStefano Zampini const PetscInt* idxs_mult; 3348dc456d91SStefano Zampini PetscInt cum; 3349dc456d91SStefano Zampini 3350dc456d91SStefano Zampini cum = 0; 3351dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3352dc456d91SStefano Zampini for (i=0;i<n;i++) { 3353dc456d91SStefano Zampini PetscInt j; 3354dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3355674ae819SStefano Zampini } 3356dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3357674ae819SStefano Zampini } else { 3358dc456d91SStefano Zampini for (i=0;i<n;i++) { 3359dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3360674ae819SStefano Zampini } 3361674ae819SStefano Zampini } 3362dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3363dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3364674ae819SStefano Zampini PetscFunctionReturn(0); 3365674ae819SStefano Zampini } 33669a7d3425SStefano Zampini 33679a7d3425SStefano Zampini #undef __FUNCT__ 33689a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33699a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33709a7d3425SStefano Zampini { 33719a7d3425SStefano Zampini PetscInt i,j; 33729a7d3425SStefano Zampini PetscScalar *alphas; 33739a7d3425SStefano Zampini PetscErrorCode ierr; 33749a7d3425SStefano Zampini 33759a7d3425SStefano Zampini PetscFunctionBegin; 33769a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3377785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33789a7d3425SStefano Zampini for (i=0;i<n;i++) { 33799a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33809a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33819a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33829a7d3425SStefano Zampini } 33839a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33849a7d3425SStefano Zampini PetscFunctionReturn(0); 33859a7d3425SStefano Zampini } 33869a7d3425SStefano Zampini 3387e7931f94SStefano Zampini #undef __FUNCT__ 338870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3389b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3390e7931f94SStefano Zampini { 339152e5ac9dSStefano Zampini IS ranks_send_to; 3392e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3393e7931f94SStefano Zampini PetscMPIInt size,rank,color; 339452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 339552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 33963837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 33972b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3398e7931f94SStefano Zampini PetscSubcomm subcomm; 339952e5ac9dSStefano Zampini PetscErrorCode ierr; 3400a57a6d2fSStefano Zampini 3401e7931f94SStefano Zampini PetscFunctionBegin; 34022b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 34032b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 34042b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3405e7931f94SStefano Zampini 3406e7931f94SStefano Zampini /* Get info on mapping */ 34073bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 34083bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3409e7931f94SStefano Zampini 3410e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3411785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3412e7931f94SStefano Zampini xadj[0] = 0; 3413e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3414785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3415785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3416e7931f94SStefano Zampini 34172b510759SStefano Zampini if (threshold) { 3418d023bfaeSStefano Zampini PetscInt xadj_count = 0; 34192b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3420d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3421d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3422d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3423d023bfaeSStefano Zampini xadj_count++; 3424e7931f94SStefano Zampini } 3425e7931f94SStefano Zampini } 3426d023bfaeSStefano Zampini xadj[1] = xadj_count; 3427c8587f34SStefano Zampini } else { 3428e7931f94SStefano Zampini if (xadj[1]) { 3429e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3430e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3431c8587f34SStefano Zampini } 3432e7931f94SStefano Zampini } 34333bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3434e7931f94SStefano Zampini if (use_square) { 3435e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3436e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3437e7931f94SStefano Zampini } 3438e7931f94SStefano Zampini } 3439e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3440e7931f94SStefano Zampini 34413837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3442e7931f94SStefano Zampini 3443e7931f94SStefano Zampini /* 3444e7931f94SStefano Zampini Restrict work on active processes only. 3445e7931f94SStefano Zampini */ 3446e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3447e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3448e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 34492b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3450d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3451e7931f94SStefano Zampini if (color) { 3452e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3453e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3454e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3455c8587f34SStefano Zampini } else { 345652e5ac9dSStefano Zampini Mat subdomain_adj; 345752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 345852e5ac9dSStefano Zampini MatPartitioning partitioner; 345952e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 346052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3461b0c7d250SStefano Zampini PetscBool aggregate; 3462b0c7d250SStefano Zampini 3463306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3464785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3465e7931f94SStefano Zampini prank = rank; 3466306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34678002ef2cSStefano Zampini /* 3468e7931f94SStefano Zampini for (i=0;i<size;i++) { 3469e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3470c8587f34SStefano Zampini } 34718002ef2cSStefano Zampini */ 3472e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3473e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3474c8587f34SStefano Zampini } 3475e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3476b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3477b0c7d250SStefano Zampini if (aggregate) { 3478b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3479b0c7d250SStefano Zampini PetscMPIInt nrank; 3480b0c7d250SStefano Zampini PetscScalar *vals; 3481b0c7d250SStefano Zampini 3482b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3483b0c7d250SStefano Zampini lrows = 0; 3484b0c7d250SStefano Zampini if (nrank<redprocs) { 3485b0c7d250SStefano Zampini lrows = size/redprocs; 3486b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3487b0c7d250SStefano Zampini } 34885fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3489b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3490b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3491b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3492b0c7d250SStefano Zampini row = nrank; 3493b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3494b0c7d250SStefano Zampini cols = adjncy; 3495b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3496b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3497b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3498b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3499b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 350052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 350152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 350252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3503b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3504b0c7d250SStefano Zampini } else { 3505306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3506b0c7d250SStefano Zampini } 350722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3508e7931f94SStefano Zampini 3509e7931f94SStefano Zampini /* Partition */ 3510306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3511e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3512e7931f94SStefano Zampini if (use_vwgt) { 35133837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3514e7931f94SStefano Zampini v_wgt[0] = local_size; 3515e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3516c8587f34SStefano Zampini } 351728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 351828143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3519e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3520e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 352122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3522e7931f94SStefano Zampini 352352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 352452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 352552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 352652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3527b0c7d250SStefano Zampini if (!redprocs) { 3528b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 352928143c3dSStefano Zampini } else { 3530b0c7d250SStefano Zampini PetscInt idxs[1]; 3531b0c7d250SStefano Zampini PetscMPIInt tag; 3532b0c7d250SStefano Zampini MPI_Request *reqs; 3533b0c7d250SStefano Zampini 3534b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3535b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3536b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3537b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 353828143c3dSStefano Zampini } 3539b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3540b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3541b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3542b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3543e7931f94SStefano Zampini } 354452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3545e7931f94SStefano Zampini /* clean up */ 3546e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 354752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3548e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3549e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3550e7931f94SStefano Zampini } 3551e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3552e7931f94SStefano Zampini 3553e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3554e7931f94SStefano Zampini i = 1; 3555e7931f94SStefano Zampini if (color) i=0; 3556e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3557e7931f94SStefano Zampini /* get back IS */ 3558e7931f94SStefano Zampini *is_sends = ranks_send_to; 3559e7931f94SStefano Zampini PetscFunctionReturn(0); 3560e7931f94SStefano Zampini } 3561e7931f94SStefano Zampini 3562e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3563e7931f94SStefano Zampini 3564e7931f94SStefano Zampini #undef __FUNCT__ 3565e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 356653a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3567e7931f94SStefano Zampini { 356870cf5478SStefano Zampini Mat local_mat; 3569e7931f94SStefano Zampini IS is_sends_internal; 35709d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 357128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 35729d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3573e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3574e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3575e7931f94SStefano Zampini const PetscInt* is_indices; 3576e7931f94SStefano Zampini MatType new_local_type; 3577e7931f94SStefano Zampini /* buffers */ 3578e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 357928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35809d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3581e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3582e7931f94SStefano Zampini /* MPI */ 358328143c3dSStefano Zampini MPI_Comm comm,comm_n; 358428143c3dSStefano Zampini PetscSubcomm subcomm; 3585e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 358628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 358728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 358828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 358928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 359028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3591e7931f94SStefano Zampini PetscErrorCode ierr; 3592e7931f94SStefano Zampini 3593e7931f94SStefano Zampini PetscFunctionBegin; 359428143c3dSStefano Zampini /* TODO: add missing checks */ 359528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 359628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 359728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 359828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3599e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 360028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3601e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3602e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3603e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3604e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3605e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 360628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 360770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 360870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 360928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 361070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 361170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 361270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 361370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 361470cf5478SStefano Zampini } 3615e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3616e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3617e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3618e7931f94SStefano Zampini if (!is_sends) { 361928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3620b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3621c8587f34SStefano Zampini } else { 3622e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3623e7931f94SStefano Zampini is_sends_internal = is_sends; 3624c8587f34SStefano Zampini } 3625e7931f94SStefano Zampini 3626e7931f94SStefano Zampini /* get comm */ 3627a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3628e7931f94SStefano Zampini 3629e7931f94SStefano Zampini /* compute number of sends */ 3630e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3631e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3632e7931f94SStefano Zampini 3633e7931f94SStefano Zampini /* compute number of receives */ 3634e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3635785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3636e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3637e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3638e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3639e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3640e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3641e7931f94SStefano Zampini 364228143c3dSStefano Zampini /* restrict comm if requested */ 364328143c3dSStefano Zampini subcomm = 0; 364428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 364528143c3dSStefano Zampini if (restrict_comm) { 3646779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3647779c1cceSStefano Zampini 364828143c3dSStefano Zampini color = 0; 364953a05cb3SStefano Zampini if (restrict_full) { 365053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 365153a05cb3SStefano Zampini } else { 365253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 365353a05cb3SStefano Zampini } 365428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 365528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 365628143c3dSStefano Zampini /* check if reuse has been requested */ 365728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 365828143c3dSStefano Zampini if (*mat_n) { 365928143c3dSStefano Zampini PetscMPIInt subcommsize2; 366028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 366128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 366228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 366328143c3dSStefano Zampini } else { 366428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 366528143c3dSStefano Zampini } 366628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3667779c1cceSStefano Zampini PetscMPIInt rank; 3668779c1cceSStefano Zampini 3669779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 367028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 367128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 367228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3673306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 367428143c3dSStefano Zampini } 367528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 367628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 367728143c3dSStefano Zampini } else { 367828143c3dSStefano Zampini comm_n = comm; 367928143c3dSStefano Zampini } 368028143c3dSStefano Zampini 3681e7931f94SStefano Zampini /* prepare send/receive buffers */ 3682785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3683e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3684785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3685e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 368628143c3dSStefano Zampini if (nis) { 3687854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 368828143c3dSStefano Zampini } 3689e7931f94SStefano Zampini 369028143c3dSStefano Zampini /* Get data from local matrices */ 3691e7931f94SStefano Zampini if (!isdense) { 3692a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3693e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3694e7931f94SStefano Zampini /* 3695e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3696e7931f94SStefano Zampini send_buffer_idxs should contain: 3697e7931f94SStefano Zampini - MatType_PRIVATE type 3698e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3699e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3700e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3701e7931f94SStefano Zampini */ 3702e7931f94SStefano Zampini } else { 3703e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 37043bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3705854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3706e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3707e7931f94SStefano Zampini send_buffer_idxs[1] = i; 37083bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3709e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 37103bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3711e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3712e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3713e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3714e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3715c8587f34SStefano Zampini } 3716c8587f34SStefano Zampini } 3717e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 371828143c3dSStefano Zampini /* additional is (if any) */ 371928143c3dSStefano Zampini if (nis) { 372028143c3dSStefano Zampini PetscMPIInt psum; 372128143c3dSStefano Zampini PetscInt j; 372228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 372328143c3dSStefano Zampini PetscInt plen; 372428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 372528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 372628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 372728143c3dSStefano Zampini } 3728854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 372928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 373028143c3dSStefano Zampini PetscInt plen; 373128143c3dSStefano Zampini const PetscInt *is_array_idxs; 373228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 373328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 373428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 373528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 373628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 373728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 373828143c3dSStefano Zampini } 373928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 374028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 374128143c3dSStefano Zampini } 374228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 374328143c3dSStefano Zampini } 374428143c3dSStefano Zampini 3745e7931f94SStefano Zampini buf_size_idxs = 0; 3746e7931f94SStefano Zampini buf_size_vals = 0; 374728143c3dSStefano Zampini buf_size_idxs_is = 0; 3748e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3749e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3750e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 375128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3752e7931f94SStefano Zampini } 3753785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3754785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 375595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3756e7931f94SStefano Zampini 3757e7931f94SStefano Zampini /* get new tags for clean communications */ 3758e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3759e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 376028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3761e7931f94SStefano Zampini 3762e7931f94SStefano Zampini /* allocate for requests */ 3763785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3764785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 376595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3766785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3767785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 376895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3769e7931f94SStefano Zampini 3770e7931f94SStefano Zampini /* communications */ 3771e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3772e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 377328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3774e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3775e7931f94SStefano Zampini source_dest = onodes[i]; 3776e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3777e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3778e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3779e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 378028143c3dSStefano Zampini if (nis) { 378128143c3dSStefano 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); 378228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 378328143c3dSStefano Zampini } 3784e7931f94SStefano Zampini } 3785e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3786e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3787e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3788e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 378928143c3dSStefano Zampini if (nis) { 379028143c3dSStefano 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); 379128143c3dSStefano Zampini } 3792e7931f94SStefano Zampini } 3793e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3794e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3795e7931f94SStefano Zampini 3796e7931f94SStefano Zampini /* assemble new l2g map */ 3797e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3798e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37999d30be91SStefano Zampini new_local_rows = 0; 3800e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 38019d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3802e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3803e7931f94SStefano Zampini } 38049d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3805e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 38069d30be91SStefano Zampini new_local_rows = 0; 3807e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 38089d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 38099d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3810e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3811e7931f94SStefano Zampini } 38129d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 38139d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3814e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3815e7931f94SStefano Zampini 3816e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3817e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3818e7931f94SStefano 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) */ 3819e7931f94SStefano Zampini if (n_recvs) { 382028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3821e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3822e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3823e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3824e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3825e7931f94SStefano Zampini break; 3826e7931f94SStefano Zampini } 3827e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3828e7931f94SStefano Zampini } 3829e7931f94SStefano Zampini switch (new_local_type_private) { 383028143c3dSStefano Zampini case MATDENSE_PRIVATE: 383128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3832e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3833e7931f94SStefano Zampini bs = 1; 383428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 383528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 383628143c3dSStefano Zampini bs = 1; 383728143c3dSStefano Zampini } 3838e7931f94SStefano Zampini break; 3839e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3840e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3841e7931f94SStefano Zampini bs = 1; 3842e7931f94SStefano Zampini break; 3843e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3844e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3845e7931f94SStefano Zampini break; 3846e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3847e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3848e7931f94SStefano Zampini break; 3849e7931f94SStefano Zampini default: 38509d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3851e7931f94SStefano Zampini break; 3852e7931f94SStefano Zampini } 385328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 385428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 385528143c3dSStefano Zampini bs = 1; 3856e7931f94SStefano Zampini } 3857e7931f94SStefano Zampini 385870cf5478SStefano Zampini /* create MATIS object if needed */ 385970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3860e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3861e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 386270cf5478SStefano Zampini } else { 386370cf5478SStefano Zampini /* it also destroys the local matrices */ 386470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 386570cf5478SStefano Zampini } 386670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3867e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 38689d30be91SStefano Zampini 38699d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 38709d30be91SStefano Zampini 38719d30be91SStefano Zampini /* Global to local map of received indices */ 38729d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 38739d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 38749d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 38759d30be91SStefano Zampini 38769d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38779d30be91SStefano Zampini buf_size_idxs = 0; 38789d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38799d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38809d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38819d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38829d30be91SStefano Zampini } 38839d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38849d30be91SStefano Zampini 38859d30be91SStefano Zampini /* set preallocation */ 38869d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38879d30be91SStefano Zampini if (!newisdense) { 38889d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38899d30be91SStefano Zampini 38909d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 38919d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38929d30be91SStefano Zampini if (n_recvs) { 38939d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38949d30be91SStefano Zampini } 38959d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38969d30be91SStefano Zampini PetscInt j; 38979d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38989d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38999d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 39009d30be91SStefano Zampini } 39019d30be91SStefano Zampini } else { 39029d30be91SStefano Zampini /* TODO */ 39039d30be91SStefano Zampini } 39049d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 39059d30be91SStefano Zampini } 39069d30be91SStefano Zampini if (new_local_nnz) { 39079d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 39089d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 39099d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 39109d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 39119d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 39129d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 39139d30be91SStefano Zampini } else { 39149d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 39159d30be91SStefano Zampini } 39169d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 39179d30be91SStefano Zampini } else { 39189d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 39199d30be91SStefano Zampini } 3920e7931f94SStefano Zampini 3921e7931f94SStefano Zampini /* set values */ 3922e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 39239d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3924e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3925e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3926e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 39279d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3928e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3929e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3930e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 393128143c3dSStefano Zampini } else { 393228143c3dSStefano Zampini /* TODO */ 3933e7931f94SStefano Zampini } 3934e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3935e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3936e7931f94SStefano Zampini } 3937e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3938e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 393970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 394070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 39419d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 39429d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3943e7931f94SStefano Zampini 3944dfd14d43SStefano Zampini #if 0 394528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3946e7931f94SStefano Zampini Vec lvec,rvec; 3947e7931f94SStefano Zampini PetscReal infty_error; 3948e7931f94SStefano Zampini 39492a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3950e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3951e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3952e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 395370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3954e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3955e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3956e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3957e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3958e7931f94SStefano Zampini } 395928143c3dSStefano Zampini #endif 3960e7931f94SStefano Zampini 396128143c3dSStefano Zampini /* assemble new additional is (if any) */ 396228143c3dSStefano Zampini if (nis) { 396328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 396428143c3dSStefano Zampini 396528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3966854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 396728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 396828143c3dSStefano Zampini psum = 0; 396928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 397028143c3dSStefano Zampini for (j=0;j<nis;j++) { 397128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 397228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 397328143c3dSStefano Zampini psum += plen; 397428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 397528143c3dSStefano Zampini } 397628143c3dSStefano Zampini } 3977854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3978854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 397928143c3dSStefano Zampini for (i=1;i<nis;i++) { 398028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 398128143c3dSStefano Zampini } 398228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 398328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 398428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 398528143c3dSStefano Zampini for (j=0;j<nis;j++) { 398628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 398728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 398828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 398928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 399028143c3dSStefano Zampini } 399128143c3dSStefano Zampini } 399228143c3dSStefano Zampini for (i=0;i<nis;i++) { 399328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 399428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 399528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 399628143c3dSStefano Zampini } 399728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 399828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 399928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 400028143c3dSStefano Zampini } 4001e7931f94SStefano Zampini /* free workspace */ 400228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4003e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4004e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4005e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4006e7931f94SStefano Zampini if (isdense) { 4007e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4008e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4009e7931f94SStefano Zampini } else { 4010e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4011e7931f94SStefano Zampini } 401228143c3dSStefano Zampini if (nis) { 401328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 401428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 401528143c3dSStefano Zampini } 4016e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4017e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 401828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4019e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4020e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 402128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4022e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4023e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4024e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4025e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4026e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 402728143c3dSStefano Zampini if (nis) { 402828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 402928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 403028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 403128143c3dSStefano Zampini } 403228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 403328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 403428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 403528143c3dSStefano Zampini for (i=0;i<nis;i++) { 403628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 403728143c3dSStefano Zampini } 403853a05cb3SStefano Zampini *mat_n = NULL; 403928143c3dSStefano Zampini } 4040e7931f94SStefano Zampini PetscFunctionReturn(0); 4041e7931f94SStefano Zampini } 4042a57a6d2fSStefano Zampini 404312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4044af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 404512edc857SStefano Zampini 4046c8587f34SStefano Zampini #undef __FUNCT__ 4047c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4048c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4049c8587f34SStefano Zampini { 4050c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4051c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 405220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 40539881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 405420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 40556e683305SStefano Zampini IS coarse_is,*isarray; 40566e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 40576e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4058f9eb5b7dSStefano Zampini PC pc_temp; 4059c8587f34SStefano Zampini PCType coarse_pc_type; 4060c8587f34SStefano Zampini KSPType coarse_ksp_type; 4061f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 40624f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 40636e683305SStefano Zampini Mat t_coarse_mat_is; 40646e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 40656e683305SStefano Zampini PetscMPIInt all_procs; 406674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 406768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 406822bc73bbSStefano Zampini PetscScalar *array; 40699881197aSStefano Zampini PetscErrorCode ierr; 4070fdc09c96SStefano Zampini 4071c8587f34SStefano Zampini PetscFunctionBegin; 4072c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 407368457ee5SStefano 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 */ 4074fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 40755a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4076fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4077f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4078f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4079f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4080fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 408151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 408251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4083dc4bcba2SStefano Zampini PC pc; 4084dc4bcba2SStefano Zampini PetscBool isbddc; 4085dc4bcba2SStefano Zampini 4086dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4087dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4088dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4089dc4bcba2SStefano Zampini if (isbddc) { 4090dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4091dc4bcba2SStefano Zampini } 4092727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4093fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4094fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4095fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4096f4ddd8eeSStefano Zampini } 4097fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4098fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4099f4ddd8eeSStefano Zampini } 410070cf5478SStefano Zampini /* reset any subassembling information */ 410170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 41026e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41036e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4104fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4105f4ddd8eeSStefano Zampini } 4106c8587f34SStefano Zampini 41076e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 41082b510759SStefano Zampini im_active = !!(pcis->n); 41092b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 41106e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 41116e683305SStefano Zampini void_procs = all_procs-active_procs; 41126e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 411374e2c79eSStefano Zampini redist = PETSC_FALSE; 411422bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 41156e683305SStefano Zampini csin_ml = PETSC_TRUE; 41166e683305SStefano Zampini ncoarse_ml = void_procs; 4117779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4118779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 41196e683305SStefano Zampini csin_ds = PETSC_TRUE; 412018a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 412118a45a71SStefano Zampini redist = PETSC_TRUE; 412218a45a71SStefano Zampini } else { 41236e683305SStefano Zampini csin_ds = PETSC_TRUE; 4124779c1cceSStefano Zampini ncoarse_ds = active_procs; 4125779c1cceSStefano Zampini redist = PETSC_TRUE; 412618a45a71SStefano Zampini } 41276e683305SStefano Zampini } else { 41286e683305SStefano Zampini csin_ml = PETSC_FALSE; 41296e683305SStefano Zampini ncoarse_ml = all_procs; 41306e683305SStefano Zampini if (void_procs) { 41316e683305SStefano Zampini csin_ds = PETSC_TRUE; 41326e683305SStefano Zampini ncoarse_ds = void_procs; 41336e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 41346e683305SStefano Zampini } else { 4135779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 413674e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 413774e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 413874e2c79eSStefano Zampini redist = PETSC_TRUE; 413974e2c79eSStefano Zampini } else { 41406e683305SStefano Zampini csin_ds = PETSC_FALSE; 41416e683305SStefano Zampini ncoarse_ds = all_procs; 41426e683305SStefano Zampini } 41436e683305SStefano Zampini } 414474e2c79eSStefano Zampini } 41456e683305SStefano Zampini 41466e683305SStefano Zampini /* 41476e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 41486e683305SStefano Zampini - we have not exceeded the number of levels requested 41496e683305SStefano Zampini - we can actually subassemble the active processes 41506e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 41516e683305SStefano Zampini */ 41526e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 41536e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 41546e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 41556e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 41566e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4157f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 41582b510759SStefano Zampini } else { 4159f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4160c8587f34SStefano Zampini } 4161c8587f34SStefano Zampini } 41626e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 41636e683305SStefano Zampini if (multilevel_allowed) { 41646e683305SStefano Zampini ncoarse = ncoarse_ml; 41656e683305SStefano Zampini csin = csin_ml; 416658da7f69SStefano Zampini redist = PETSC_FALSE; 41676e683305SStefano Zampini } else { 41686e683305SStefano Zampini ncoarse = ncoarse_ds; 41696e683305SStefano Zampini csin = csin_ds; 41706e683305SStefano Zampini } 4171e7931f94SStefano Zampini 4172abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4173abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4174abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4175abbbba34SStefano Zampini 4176abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 417722bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 417822bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 417922bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 418022bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4181b9b85e73SStefano Zampini #if 0 4182b9b85e73SStefano Zampini { 4183b9b85e73SStefano Zampini PetscViewer viewer; 4184b9b85e73SStefano Zampini char filename[256]; 4185b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4186b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4187b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4188b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4189b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4190b9b85e73SStefano Zampini } 4191b9b85e73SStefano Zampini #endif 4192e176bc59SStefano 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); 41936e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41946e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41956e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4196abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4197abbbba34SStefano Zampini 41986e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41996e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 42006e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 42016e683305SStefano Zampini const PetscInt *idxs; 42026e683305SStefano Zampini ISLocalToGlobalMapping tmap; 42036e683305SStefano Zampini 42046e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 42050be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 42066e683305SStefano Zampini /* allocate space for temporary storage */ 4207854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4208854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 42096e683305SStefano Zampini /* allocate for IS array */ 42106e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 42116e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 42126e683305SStefano Zampini nis = nisdofs + nisneu; 4213854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 42146e683305SStefano Zampini /* dofs splitting */ 42156e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 42166e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 42176e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 42186e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 42196e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 42206e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 42216e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 42226e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 42236e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 42246e683305SStefano Zampini } 42256e683305SStefano Zampini /* neumann boundaries */ 42266e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 42276e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 42286e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 42296e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 42306e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 42316e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 42326e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 42336e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 42346e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 42356e683305SStefano Zampini } 42366e683305SStefano Zampini /* free memory */ 42376e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 42386e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 42396e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 42406e683305SStefano Zampini } else { 42416e683305SStefano Zampini nis = 0; 42426e683305SStefano Zampini nisdofs = 0; 42436e683305SStefano Zampini nisneu = 0; 42446e683305SStefano Zampini isarray = NULL; 42456e683305SStefano Zampini } 42466e683305SStefano Zampini /* destroy no longer needed map */ 42476e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 42486e683305SStefano Zampini 42496e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 42506e683305SStefano Zampini coarse_mat_is = NULL; 42516e683305SStefano Zampini if (csin) { 42526e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 425374e2c79eSStefano Zampini if (redist) { 425474e2c79eSStefano Zampini PetscMPIInt rank; 4255779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 425674e2c79eSStefano Zampini 425774e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 425858da7f69SStefano Zampini spc = active_procs/ncoarse; 425958da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4260779c1cceSStefano Zampini if (im_active) { 4261779c1cceSStefano Zampini destsize = 1; 426274e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 426374e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 426474e2c79eSStefano Zampini } else { 426574e2c79eSStefano Zampini dest[0] = rank/(spc+1); 426674e2c79eSStefano Zampini } 426774e2c79eSStefano Zampini } else { 4268779c1cceSStefano Zampini destsize = 0; 42696e683305SStefano Zampini } 4270779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4271779c1cceSStefano Zampini } else if (csin_type_simple) { 42726e683305SStefano Zampini PetscMPIInt rank; 42736e683305SStefano Zampini PetscInt issize,isidx; 4274779c1cceSStefano Zampini 42756e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42766e683305SStefano Zampini if (im_active) { 42776e683305SStefano Zampini issize = 1; 42786e683305SStefano Zampini isidx = (PetscInt)rank; 42796e683305SStefano Zampini } else { 42806e683305SStefano Zampini issize = 0; 42816e683305SStefano Zampini isidx = -1; 42826e683305SStefano Zampini } 42836e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4284779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4285b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42866e683305SStefano Zampini } 4287779c1cceSStefano Zampini 4288779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4289779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4290779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4291779c1cceSStefano Zampini PetscInt *coarse_candidates; 4292779c1cceSStefano Zampini const PetscInt* tisindices; 4293779c1cceSStefano Zampini 4294779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4295779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4296779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4297779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4298779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4299779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4300779c1cceSStefano Zampini } 4301779c1cceSStefano Zampini } 4302779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4303779c1cceSStefano Zampini 4304779c1cceSStefano Zampini 43056e683305SStefano Zampini if (pcbddc->dbg_flag) { 43066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 43086e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 43096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4310779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 43116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 43126e683305SStefano Zampini } 43136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 43146e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43156e683305SStefano Zampini } 43166e683305SStefano Zampini /* shift the pattern on coarse candidates */ 43176e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 43186e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4319854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 43206e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 43216e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 43226e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 43236e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 43246e683305SStefano Zampini } 43256e683305SStefano Zampini if (pcbddc->dbg_flag) { 43266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 43286e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 43296e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43306e683305SStefano Zampini } 4331779c1cceSStefano Zampini } 43326e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 433353a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 433453a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 433553a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 433653a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 433753a05cb3SStefano Zampini } 43386e683305SStefano Zampini } else { 43396e683305SStefano Zampini if (pcbddc->dbg_flag) { 43406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 43426e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43436e683305SStefano Zampini } 43446e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 43456e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 43466e683305SStefano Zampini } 43476e683305SStefano Zampini 43486e683305SStefano Zampini /* create local to global scatters for coarse problem */ 434968457ee5SStefano Zampini if (compute_vecs) { 43506e683305SStefano Zampini PetscInt lrows; 43516e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43526e683305SStefano Zampini if (coarse_mat_is) { 43536e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 43546e683305SStefano Zampini } else { 43556e683305SStefano Zampini lrows = 0; 43566e683305SStefano Zampini } 43576e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 43586e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 43596e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 43606e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43616e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 43626e683305SStefano Zampini } 43636e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 43646e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4365c8587f34SStefano Zampini 4366f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4367f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4368f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4369f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4370f9eb5b7dSStefano Zampini } else { 4371f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4372f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4373c8587f34SStefano Zampini } 4374c8587f34SStefano Zampini 43756e683305SStefano Zampini /* print some info if requested */ 43766e683305SStefano Zampini if (pcbddc->dbg_flag) { 43776e683305SStefano Zampini if (!multilevel_allowed) { 43786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43796e683305SStefano Zampini if (multilevel_requested) { 43806e683305SStefano 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); 43816e683305SStefano Zampini } else if (pcbddc->max_levels) { 43826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43836e683305SStefano Zampini } 43846e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43856e683305SStefano Zampini } 43866e683305SStefano Zampini } 43876e683305SStefano Zampini 4388f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43896e683305SStefano Zampini if (coarse_mat_is) { 43906e683305SStefano Zampini MatReuse coarse_mat_reuse; 43916a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43926e683305SStefano Zampini if (pcbddc->dbg_flag) { 43936e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43946e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43956e683305SStefano Zampini } 4396f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4397312be037SStefano Zampini char prefix[256],str_level[16]; 4398e604994aSStefano Zampini size_t len; 43996e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4400422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4401c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4402f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 44035f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4404c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 44056e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4406c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4407c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4408e604994aSStefano Zampini /* prefix */ 4409e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4410e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4411e604994aSStefano Zampini if (!pcbddc->current_level) { 4412e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4413e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4414c8587f34SStefano Zampini } else { 4415e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4416312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4417312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 441834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4419312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4420e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4421e604994aSStefano Zampini } 4422e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 44233e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 44243e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 44253e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 44263e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4427f9eb5b7dSStefano Zampini /* allow user customization */ 4428f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 44293e3c6dadSStefano Zampini } 44303e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 443151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 44323e3c6dadSStefano Zampini if (nisdofs) { 44333e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 44343e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 44353e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 44363e3c6dadSStefano Zampini } 44373e3c6dadSStefano Zampini } 44383e3c6dadSStefano Zampini if (nisneu) { 44393e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 44403e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4441312be037SStefano Zampini } 4442f9eb5b7dSStefano Zampini 4443f9eb5b7dSStefano Zampini /* get some info after set from options */ 4444f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4445f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 44464f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 44476e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4448f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4449f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4450f9eb5b7dSStefano Zampini } 445139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 44524f3a063dSStefano Zampini if (isredundant) { 44534f3a063dSStefano Zampini KSP inner_ksp; 44544f3a063dSStefano Zampini PC inner_pc; 44554f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 44564f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 44574f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 44584f3a063dSStefano Zampini } 4459f9eb5b7dSStefano Zampini 4460f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4461fa7f1dd8SStefano Zampini if (coarse_reuse) { 446281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4463fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 44646e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4465fa7f1dd8SStefano Zampini } else { 44666e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4467fa7f1dd8SStefano Zampini } 4468c8587f34SStefano Zampini if (isbddc || isnn) { 446922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 447070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4471b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 447222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 44736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 44756e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44766e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 447722b6e8a2SStefano Zampini } 447870cf5478SStefano Zampini } 447953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 448070cf5478SStefano Zampini } else { 448122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 448222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 448322bc73bbSStefano Zampini } 448422bc73bbSStefano Zampini } else { 44852e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4486c8587f34SStefano Zampini } 4487c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4488c8587f34SStefano Zampini 44893301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44905a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44913301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44923301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44933301b35fSStefano Zampini } 44943301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44953301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44963301b35fSStefano Zampini } 44973301b35fSStefano Zampini if (pc->pmat->spd_set) { 44983301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44993301b35fSStefano Zampini } 45006e683305SStefano Zampini /* set operators */ 45015f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 45026e683305SStefano Zampini if (pcbddc->dbg_flag) { 45036e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 45046e683305SStefano Zampini } 45056e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 45066e683305SStefano Zampini coarse_mat = 0; 45076e683305SStefano Zampini } 45086e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4509b9b85e73SStefano Zampini #if 0 4510b9b85e73SStefano Zampini { 4511b9b85e73SStefano Zampini PetscViewer viewer; 4512b9b85e73SStefano Zampini char filename[256]; 4513b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4514b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4515b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4516b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4517b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4518b9b85e73SStefano Zampini } 4519b9b85e73SStefano Zampini #endif 4520c8587f34SStefano Zampini 4521c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4522298c0119SStefano Zampini #if 0 4523c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4524c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 452598a51de6SStefano Zampini } 4526298c0119SStefano Zampini #endif 452798a51de6SStefano Zampini 452898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 452998a51de6SStefano Zampini Vec crhs,csol; 453098a51de6SStefano Zampini PetscBool ispreonly; 453104708bb6SStefano Zampini 453298a51de6SStefano Zampini if (CoarseNullSpace) { 4533c8587f34SStefano Zampini if (isbddc) { 4534c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4535c8587f34SStefano Zampini } else { 45365fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4537c8587f34SStefano Zampini } 4538c8587f34SStefano Zampini } 4539f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4540f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4541f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4542f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 45436e683305SStefano Zampini /* hack */ 4544f347579bSStefano Zampini if (!csol) { 45452a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4546f9eb5b7dSStefano Zampini } 4547f347579bSStefano Zampini if (!crhs) { 45482a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4549f347579bSStefano Zampini } 4550cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4551cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 45526e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4553c8587f34SStefano Zampini KSP check_ksp; 45542b510759SStefano Zampini KSPType check_ksp_type; 4555c8587f34SStefano Zampini PC check_pc; 45566e683305SStefano Zampini Vec check_vec,coarse_vec; 45576a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 45582b510759SStefano Zampini PetscInt its; 45596e683305SStefano Zampini PetscBool compute_eigs; 45606e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 45616e683305SStefano Zampini PetscInt neigs; 45628e185a42SStefano Zampini const char *prefix; 4563c8587f34SStefano Zampini 45642b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 45656e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4566422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 456723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4568f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 45692b510759SStefano Zampini if (ispreonly) { 45702b510759SStefano Zampini check_ksp_type = KSPPREONLY; 45716e683305SStefano Zampini compute_eigs = PETSC_FALSE; 45722b510759SStefano Zampini } else { 4573cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 45746e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4575c8587f34SStefano Zampini } 4576c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45776e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45786e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45796e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4580a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4581a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4582a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4583a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4584c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4585c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4586c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4587c8587f34SStefano Zampini /* create random vec */ 45886e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45896e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4590c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4591c8587f34SStefano Zampini if (CoarseNullSpace) { 4592c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4593c8587f34SStefano Zampini } 45946e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4595c8587f34SStefano Zampini /* solve coarse problem */ 45966e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4597c8587f34SStefano Zampini if (CoarseNullSpace) { 45986e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4599c8587f34SStefano Zampini } 4600cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 46016e683305SStefano Zampini if (compute_eigs) { 4602854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4603854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 46046e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 46056e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 46066e683305SStefano Zampini lambda_min = eigs_r[0]; 46076e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 46086e683305SStefano Zampini if (lambda_max>lambda_min) { 4609cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4610cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4611cbcc2c2aSStefano Zampini } 4612c8587f34SStefano Zampini } 4613c8587f34SStefano Zampini } 4614cbcc2c2aSStefano Zampini 4615c8587f34SStefano Zampini /* check coarse problem residual error */ 46166e683305SStefano Zampini if (pcbddc->dbg_flag) { 46176e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 46186e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 46196e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4620c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 46216e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 46226e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4623c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4624779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 46256e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 46266e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 46276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 46286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 46296e683305SStefano Zampini if (compute_eigs) { 46306e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4631deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4632c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 46336e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 46346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem eigenvalues (estimated with %d iterations of %s): %1.6e %1.6e (%1.6e %1.6e)\n",its,check_ksp_type,lambda_min,lambda_max,lambda_min_s,lambda_max_s);CHKERRQ(ierr); 46356e683305SStefano Zampini for (i=0;i<neigs;i++) { 46366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4637c8587f34SStefano Zampini } 46386e683305SStefano Zampini } 46396e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 46406e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 46416e683305SStefano Zampini } 4642c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 46436e683305SStefano Zampini if (compute_eigs) { 46446e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 46456e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4646c8587f34SStefano Zampini } 46476e683305SStefano Zampini } 46486e683305SStefano Zampini } 4649cbcc2c2aSStefano Zampini /* print additional info */ 4650cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 46516e683305SStefano Zampini /* waits until all processes reaches this point */ 46526e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4653cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4654cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4655cbcc2c2aSStefano Zampini } 4656cbcc2c2aSStefano Zampini 46572b510759SStefano Zampini /* free memory */ 4658c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4659fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4660c8587f34SStefano Zampini PetscFunctionReturn(0); 4661c8587f34SStefano Zampini } 4662674ae819SStefano Zampini 4663f34684f1SStefano Zampini #undef __FUNCT__ 4664f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4665f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4666f34684f1SStefano Zampini { 4667f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4668f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4669f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4670dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4671dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 467273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4673dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4674f34684f1SStefano Zampini PetscErrorCode ierr; 4675f34684f1SStefano Zampini 4676f34684f1SStefano Zampini PetscFunctionBegin; 4677f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46780e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 46790e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4680727cdba6SStefano Zampini } 4681dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 46823bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4683dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4684dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4685dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4686dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4687dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4688dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46890e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 46900e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 46910e6343abSStefano Zampini } 4692dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4693dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4694dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4695dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4696dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4697f34684f1SStefano Zampini 4698f34684f1SStefano Zampini /* check numbering */ 4699f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4700f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4701dc456d91SStefano Zampini PetscInt i; 4702b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4703f34684f1SStefano Zampini 4704f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4705f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4706f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 47070fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4708f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4709f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4710727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4711f34684f1SStefano Zampini } 4712f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4713f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4714f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4715e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4716e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4717e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4718e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4719f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4720f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4721f34684f1SStefano Zampini if (array[i] == 1.0) { 4722b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4723f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4724f34684f1SStefano Zampini } 4725f34684f1SStefano Zampini } 4726b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4727f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4728f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4729f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4730f34684f1SStefano Zampini } 4731f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4732f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4733e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4734e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4735f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4736f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4737b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4738ca8b9ea9SStefano Zampini PetscInt *gidxs; 4739ca8b9ea9SStefano Zampini 4740ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 47413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4742f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4743f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4744f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4745f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 47464bc2dc4bSStefano 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); 4747f34684f1SStefano Zampini } 4748f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4749ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4750f34684f1SStefano Zampini } 4751f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4752302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4753f34684f1SStefano Zampini } 47548bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4755f34684f1SStefano Zampini /* get back data */ 4756f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4757f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4758674ae819SStefano Zampini PetscFunctionReturn(0); 4759674ae819SStefano Zampini } 4760674ae819SStefano Zampini 4761e456f2a8SStefano Zampini #undef __FUNCT__ 4762e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4763a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4764e456f2a8SStefano Zampini { 4765e456f2a8SStefano Zampini IS localis_t; 4766a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4767e456f2a8SStefano Zampini PetscScalar *vals; 4768e456f2a8SStefano Zampini PetscErrorCode ierr; 4769e456f2a8SStefano Zampini 4770e456f2a8SStefano Zampini PetscFunctionBegin; 4771a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4772e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4773854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4774e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4775e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4776a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4777a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47781035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4779a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47801035eff8SStefano Zampini } 4781a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4782e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4783e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4784a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4785a7dc3881SStefano Zampini /* now compute set in local ordering */ 4786a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4787a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4788a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4789a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4790a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4791ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4792e456f2a8SStefano Zampini lsize++; 4793e456f2a8SStefano Zampini } 4794e456f2a8SStefano Zampini } 4795854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4796a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4797ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4798e456f2a8SStefano Zampini idxs[lsize++] = i; 4799e456f2a8SStefano Zampini } 4800e456f2a8SStefano Zampini } 4801a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4802a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4803e456f2a8SStefano Zampini *localis = localis_t; 4804e456f2a8SStefano Zampini PetscFunctionReturn(0); 4805e456f2a8SStefano Zampini } 4806906d46d4SStefano Zampini 4807906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4808906d46d4SStefano Zampini #undef __FUNCT__ 4809906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4810906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4811906d46d4SStefano Zampini { 4812906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4813906d46d4SStefano Zampini PetscErrorCode ierr; 4814906d46d4SStefano Zampini 4815906d46d4SStefano Zampini PetscFunctionBegin; 4816906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4817906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4818906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4819906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4820906d46d4SStefano Zampini PetscFunctionReturn(0); 4821906d46d4SStefano Zampini } 4822906d46d4SStefano Zampini 4823906d46d4SStefano Zampini #undef __FUNCT__ 4824906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4825906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4826906d46d4SStefano Zampini { 4827906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4828906d46d4SStefano Zampini PetscErrorCode ierr; 4829906d46d4SStefano Zampini 4830906d46d4SStefano Zampini PetscFunctionBegin; 4831906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4832906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4833906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4834906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4835906d46d4SStefano Zampini PetscFunctionReturn(0); 4836906d46d4SStefano Zampini } 4837b96c3477SStefano Zampini 4838b96c3477SStefano Zampini #undef __FUNCT__ 4839b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 484008122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4841b96c3477SStefano Zampini { 4842a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4843b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4844b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4845a64f4aa4SStefano Zampini Mat S_j; 4846b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4847b96c3477SStefano Zampini PetscBool free_used_adj; 4848b96c3477SStefano Zampini PetscErrorCode ierr; 4849b96c3477SStefano Zampini 4850b96c3477SStefano Zampini PetscFunctionBegin; 4851b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4852b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 485308122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4854b96c3477SStefano Zampini used_xadj = NULL; 4855b96c3477SStefano Zampini used_adjncy = NULL; 4856b96c3477SStefano Zampini } else { 485708122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 485808122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 485908122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 486008122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4861b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4862b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4863b96c3477SStefano Zampini } else { 48642fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4865b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4866b96c3477SStefano Zampini PetscInt nvtxs; 4867b96c3477SStefano Zampini 48682fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 48692fffb893SStefano Zampini if (flg_row) { 4870b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4871b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4872b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4873b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 48742fffb893SStefano Zampini } else { 48752fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 48762fffb893SStefano Zampini used_xadj = NULL; 48772fffb893SStefano Zampini used_adjncy = NULL; 48782fffb893SStefano Zampini } 48792fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4880b96c3477SStefano Zampini } 4881b96c3477SStefano Zampini } 4882d5574798SStefano Zampini 4883d5574798SStefano Zampini /* setup sub_schurs data */ 4884a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4885a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4886a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4887a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 488806a4e24aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE);CHKERRQ(ierr); 4889a64f4aa4SStefano Zampini } else { 48906816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 489104708bb6SStefano Zampini PetscBool isseqaij; 48925feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48935feab87aSStefano Zampini PetscInt n_vertices; 48945feab87aSStefano Zampini 48955feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48962034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48975feab87aSStefano Zampini } 489804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 489904708bb6SStefano Zampini if (!isseqaij) { 490004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 490104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 490204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 490304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 490404708bb6SStefano Zampini } else { 490504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 490604708bb6SStefano Zampini } 490704708bb6SStefano Zampini } 490806a4e24aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point);CHKERRQ(ierr); 4909a64f4aa4SStefano Zampini } 4910a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4911b96c3477SStefano Zampini 4912b96c3477SStefano Zampini /* free adjacency */ 4913b96c3477SStefano Zampini if (free_used_adj) { 4914b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4915b96c3477SStefano Zampini } 4916b96c3477SStefano Zampini PetscFunctionReturn(0); 4917b96c3477SStefano Zampini } 4918b96c3477SStefano Zampini 4919b96c3477SStefano Zampini #undef __FUNCT__ 4920b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 492108122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4922b96c3477SStefano Zampini { 4923b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4924b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4925b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4926b96c3477SStefano Zampini PCBDDCGraph graph; 4927b96c3477SStefano Zampini PetscErrorCode ierr; 4928b96c3477SStefano Zampini 4929b96c3477SStefano Zampini PetscFunctionBegin; 4930b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 493108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 49323301b35fSStefano Zampini IS verticesIS,verticescomm; 49333301b35fSStefano Zampini PetscInt vsize,*idxs; 4934b96c3477SStefano Zampini 4935b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 49363301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 49373301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 49383301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 49393301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 49403301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4941b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 49427fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 49433301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 49443301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4945b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4946b96c3477SStefano Zampini /* 4947b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4948b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4949b96c3477SStefano Zampini } 4950b96c3477SStefano Zampini */ 4951b96c3477SStefano Zampini } else { 4952b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4953b96c3477SStefano Zampini } 4954b96c3477SStefano Zampini 4955b96c3477SStefano Zampini /* sub_schurs init */ 4956a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4957a64f4aa4SStefano Zampini 4958b96c3477SStefano Zampini /* free graph struct */ 495908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4960b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4961b96c3477SStefano Zampini } 4962b96c3477SStefano Zampini PetscFunctionReturn(0); 4963b96c3477SStefano Zampini } 4964