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); 477b026bddSStefano 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; 525*019a44ceSStefano Zampini PetscInt n_constraints,n_R,old_size,n_benign; 5266bfb1811SStefano Zampini PetscErrorCode ierr; 5276bfb1811SStefano Zampini 5286bfb1811SStefano Zampini PetscFunctionBegin; 529f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 530*019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 531f4ddd8eeSStefano Zampini } 532e7b262bdSStefano Zampini /* get sizes */ 533*019a44ceSStefano Zampini n_benign = (PetscInt)(!!pcbddc->zerodiag); 534*019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - pcbddc->n_vertices; 535*019a44ceSStefano Zampini if (n_benign && n_constraints) { 536*019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot handle benign susbspace trick without change of basis"); 537*019a44ceSStefano Zampini } 538b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 5396bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 540e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 541e7b262bdSStefano Zampini /* R nodes */ 542e7b262bdSStefano Zampini old_size = -1; 543e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 544e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 545e7b262bdSStefano Zampini } 546e7b262bdSStefano Zampini if (n_R != old_size) { 547e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 548e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5496bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5506bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5516bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5526bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 553e7b262bdSStefano Zampini } 554e7b262bdSStefano Zampini /* local primal dofs */ 555e7b262bdSStefano Zampini old_size = -1; 556e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 557e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 558e7b262bdSStefano Zampini } 559e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 560e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 56183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 562e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5636bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 564e7b262bdSStefano Zampini } 565e7b262bdSStefano Zampini /* local explicit constraints */ 566e7b262bdSStefano Zampini old_size = -1; 567e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 568e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 569e7b262bdSStefano Zampini } 570e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 571e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 57283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 57383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 57483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 57583b7ccabSStefano Zampini } 5766bfb1811SStefano Zampini PetscFunctionReturn(0); 5776bfb1811SStefano Zampini } 5786bfb1811SStefano Zampini 5796bfb1811SStefano Zampini #undef __FUNCT__ 58047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 58147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 58288ebb749SStefano Zampini { 58325084f0cSStefano Zampini PetscErrorCode ierr; 58425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 58588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 58688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 587d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 58825084f0cSStefano Zampini /* submatrices of local problem */ 58980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 590*019a44ceSStefano Zampini /* submatrices of benign trick */ 591*019a44ceSStefano Zampini Mat B0_R,B0_V; 59206656605SStefano Zampini /* submatrices of local coarse problem */ 59306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 59425084f0cSStefano Zampini /* working matrices */ 59506656605SStefano Zampini Mat C_CR; 59625084f0cSStefano Zampini /* additional working stuff */ 59706656605SStefano Zampini PC pc_R; 59806656605SStefano Zampini Mat F; 599*019a44ceSStefano Zampini Vec vec_R_benign; 60006656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 60106656605SStefano Zampini 60225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 60306656605SStefano Zampini PetscScalar *work; 60406656605SStefano Zampini PetscInt *idx_V_B; 605*019a44ceSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign; 60606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 607b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 60845a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 60988ebb749SStefano Zampini MatType impMatType; 61025084f0cSStefano Zampini /* some shortcuts to scalars */ 61106656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 61288ebb749SStefano Zampini 61388ebb749SStefano Zampini PetscFunctionBegin; 614*019a44ceSStefano Zampini n_benign = (PetscInt)(!!pcbddc->zerodiag); 615b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 616*019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - n_vertices; 61788ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 618b371cd4fSStefano Zampini n_B = pcis->n_B; 619b371cd4fSStefano Zampini n_D = pcis->n - n_B; 62088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 62188ebb749SStefano Zampini 62288ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 62388ebb749SStefano Zampini impMatType = MATSEQDENSE; 62488ebb749SStefano Zampini 62588ebb749SStefano Zampini /* vertices in boundary numbering */ 626785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 6270e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 62888ebb749SStefano Zampini if (i != n_vertices) { 62922d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 63088ebb749SStefano Zampini } 63188ebb749SStefano Zampini 63206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 633*019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 634*019a44ceSStefano Zampini if (!n_benign) { 63506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 63606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 63706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 63806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 63906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 64006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 64106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 64206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 643*019a44ceSStefano Zampini } else { /* these wrappers are not needed with the benign trick */ 644*019a44ceSStefano Zampini S_VV = S_CV = S_VC = S_CC = NULL; 645*019a44ceSStefano Zampini } 64606656605SStefano Zampini 64706656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 64806656605SStefano Zampini /* allocate workspace */ 64906656605SStefano Zampini n = 0; 65006656605SStefano Zampini if (n_constraints) { 65106656605SStefano Zampini n += n_R*n_constraints; 65206656605SStefano Zampini } 65306656605SStefano Zampini if (n_vertices) { 65406656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 65580677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 65606656605SStefano Zampini } 6573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 65806656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 65906656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 66006656605SStefano Zampini } 66106656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 66206656605SStefano Zampini 66306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 66406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 66506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 66606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 66706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 66806656605SStefano Zampini if (isLU || isILU || isCHOL) { 66906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 670d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 671d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 672d62866d3SStefano Zampini MatFactorType type; 673d62866d3SStefano Zampini 6746816873aSStefano Zampini F = reuse_mumps->F; 6756816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 676d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 67706656605SStefano Zampini } else { 67806656605SStefano Zampini F = NULL; 67906656605SStefano Zampini } 68006656605SStefano Zampini 68188ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 68288ebb749SStefano Zampini if (n_constraints) { 68306656605SStefano Zampini Mat M1,M2,M3; 68480677318SStefano Zampini Mat auxmat; 68506656605SStefano Zampini IS is_aux; 68680677318SStefano Zampini PetscScalar *array,*array2; 68706656605SStefano Zampini 688f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 68980677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 69088ebb749SStefano Zampini 69125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 69225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6938ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 69480677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 69588ebb749SStefano Zampini 69680677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 69780677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 69806656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 69988ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 70006656605SStefano Zampini const PetscScalar *row_cmat_values; 70106656605SStefano Zampini const PetscInt *row_cmat_indices; 70206656605SStefano Zampini PetscInt size_of_constraint,j; 70388ebb749SStefano Zampini 70406656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 70506656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 70606656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 70706656605SStefano Zampini } 70806656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 70906656605SStefano Zampini } 71080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 71106656605SStefano Zampini if (F) { 71206656605SStefano Zampini Mat B; 71306656605SStefano Zampini 71406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 71580677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 71606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 71706656605SStefano Zampini } else { 71880677318SStefano Zampini PetscScalar *marr; 71980677318SStefano Zampini 72080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 72106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 72206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 72380677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 72406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 72506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 72606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 72706656605SStefano Zampini } 72880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 72906656605SStefano Zampini } 73080677318SStefano Zampini if (!pcbddc->switch_static) { 73180677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 73280677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 73380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 73480677318SStefano Zampini for (i=0;i<n_constraints;i++) { 73580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 73680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 73780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 73880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 73980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 74080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 74180677318SStefano Zampini } 74280677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 74380677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 74480677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 74580677318SStefano Zampini } else { 74680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 74780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 74825084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 74980677318SStefano Zampini } 75080677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 75180677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 75280677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 75306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 75406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 75580677318SStefano Zampini if (isCHOL) { 75680677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 75780677318SStefano Zampini } else { 75825084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 75980677318SStefano Zampini } 76080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 76106656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 76225084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 76325084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 76425084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 76580677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 76680677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 76780677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 76806656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 76906656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 770f4ddd8eeSStefano Zampini } 77188ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 77288ebb749SStefano Zampini if (n_vertices) { 77306656605SStefano Zampini IS is_aux; 7743a50541eSStefano Zampini 7756816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7766816873aSStefano Zampini IS tis; 7776816873aSStefano Zampini 7786816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7796816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7806816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7816816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7826816873aSStefano Zampini } else { 7833a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7846816873aSStefano Zampini } 7859577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7869577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 78704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 788*019a44ceSStefano Zampini if (n_benign) { 789*019a44ceSStefano Zampini Mat B0; 790*019a44ceSStefano Zampini IS dummy; 791*019a44ceSStefano Zampini PetscInt ii[2]; 792*019a44ceSStefano Zampini 793*019a44ceSStefano Zampini ii[0] = 0; 794*019a44ceSStefano Zampini ii[1] = pcbddc->B0_ncol; 795*019a44ceSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 796*019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 797*019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 798*019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 799*019a44ceSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 800*019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 801*019a44ceSStefano Zampini } 80225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 80388ebb749SStefano Zampini } 80488ebb749SStefano Zampini 80588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 806f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 80706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 80806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 80906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 81006656605SStefano Zampini } 811f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 81206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 81306656605SStefano Zampini PetscScalar *marray; 81406656605SStefano Zampini 81506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 81606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 817f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 818f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 819f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 820f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 821f4ddd8eeSStefano Zampini } 822f4ddd8eeSStefano Zampini } 82306656605SStefano Zampini 824f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 82506656605SStefano Zampini PetscScalar *marray; 82688ebb749SStefano Zampini 82706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8288eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 82906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 83088ebb749SStefano Zampini } 8313301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 83206656605SStefano Zampini n *= 2; 83388ebb749SStefano Zampini } 83406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 83606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8378eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 83806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 83906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 84088ebb749SStefano Zampini } 8413301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 84206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 8438eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 84406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 84506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 84688ebb749SStefano Zampini } 84788ebb749SStefano Zampini } else { 848c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 849c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8501b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 851c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 852c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 853c0553b1fSStefano Zampini } 85488ebb749SStefano Zampini } 85506656605SStefano Zampini } 856*019a44ceSStefano Zampini 85706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 85806656605SStefano Zampini /* vertices */ 85906656605SStefano Zampini if (n_vertices) { 86016f15bc4SStefano Zampini 86104708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 86204708bb6SStefano Zampini 86316f15bc4SStefano Zampini if (n_R) { 86406656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 865*019a44ceSStefano Zampini Mat S_V_benign; 86606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 86716f15bc4SStefano Zampini PetscScalar *x,*y; 86804708bb6SStefano Zampini PetscBool isseqaij; 86906656605SStefano Zampini 87021eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 87204708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 8736816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 87406656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 87506656605SStefano Zampini } else { 87606656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 87706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 87806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 87906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 88006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 88106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 88206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 88306656605SStefano Zampini } 88406656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 88506656605SStefano Zampini } 886*019a44ceSStefano Zampini if (n_benign) { 887*019a44ceSStefano Zampini const PetscScalar *vals; 888*019a44ceSStefano Zampini const PetscInt *idxs; 889*019a44ceSStefano Zampini PetscInt n; 890*019a44ceSStefano Zampini 891*019a44ceSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign);CHKERRQ(ierr); 892*019a44ceSStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 893*019a44ceSStefano Zampini ierr = MatGetRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 894*019a44ceSStefano Zampini ierr = VecSetValues(pcbddc->vec1_R,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 895*019a44ceSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_R);CHKERRQ(ierr); 896*019a44ceSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_R);CHKERRQ(ierr); 897*019a44ceSStefano Zampini ierr = MatRestoreRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 898*019a44ceSStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,vec_R_benign);CHKERRQ(ierr); 899*019a44ceSStefano Zampini ierr = VecScale(vec_R_benign,m_one);CHKERRQ(ierr); 900*019a44ceSStefano Zampini } 90180677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 90206656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 90306656605SStefano Zampini if (n_constraints) { 90406656605SStefano Zampini Mat B; 90580677318SStefano Zampini 906b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 90780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 90880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 90980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 91080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91480677318SStefano Zampini } 91580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 91680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 91780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 91980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 92006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 92106656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 92206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 92306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 92406656605SStefano Zampini } 92504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 92604708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 92704708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 92804708bb6SStefano Zampini } 92906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 930*019a44ceSStefano Zampini if (n_benign) { 931*019a44ceSStefano Zampini ierr = MatMatMult(B0_R,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_V_benign);CHKERRQ(ierr); 932*019a44ceSStefano Zampini } 93380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 93606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 93706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 93806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 93906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 940*019a44ceSStefano Zampini if (!n_benign) { 94106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 942*019a44ceSStefano Zampini } else { 943*019a44ceSStefano Zampini const PetscScalar *vals; 944*019a44ceSStefano Zampini PetscScalar val; 945*019a44ceSStefano Zampini const PetscInt *idxs; 946*019a44ceSStefano Zampini PetscInt n,j; 947*019a44ceSStefano Zampini 948*019a44ceSStefano Zampini /* Use explicit loops to populate the local coarse matrix */ 949*019a44ceSStefano Zampini ierr = MatDenseGetArray(S_VVt,&x);CHKERRQ(ierr); 950*019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 951*019a44ceSStefano Zampini for (i=0;i<n_vertices;i++) { 952*019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+i] = x[j*n_vertices+i]; 953*019a44ceSStefano Zampini } 954*019a44ceSStefano Zampini } 955*019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&x);CHKERRQ(ierr); 956*019a44ceSStefano Zampini ierr = MatDenseGetArray(S_V_benign,&x);CHKERRQ(ierr); 957*019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 958*019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 959*019a44ceSStefano Zampini coarse_submat_vals[n_vertices*pcbddc->local_primal_size+j] = x[j]+vals[j]; 960*019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices] = x[j]+vals[j]; 961*019a44ceSStefano Zampini } 962*019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 963*019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_V_benign,&x);CHKERRQ(ierr); 964*019a44ceSStefano Zampini ierr = MatDestroy(&S_V_benign);CHKERRQ(ierr); 965*019a44ceSStefano Zampini ierr = VecDot(vec_R_benign,vec_R_benign,&val);CHKERRQ(ierr); 966*019a44ceSStefano Zampini coarse_submat_vals[pcbddc->local_primal_size*pcbddc->local_primal_size-1] = -val; 967*019a44ceSStefano Zampini } 96806656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 96916f15bc4SStefano Zampini } else { 970*019a44ceSStefano Zampini /* TODO benign case also here */ 971*019a44ceSStefano Zampini if (n_benign) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not yet implemented"); 97216f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 97316f15bc4SStefano Zampini } 97421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 97506656605SStefano Zampini /* coarse basis functions */ 97606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 97716f15bc4SStefano Zampini PetscScalar *y; 97816f15bc4SStefano Zampini 97906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 98006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 98106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 98506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 98606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 98706656605SStefano Zampini 98806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 98906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 99506656605SStefano Zampini } 99606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 99706656605SStefano Zampini } 998*019a44ceSStefano Zampini if (n_benign) { 999*019a44ceSStefano Zampini PetscScalar *y; 1000*019a44ceSStefano Zampini 1001*019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1002*019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*n_vertices);CHKERRQ(ierr); 1003*019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1004*019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1005*019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1006*019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 1007*019a44ceSStefano Zampini 1008*019a44ceSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1009*019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1010*019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 1011*019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1012*019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1013*019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1014*019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1015*019a44ceSStefano Zampini } 1016*019a44ceSStefano Zampini } 101704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 101804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 101906656605SStefano Zampini } 102006656605SStefano Zampini 102106656605SStefano Zampini if (n_constraints) { 102206656605SStefano Zampini Mat B; 102306656605SStefano Zampini 102406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 102506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 102680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 102706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 102806656605SStefano Zampini if (n_vertices) { 102980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 103080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 103180677318SStefano Zampini } else { 103280677318SStefano Zampini Mat S_VCt; 103380677318SStefano Zampini 103480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 103580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 103680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 103780677318SStefano Zampini } 103806656605SStefano Zampini } 103906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 104006656605SStefano Zampini /* coarse basis functions */ 104106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 104206656605SStefano Zampini PetscScalar *y; 104306656605SStefano Zampini 104406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 104506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 104606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 104706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 104806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 104906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 105006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 105106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 105206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 105306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 105406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 105506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 105606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 105706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 105806656605SStefano Zampini } 105906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 106006656605SStefano Zampini } 106106656605SStefano Zampini } 106280677318SStefano Zampini if (n_constraints) { 106380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 106480677318SStefano Zampini } 106506656605SStefano Zampini 1066*019a44ceSStefano Zampini if (n_benign) { 1067*019a44ceSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1068*019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1069*019a44ceSStefano Zampini ierr = VecDestroy(&vec_R_benign);CHKERRQ(ierr); 1070*019a44ceSStefano Zampini } 1071*019a44ceSStefano Zampini 107206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 10733301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 107406656605SStefano Zampini 107506656605SStefano Zampini if (n_constraints) { 107616f15bc4SStefano Zampini Mat S_CCT,B_C; 107706656605SStefano Zampini 107880677318SStefano Zampini /* this is a lazy thing */ 107980677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 108006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 108206656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 108316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 108406656605SStefano Zampini if (n_vertices) { 108516f15bc4SStefano Zampini Mat B_V,S_VCT; 108606656605SStefano Zampini 108706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 108806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 108906656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 109006656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 109116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 109206656605SStefano Zampini } 109306656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 109480677318SStefano Zampini } else { /* if there are no constraints, reset work */ 109580677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 109606656605SStefano Zampini } 109716f15bc4SStefano Zampini if (n_vertices && n_R) { 109806656605SStefano Zampini Mat A_VRT; 109980677318SStefano Zampini PetscScalar *marray; 110006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 110106656605SStefano Zampini 110280677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 110380677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 110480677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 110506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 110680677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 110780677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 110816f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 110906656605SStefano Zampini } 111006656605SStefano Zampini 111106656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 111206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 111306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 111406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 111506656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 111606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 111706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 111806656605SStefano Zampini } 111906656605SStefano Zampini } else { 112006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 112106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 112206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 112306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 112406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 112506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 112606656605SStefano Zampini } 112706656605SStefano Zampini } 112806656605SStefano Zampini /* coarse basis functions */ 112906656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 113006656605SStefano Zampini PetscScalar *y; 113106656605SStefano Zampini 113206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 113306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 113406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 113506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 113606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 113706656605SStefano Zampini if (i<n_vertices) { 113806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 113906656605SStefano Zampini } 114006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 114106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 114206656605SStefano Zampini 114306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 114406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 114506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 114706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 114906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 115006656605SStefano Zampini } 115106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 115206656605SStefano Zampini } 115306656605SStefano Zampini } 1154d62866d3SStefano Zampini /* free memory */ 115588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 115606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 115806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 115906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1160d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1161d62866d3SStefano Zampini if (n_vertices) { 1162d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1163d62866d3SStefano Zampini } 1164d62866d3SStefano Zampini if (n_constraints) { 1165d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1166d62866d3SStefano Zampini } 116788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 116888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 116988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1170*019a44ceSStefano Zampini //if (pcbddc->dbg_flag) { 1171*019a44ceSStefano Zampini if (0) { 117288ebb749SStefano Zampini Mat coarse_sub_mat; 117325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 117488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 117588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 117688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 11778bec7fa6SStefano Zampini Mat C_B,CPHI; 11788bec7fa6SStefano Zampini IS is_dummy; 11798bec7fa6SStefano Zampini Vec mones; 118088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 118188ebb749SStefano Zampini PetscReal real_value; 118288ebb749SStefano Zampini 118388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 118488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 118588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 118688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 118788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 118888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1189c0553b1fSStefano Zampini if (unsymmetric_check) { 119088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 119188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 119288ebb749SStefano Zampini } 119388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 119488ebb749SStefano Zampini 119525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 11963301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 119725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1198c0553b1fSStefano Zampini if (unsymmetric_check) { 119988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 120188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 120288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 120488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 120588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 120788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 120888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 121088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 121188ebb749SStefano Zampini } else { 121288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 121388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 121488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 121588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 121688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 121788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 121888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 121988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 122088ebb749SStefano Zampini } 122188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 122288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 122388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 122488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 122581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 12268bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 12270fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 122806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 12298bec7fa6SStefano Zampini 12308bec7fa6SStefano Zampini /* check constraints */ 12318bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 12328bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 12338bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 12348bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 12358bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 12368bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 12378bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1238bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1239c0553b1fSStefano Zampini if (unsymmetric_check) { 1240bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1241bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1242bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1243bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1244bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 124588ebb749SStefano Zampini } 12468bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 12478bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 12488bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 12498bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 12508bec7fa6SStefano Zampini 125125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 125288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 125388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 125488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 125588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 125688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 125788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 125888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 125988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 126088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 126188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1262c0553b1fSStefano Zampini if (unsymmetric_check) { 126388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 126488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 126588ebb749SStefano Zampini } 126688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 126788ebb749SStefano Zampini } 12688629588bSStefano Zampini /* get back data */ 12698629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 127088ebb749SStefano Zampini PetscFunctionReturn(0); 127188ebb749SStefano Zampini } 127288ebb749SStefano Zampini 127388ebb749SStefano Zampini #undef __FUNCT__ 1274d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1275d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1276aa0d41d4SStefano Zampini { 1277d65f70fdSStefano Zampini Mat *work_mat; 1278d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1279d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1280d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1281aa0d41d4SStefano Zampini PetscErrorCode ierr; 1282aa0d41d4SStefano Zampini 1283aa0d41d4SStefano Zampini PetscFunctionBegin; 1284d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1285d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1286d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1287d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1288aa0d41d4SStefano Zampini 1289d65f70fdSStefano Zampini if (!rsorted) { 1290906d46d4SStefano Zampini const PetscInt *idxs; 1291906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1292aa0d41d4SStefano Zampini 1293d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1294d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1295d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1296d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1297aa0d41d4SStefano Zampini } 1298d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1299d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1300d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1301d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1302aa0d41d4SStefano Zampini } 1303d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1304d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1305d65f70fdSStefano Zampini } else { 1306d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1307d65f70fdSStefano Zampini isrow_s = isrow; 1308aa0d41d4SStefano Zampini } 1309906d46d4SStefano Zampini 1310d65f70fdSStefano Zampini if (!csorted) { 1311d65f70fdSStefano Zampini if (isrow == iscol) { 1312d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1313d65f70fdSStefano Zampini iscol_s = isrow_s; 1314d65f70fdSStefano Zampini } else { 1315d65f70fdSStefano Zampini const PetscInt *idxs; 1316d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1317906d46d4SStefano Zampini 1318d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1319d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1320d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1321d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1322d65f70fdSStefano Zampini } 1323d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1324d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1325d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1326d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1327d65f70fdSStefano Zampini } 1328d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1329d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1330d65f70fdSStefano Zampini } 1331d65f70fdSStefano Zampini } else { 1332d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1333d65f70fdSStefano Zampini iscol_s = iscol; 1334d65f70fdSStefano Zampini } 1335d65f70fdSStefano Zampini 1336d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1337d65f70fdSStefano Zampini 1338d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1339906d46d4SStefano Zampini Mat new_mat; 1340d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1341906d46d4SStefano Zampini 1342d65f70fdSStefano Zampini if (!rsorted) { 1343d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1344d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1345d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1346d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1347906d46d4SStefano Zampini } 1348d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1349d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1350d65f70fdSStefano Zampini } else { 1351d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1352906d46d4SStefano Zampini } 1353d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1354d65f70fdSStefano Zampini 1355d65f70fdSStefano Zampini if (!csorted) { 1356d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1357d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1358d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1359d65f70fdSStefano Zampini } else { 1360d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1361d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1362d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1363d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1364d65f70fdSStefano Zampini } 1365d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1366d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1367d65f70fdSStefano Zampini } 1368d65f70fdSStefano Zampini } else { 1369d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1370d65f70fdSStefano Zampini } 1371d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1372d65f70fdSStefano Zampini 1373d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1374d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1375d65f70fdSStefano Zampini work_mat[0] = new_mat; 1376d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1377d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1378d65f70fdSStefano Zampini } 1379d65f70fdSStefano Zampini 1380d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1381d65f70fdSStefano Zampini *B = work_mat[0]; 1382d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1383d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1384d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1385d65f70fdSStefano Zampini PetscFunctionReturn(0); 1386d65f70fdSStefano Zampini } 1387d65f70fdSStefano Zampini 1388d65f70fdSStefano Zampini #undef __FUNCT__ 13895e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 13905e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1391aa0d41d4SStefano Zampini { 1392aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 13933bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 13945e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1395d65f70fdSStefano Zampini Mat new_mat; 13965e8657edSStefano Zampini IS is_local,is_global; 1397d65f70fdSStefano Zampini PetscInt local_size; 1398d65f70fdSStefano Zampini PetscBool isseqaij; 1399aa0d41d4SStefano Zampini PetscErrorCode ierr; 1400aa0d41d4SStefano Zampini 1401aa0d41d4SStefano Zampini PetscFunctionBegin; 1402aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 14035e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 14045e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 14053bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1406aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1407d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1408aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1409906d46d4SStefano Zampini 1410906d46d4SStefano Zampini /* check */ 1411906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1412906d46d4SStefano Zampini Vec x,x_change; 1413906d46d4SStefano Zampini PetscReal error; 1414906d46d4SStefano Zampini 14155e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1416906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 14175e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1418e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1419e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1420d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1421e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1422e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1423906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1424906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1425906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1426906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1427906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1428906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1429906d46d4SStefano Zampini } 1430906d46d4SStefano Zampini 143122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 14329b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 143322d5777bSStefano Zampini if (isseqaij) { 1434d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1435aa0d41d4SStefano Zampini } else { 1436aa0d41d4SStefano Zampini Mat work_mat; 1437aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1438d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1439aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1440aa0d41d4SStefano Zampini } 14413301b35fSStefano Zampini if (matis->A->symmetric_set) { 14423301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1443e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 14443301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1445e496cd5dSStefano Zampini #endif 14463301b35fSStefano Zampini } 144745a1bb75SStefano Zampini /* 144845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1449d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 145045a1bb75SStefano Zampini */ 1451d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1452aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1453aa0d41d4SStefano Zampini } 1454aa0d41d4SStefano Zampini 1455aa0d41d4SStefano Zampini #undef __FUNCT__ 1456a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 14578ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1458a64d13efSStefano Zampini { 1459a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1460a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1461d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 146253892102SStefano Zampini PetscInt *idx_R_local=NULL; 14633a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 14643a50541eSStefano Zampini PetscInt vbs,bs; 14656816873aSStefano Zampini PetscBT bitmask=NULL; 1466a64d13efSStefano Zampini PetscErrorCode ierr; 1467a64d13efSStefano Zampini 1468a64d13efSStefano Zampini PetscFunctionBegin; 1469b23d619eSStefano Zampini /* 1470b23d619eSStefano Zampini No need to setup local scatters if 1471b23d619eSStefano Zampini - primal space is unchanged 1472b23d619eSStefano Zampini AND 1473b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1474b23d619eSStefano Zampini AND 1475b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1476b23d619eSStefano Zampini */ 1477b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1478f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1479f4ddd8eeSStefano Zampini } 1480f4ddd8eeSStefano Zampini /* destroy old objects */ 1481f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1482f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1483f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1484a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1485b371cd4fSStefano Zampini n_B = pcis->n_B; 1486b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1487b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14883a50541eSStefano Zampini 1489a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 14906816873aSStefano Zampini 149153892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 14926816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1493854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1494a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1495a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 14960e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1497a64d13efSStefano Zampini } 1498a64d13efSStefano Zampini 1499a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 15004641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 15016816873aSStefano Zampini idx_R_local[n_R++] = i; 1502a64d13efSStefano Zampini } 1503a64d13efSStefano Zampini } 150453892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 15056816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15066816873aSStefano Zampini 150753892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 150853892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 15096816873aSStefano Zampini } 15103a50541eSStefano Zampini 15113a50541eSStefano Zampini /* Block code */ 15123a50541eSStefano Zampini vbs = 1; 15133a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 15143a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 15153a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 15163a50541eSStefano Zampini PetscInt *vary; 1517d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1518785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 15193a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1520d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1521d3df7717SStefano 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 */ 15220e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1523d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 15243a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 15253a50541eSStefano Zampini is_blocked = PETSC_FALSE; 15263a50541eSStefano Zampini break; 15273a50541eSStefano Zampini } 15283a50541eSStefano Zampini } 1529d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1530d3df7717SStefano Zampini } else { 1531d3df7717SStefano Zampini /* Verify directly the R set */ 1532d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1533d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1534d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1535d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1536d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1537d3df7717SStefano Zampini break; 1538d3df7717SStefano Zampini } 1539d3df7717SStefano Zampini } 1540d3df7717SStefano Zampini } 1541d3df7717SStefano Zampini } 15423a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 15433a50541eSStefano Zampini vbs = bs; 15443a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 15453a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 15463a50541eSStefano Zampini } 15473a50541eSStefano Zampini } 15483a50541eSStefano Zampini } 15493a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 155053892102SStefano Zampini if (sub_schurs->reuse_mumps) { 155153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 155253892102SStefano Zampini 155353892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 155453892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 155553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 155653892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 155753892102SStefano Zampini } else { 15583a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 155953892102SStefano Zampini } 1560a64d13efSStefano Zampini 1561a64d13efSStefano Zampini /* print some info if requested */ 1562a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1563*019a44ceSStefano Zampini PetscInt benign; 1564*019a44ceSStefano Zampini 1565*019a44ceSStefano Zampini benign = (PetscInt)(!!pcbddc->zerodiag); 1566a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1567a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 15680fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1569a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1570a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1571*019a44ceSStefano 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-benign,pcbddc->local_primal_size);CHKERRQ(ierr); 1572a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1573a64d13efSStefano Zampini } 1574a64d13efSStefano Zampini 1575a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 15766816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 15776816873aSStefano Zampini IS is_aux1,is_aux2; 15786816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 15796816873aSStefano Zampini 15803a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1581854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1582854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1583a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 15844641a718SStefano Zampini for (i=0; i<n_D; i++) { 15854641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 15864641a718SStefano Zampini } 1587a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1588a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 15894641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 15904641a718SStefano Zampini aux_array1[j++] = i; 1591a64d13efSStefano Zampini } 1592a64d13efSStefano Zampini } 1593a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1594a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1595a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 15964641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 15974641a718SStefano Zampini aux_array2[j++] = i; 1598a64d13efSStefano Zampini } 1599a64d13efSStefano Zampini } 1600a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1601a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1602a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1603a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1604a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1605a64d13efSStefano Zampini 16068eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1607785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1608a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16094641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 16104641a718SStefano Zampini aux_array1[j++] = i; 1611a64d13efSStefano Zampini } 1612a64d13efSStefano Zampini } 1613a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1614a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1615a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1616a64d13efSStefano Zampini } 16174641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 16183a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1619d62866d3SStefano Zampini } else { 162053892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16216816873aSStefano Zampini IS tis; 16226816873aSStefano Zampini PetscInt schur_size; 16236816873aSStefano Zampini 162453892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 16256816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 162653892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 16276816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16286816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 16296816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 16306816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 16316816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1632d62866d3SStefano Zampini } 1633d62866d3SStefano Zampini } 1634a64d13efSStefano Zampini PetscFunctionReturn(0); 1635a64d13efSStefano Zampini } 1636a64d13efSStefano Zampini 1637304d26faSStefano Zampini 1638304d26faSStefano Zampini #undef __FUNCT__ 1639304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1640684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1641304d26faSStefano Zampini { 1642304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1643304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1644304d26faSStefano Zampini PC pc_temp; 1645304d26faSStefano Zampini Mat A_RR; 1646f4ddd8eeSStefano Zampini MatReuse reuse; 1647304d26faSStefano Zampini PetscScalar m_one = -1.0; 1648304d26faSStefano Zampini PetscReal value; 164904708bb6SStefano Zampini PetscInt n_D,n_R; 16509577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1651304d26faSStefano Zampini PetscErrorCode ierr; 1652e604994aSStefano Zampini /* prefixes stuff */ 1653312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1654e604994aSStefano Zampini size_t len; 1655304d26faSStefano Zampini 1656304d26faSStefano Zampini PetscFunctionBegin; 1657304d26faSStefano Zampini 1658e604994aSStefano Zampini /* compute prefixes */ 1659e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1660e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1661e604994aSStefano Zampini if (!pcbddc->current_level) { 1662e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1663e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1664e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1665e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1666e604994aSStefano Zampini } else { 1667e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1668312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1669e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1670e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1671312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1672312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 167334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 167434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1675e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1676e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1677e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1678e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1679e604994aSStefano Zampini } 1680e604994aSStefano Zampini 1681304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1682684f6988SStefano Zampini if (dirichlet) { 1683d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 16843301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16853301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1686964fefecSStefano Zampini } 1687ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1688964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1689304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1690304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1691304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1692304d26faSStefano Zampini /* default */ 1693304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1694e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 16959577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1696304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 16979577ea80SStefano Zampini if (issbaij) { 16989577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16999577ea80SStefano Zampini } else { 1700304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 17019577ea80SStefano Zampini } 1702304d26faSStefano Zampini /* Allow user's customization */ 1703304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1704304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1705304d26faSStefano Zampini } 1706d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1707d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1708d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1709d62866d3SStefano Zampini 1710d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1711d5574798SStefano Zampini } 1712304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1713304d26faSStefano Zampini if (!n_D) { 1714304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1715304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1716304d26faSStefano Zampini } 1717304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1718304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1719304d26faSStefano Zampini /* set ksp_D into pcis data */ 1720304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1721304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1722304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1723684f6988SStefano Zampini } 1724304d26faSStefano Zampini 1725304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1726684f6988SStefano Zampini A_RR = 0; 1727684f6988SStefano Zampini if (neumann) { 1728d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 172904708bb6SStefano Zampini PetscInt ibs,mbs; 173004708bb6SStefano Zampini PetscBool issbaij; 173104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1732f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 17338ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1734f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1735f4ddd8eeSStefano Zampini PetscInt nn_R; 173681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1737f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1738f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1739f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1740f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1741f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1742f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1743f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1744727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1745f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1746f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1747f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1748f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1749f4ddd8eeSStefano Zampini } 1750f4ddd8eeSStefano Zampini } 1751f4ddd8eeSStefano Zampini /* last check */ 1752d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1753f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1754f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1755f4ddd8eeSStefano Zampini } 1756f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1757f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1758f4ddd8eeSStefano Zampini } 1759f4ddd8eeSStefano Zampini /* extract A_RR */ 1760af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1761af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 176204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 176304708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 176404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 176504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 176604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1767af732b37SStefano Zampini } else { 176804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 17696816873aSStefano Zampini } 177004708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 177104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 177204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 177304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 177404708bb6SStefano Zampini } else { 177504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 177604708bb6SStefano Zampini } 177704708bb6SStefano Zampini } 177804708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1779f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 17803301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 17813301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 17826816873aSStefano Zampini } 17836816873aSStefano Zampini } else { 17846816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 17856816873aSStefano Zampini 17866816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 17876816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 17886816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1789af732b37SStefano Zampini } 1790f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1791304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1792304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1793304d26faSStefano Zampini /* default */ 1794304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1795e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1796304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 17979577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 17989577ea80SStefano Zampini if (issbaij) { 17999577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 18009577ea80SStefano Zampini } else { 1801304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 18029577ea80SStefano Zampini } 1803304d26faSStefano Zampini /* Allow user's customization */ 1804304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1805304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1806304d26faSStefano Zampini } 1807d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1808304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1809304d26faSStefano Zampini if (!n_R) { 1810304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1811304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1812304d26faSStefano Zampini } 1813d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1814d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1815d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1816d62866d3SStefano Zampini 1817d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1818d62866d3SStefano Zampini } 1819304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1820304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1821684f6988SStefano Zampini } 18226816873aSStefano Zampini /* free Neumann problem's matrix */ 18236816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1824304d26faSStefano Zampini 1825304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 18260fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1827684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1828684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1829684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1830684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1831684f6988SStefano Zampini } 1832684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 18330fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 18340fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 18350fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 18360fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 18370fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1838304d26faSStefano Zampini /* need to be adapted? */ 1839b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1840b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1841b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1842304d26faSStefano Zampini /* print info */ 1843304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1844e604994aSStefano 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); 1845304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1846304d26faSStefano Zampini } 1847b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1848298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1849304d26faSStefano Zampini } 1850684f6988SStefano Zampini } 1851684f6988SStefano Zampini if (neumann) { /* Neumann */ 18526816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 18530fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 18540fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 18550fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 18560fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 18570fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1858304d26faSStefano Zampini /* need to be adapted? */ 1859b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1860b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1861304d26faSStefano Zampini /* print info */ 1862304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1863e604994aSStefano 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); 1864304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1865304d26faSStefano Zampini } 1866b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1867298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1868304d26faSStefano Zampini } 18690fccc4e9SStefano Zampini } 1870684f6988SStefano Zampini } 1871304d26faSStefano Zampini PetscFunctionReturn(0); 1872304d26faSStefano Zampini } 1873304d26faSStefano Zampini 1874304d26faSStefano Zampini #undef __FUNCT__ 1875ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 187680677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1877674ae819SStefano Zampini { 1878674ae819SStefano Zampini PetscErrorCode ierr; 1879674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1880be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1881674ae819SStefano Zampini 1882674ae819SStefano Zampini PetscFunctionBegin; 1883be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 188480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 188520c7b377SStefano Zampini } 188680677318SStefano Zampini if (!pcbddc->switch_static) { 188780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 188880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 188980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 189020c7b377SStefano Zampini } 1891be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 189280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 189380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 189420c7b377SStefano Zampini } else { 1895be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1896be83ff47SStefano Zampini 189753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 189853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 189920c7b377SStefano Zampini } 1900be83ff47SStefano Zampini } else { 190180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 190280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 190380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 190480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 190580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 190680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 190780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 190880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 190980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1910674ae819SStefano Zampini } 1911674ae819SStefano Zampini } 1912be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 191380677318SStefano Zampini if (applytranspose) { 191480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 191580677318SStefano Zampini } else { 191680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 191780677318SStefano Zampini } 191853892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1919be83ff47SStefano Zampini } else { 1920be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1921be83ff47SStefano Zampini 1922be83ff47SStefano Zampini if (applytranspose) { 192353892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1924be83ff47SStefano Zampini } else { 192553892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1926be83ff47SStefano Zampini } 192753892102SStefano Zampini #endif 1928be83ff47SStefano Zampini } 192980677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 193080677318SStefano Zampini if (!pcbddc->switch_static) { 1931be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 193280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1934be83ff47SStefano Zampini } else { 1935be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1936be83ff47SStefano Zampini 193753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 193853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1939be83ff47SStefano Zampini } 194080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 194180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 194280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 194380677318SStefano Zampini } 194480677318SStefano Zampini } else { 194580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 195080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 195180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 195280677318SStefano Zampini } 195380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1957674ae819SStefano Zampini } 1958674ae819SStefano Zampini PetscFunctionReturn(0); 1959674ae819SStefano Zampini } 1960674ae819SStefano Zampini 1961dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1962674ae819SStefano Zampini #undef __FUNCT__ 1963674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1964dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1965674ae819SStefano Zampini { 1966674ae819SStefano Zampini PetscErrorCode ierr; 1967674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1968674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1969674ae819SStefano Zampini const PetscScalar zero = 0.0; 1970674ae819SStefano Zampini 1971674ae819SStefano Zampini PetscFunctionBegin; 1972dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1973dc359a40SStefano Zampini if (applytranspose) { 1974674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 19758eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1976dc359a40SStefano Zampini } else { 1977674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1978674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 197915aaf578SStefano Zampini } 198012edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 198112edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 198212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198412edc857SStefano Zampini 19859f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 198612edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 198712edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1988964fefecSStefano Zampini Vec rhs,sol; 1989964fefecSStefano Zampini 1990964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1991964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 199212edc857SStefano Zampini if (applytranspose) { 1993964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 199412edc857SStefano Zampini } else { 1995964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 199612edc857SStefano Zampini } 199712edc857SStefano Zampini } 1998674ae819SStefano Zampini 1999674ae819SStefano Zampini /* Local solution on R nodes */ 200080677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 200180677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 20029f00e9b4SStefano Zampini } 2003674ae819SStefano Zampini 20049f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 20059f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 200612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2007674ae819SStefano Zampini 2008674ae819SStefano Zampini /* Sum contributions from two levels */ 2009dc359a40SStefano Zampini if (applytranspose) { 2010dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2011dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2012dc359a40SStefano Zampini } else { 2013674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 20148eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2015dc359a40SStefano Zampini } 2016674ae819SStefano Zampini PetscFunctionReturn(0); 2017674ae819SStefano Zampini } 2018674ae819SStefano Zampini 2019674ae819SStefano Zampini #undef __FUNCT__ 2020674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 202112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2022674ae819SStefano Zampini { 2023674ae819SStefano Zampini PetscErrorCode ierr; 2024674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 202558da7f69SStefano Zampini PetscScalar *array; 202612edc857SStefano Zampini Vec from,to; 2027674ae819SStefano Zampini 2028674ae819SStefano Zampini PetscFunctionBegin; 202912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 203012edc857SStefano Zampini from = pcbddc->coarse_vec; 203112edc857SStefano Zampini to = pcbddc->vec1_P; 203212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 203312edc857SStefano Zampini Vec tvec; 203458da7f69SStefano Zampini 203558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 203658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 203712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 203858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 203958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 204058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 204112edc857SStefano Zampini } 204212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 204312edc857SStefano Zampini from = pcbddc->vec1_P; 204412edc857SStefano Zampini to = pcbddc->coarse_vec; 204512edc857SStefano Zampini } 204612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2047674ae819SStefano Zampini PetscFunctionReturn(0); 2048674ae819SStefano Zampini } 2049674ae819SStefano Zampini 2050674ae819SStefano Zampini #undef __FUNCT__ 2051674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 205212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2053674ae819SStefano Zampini { 2054674ae819SStefano Zampini PetscErrorCode ierr; 2055674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 205658da7f69SStefano Zampini PetscScalar *array; 205712edc857SStefano Zampini Vec from,to; 2058674ae819SStefano Zampini 2059674ae819SStefano Zampini PetscFunctionBegin; 206012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 206112edc857SStefano Zampini from = pcbddc->coarse_vec; 206212edc857SStefano Zampini to = pcbddc->vec1_P; 206312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 206412edc857SStefano Zampini from = pcbddc->vec1_P; 206512edc857SStefano Zampini to = pcbddc->coarse_vec; 206612edc857SStefano Zampini } 206712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 206812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 206912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 207012edc857SStefano Zampini Vec tvec; 207158da7f69SStefano Zampini 207212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 207358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 207458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 207558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 207658da7f69SStefano Zampini } 207758da7f69SStefano Zampini } else { 207858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 207958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 208012edc857SStefano Zampini } 208112edc857SStefano Zampini } 2082674ae819SStefano Zampini PetscFunctionReturn(0); 2083674ae819SStefano Zampini } 2084674ae819SStefano Zampini 2085984c4197SStefano Zampini /* uncomment for testing purposes */ 2086984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2087674ae819SStefano Zampini #undef __FUNCT__ 2088674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2089674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2090674ae819SStefano Zampini { 2091674ae819SStefano Zampini PetscErrorCode ierr; 2092674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2093674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2094674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2095984c4197SStefano Zampini /* one and zero */ 2096984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2097984c4197SStefano Zampini /* space to store constraints and their local indices */ 20989162d606SStefano Zampini PetscScalar *constraints_data; 20999162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 21009162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 21019162d606SStefano Zampini PetscInt *constraints_n; 2102984c4197SStefano Zampini /* iterators */ 2103b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2104984c4197SStefano Zampini /* BLAS integers */ 2105e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2106e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2107c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2108727cdba6SStefano Zampini /* reuse */ 21090e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 21100e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2111984c4197SStefano Zampini /* change of basis */ 2112b3d85658SStefano Zampini PetscBool qr_needed; 21139162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2114984c4197SStefano Zampini /* auxiliary stuff */ 211564efe560SStefano Zampini PetscInt *nnz,*is_indices; 21168a0068c3SStefano Zampini PetscInt ncc; 2117984c4197SStefano Zampini /* some quantities */ 211845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2119a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2120984c4197SStefano Zampini 2121674ae819SStefano Zampini PetscFunctionBegin; 21228e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 21238e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 21248e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2125088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2126088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 21270e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 21280e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 21290e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21300e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21310e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2132088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2133cf5a6209SStefano Zampini 2134cf5a6209SStefano Zampini /* print some info */ 2135cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2136cf5a6209SStefano Zampini IS vertices; 2137cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2138cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2139cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2140cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2141cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2142cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2143cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2144fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2145fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2146cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2147cf5a6209SStefano Zampini } 2148cf5a6209SStefano Zampini 2149cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 21509162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2151cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2152cf5a6209SStefano Zampini const Vec *nearnullvecs; 2153cf5a6209SStefano Zampini Vec *localnearnullsp; 2154cf5a6209SStefano Zampini PetscScalar *array; 2155cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2156cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2157674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2158b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2159674ae819SStefano Zampini PetscScalar *work; 2160674ae819SStefano Zampini PetscReal *singular_vals; 2161674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2162674ae819SStefano Zampini PetscReal *rwork; 2163674ae819SStefano Zampini #endif 2164674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2165674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2166674ae819SStefano Zampini #else 2167964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2168964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2169674ae819SStefano Zampini #endif 2170674ae819SStefano Zampini 2171674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2172d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2173d06fc5fdSStefano Zampini /* free unneeded index sets */ 2174d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2175d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2176674ae819SStefano Zampini } 2177d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2178d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2179d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2180d06fc5fdSStefano Zampini } 2181d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2182d06fc5fdSStefano Zampini n_ISForEdges = 0; 2183d06fc5fdSStefano Zampini } 2184d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2185d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2186d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2187d06fc5fdSStefano Zampini } 2188d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2189d06fc5fdSStefano Zampini n_ISForFaces = 0; 2190d06fc5fdSStefano Zampini } 219170022509SStefano Zampini 219270022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 219370022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 219470022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 219570022509SStefano Zampini if (pcbddc->NullSpace) { 219670022509SStefano Zampini PetscBool tbool[2],gbool[2]; 219770022509SStefano Zampini 219870022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2199b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2200d06fc5fdSStefano Zampini if (!ISForEdges) { 2201d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2202d06fc5fdSStefano Zampini } 2203b8ffe317SStefano Zampini } 2204d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2205d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2206d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2207d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2208d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 220998a51de6SStefano Zampini } 221070022509SStefano Zampini #endif 221108122e43SStefano Zampini 2212674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2213674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2214674ae819SStefano Zampini if (nearnullsp) { 2215674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2216f4ddd8eeSStefano Zampini /* remove any stored info */ 2217f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2218f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2219f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2220f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2221f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2222473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2223f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2224f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2225f4ddd8eeSStefano Zampini } 2226984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2227984c4197SStefano Zampini nnsp_size = 0; 2228674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2229674ae819SStefano Zampini } 2230984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2231984c4197SStefano Zampini max_constraints = nnsp_size; 2232984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2233984c4197SStefano Zampini 2234674ae819SStefano Zampini /* 2235674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 22369162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 22379162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 22389162d606SStefano Zampini There can be multiple constraints per connected component 2239674ae819SStefano Zampini */ 2240674ae819SStefano Zampini n_vertices = 0; 2241674ae819SStefano Zampini if (ISForVertices) { 2242674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2243674ae819SStefano Zampini } 22449162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 22459162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 22469162d606SStefano Zampini 22479162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 22489162d606SStefano Zampini total_counts *= max_constraints; 2249674ae819SStefano Zampini total_counts += n_vertices; 22504641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 22519162d606SStefano Zampini 2252674ae819SStefano Zampini total_counts = 0; 2253674ae819SStefano Zampini max_size_of_constraint = 0; 2254674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 22559162d606SStefano Zampini IS used_is; 2256674ae819SStefano Zampini if (i<n_ISForEdges) { 22579162d606SStefano Zampini used_is = ISForEdges[i]; 2258674ae819SStefano Zampini } else { 22599162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2260674ae819SStefano Zampini } 22619162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2262674ae819SStefano Zampini total_counts += j; 2263674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2264674ae819SStefano Zampini } 22659162d606SStefano 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); 22669162d606SStefano Zampini 2267984c4197SStefano Zampini /* get local part of global near null space vectors */ 2268785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2269984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2270984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2271e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2272e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2273984c4197SStefano Zampini } 2274674ae819SStefano Zampini 2275242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2276242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2277a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2278242a89d7SStefano Zampini 2279984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2280a773dcb8SStefano Zampini if (!skip_lapack) { 2281674ae819SStefano Zampini PetscScalar temp_work; 2282911cabfeSStefano Zampini 2283674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2284984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2285785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2286785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2287785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2288674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2289785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2290674ae819SStefano Zampini #endif 2291674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2292c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2293c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2294674ae819SStefano Zampini lwork = -1; 2295674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2296674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2297c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2298674ae819SStefano Zampini #else 2299c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2300674ae819SStefano Zampini #endif 2301674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2302984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2303674ae819SStefano Zampini #else /* on missing GESVD */ 2304674ae819SStefano Zampini /* SVD */ 2305674ae819SStefano Zampini PetscInt max_n,min_n; 2306674ae819SStefano Zampini max_n = max_size_of_constraint; 2307984c4197SStefano Zampini min_n = max_constraints; 2308984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2309674ae819SStefano Zampini min_n = max_size_of_constraint; 2310984c4197SStefano Zampini max_n = max_constraints; 2311674ae819SStefano Zampini } 2312785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2313674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2314785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2315674ae819SStefano Zampini #endif 2316674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2317674ae819SStefano Zampini lwork = -1; 2318e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2319e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2320b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2321674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2322674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23239162d606SStefano 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)); 2324674ae819SStefano Zampini #else 23259162d606SStefano 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)); 2326674ae819SStefano Zampini #endif 2327674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2328984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2329984c4197SStefano Zampini #endif /* on missing GESVD */ 2330674ae819SStefano Zampini /* Allocate optimal workspace */ 2331674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2332854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2333674ae819SStefano Zampini } 2334674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2335674ae819SStefano Zampini total_counts = 0; 23369162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 23379162d606SStefano Zampini constraints_data_ptr[0] = 0; 2338674ae819SStefano Zampini /* vertices */ 23399162d606SStefano Zampini if (n_vertices) { 2340674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 23419162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 23429162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2343674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 23449162d606SStefano Zampini constraints_n[total_counts] = 1; 23459162d606SStefano Zampini constraints_data[total_counts] = 1.0; 23469162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 23479162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2348674ae819SStefano Zampini total_counts++; 2349674ae819SStefano Zampini } 2350674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2351984c4197SStefano Zampini PetscBool used_vertex; 2352674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2353674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2354674ae819SStefano Zampini k = 0; 2355674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2356984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2357984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 23589162d606SStefano Zampini constraints_n[total_counts] = 1; 23599162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 23609162d606SStefano Zampini constraints_data[total_counts] = 1.0; 23619162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 23629162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2363674ae819SStefano Zampini total_counts++; 2364674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2365674ae819SStefano Zampini } 2366984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2367674ae819SStefano Zampini k++; 2368674ae819SStefano Zampini } 2369674ae819SStefano Zampini } 2370674ae819SStefano Zampini } 2371674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2372674ae819SStefano Zampini n_vertices = total_counts; 2373674ae819SStefano Zampini } 2374984c4197SStefano Zampini 2375674ae819SStefano Zampini /* edges and faces */ 23769162d606SStefano Zampini total_counts_cc = total_counts; 2377911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 23789162d606SStefano Zampini IS used_is; 23799162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 23809162d606SStefano Zampini 2381911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 23829162d606SStefano Zampini used_is = ISForEdges[ncc]; 2383984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2384674ae819SStefano Zampini } else { 23859162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2386984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2387674ae819SStefano Zampini } 2388674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 23899162d606SStefano Zampini 23909162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 23919162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2392984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2393984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2394674ae819SStefano Zampini if (nnsp_has_cnst) { 23955b08dc53SStefano Zampini PetscScalar quad_value; 23969162d606SStefano Zampini 23979162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 23989162d606SStefano Zampini idxs_copied = PETSC_TRUE; 23999162d606SStefano Zampini 2400a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2401674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2402a773dcb8SStefano Zampini } else { 2403a773dcb8SStefano Zampini quad_value = 1.0; 2404a773dcb8SStefano Zampini } 2405674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24069162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2407674ae819SStefano Zampini } 24089162d606SStefano Zampini temp_constraints++; 2409674ae819SStefano Zampini total_counts++; 2410674ae819SStefano Zampini } 2411674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2412984c4197SStefano Zampini PetscReal real_value; 24139162d606SStefano Zampini PetscScalar *ptr_to_data; 24149162d606SStefano Zampini 2415984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 24169162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2417674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24189162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2419674ae819SStefano Zampini } 2420984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2421984c4197SStefano Zampini /* check if array is null on the connected component */ 2422e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24239162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 24245b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2425674ae819SStefano Zampini temp_constraints++; 2426674ae819SStefano Zampini total_counts++; 24279162d606SStefano Zampini if (!idxs_copied) { 24289162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24299162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2430674ae819SStefano Zampini } 2431674ae819SStefano Zampini } 24329162d606SStefano Zampini } 24339162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 243445a1bb75SStefano Zampini valid_constraints = temp_constraints; 2435eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2436a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 24379162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 24389162d606SStefano Zampini 24399162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2440a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24419162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2442a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 24439162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2444a773dcb8SStefano Zampini } else { /* perform SVD */ 2445984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 24469162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2447674ae819SStefano Zampini 2448674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2449984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2450984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2451984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2452984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2453984c4197SStefano Zampini from that computed using LAPACKgesvd 2454984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2455984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2456984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2457674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2458e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2459984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2460674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2461674ae819SStefano Zampini for (k=0;k<j+1;k++) { 24629162d606SStefano 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)); 2463674ae819SStefano Zampini } 2464674ae819SStefano Zampini } 2465e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2466e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2467e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2468674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2469c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2470674ae819SStefano Zampini #else 2471c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2472674ae819SStefano Zampini #endif 2473674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2474984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2475984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2476674ae819SStefano Zampini j = 0; 2477984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2478674ae819SStefano Zampini total_counts = total_counts-j; 247945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2480e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2481c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2482c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2483c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2484c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2485c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2486c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2487674ae819SStefano Zampini if (j<temp_constraints) { 2488984c4197SStefano Zampini PetscInt ii; 2489984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2490674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 24919162d606SStefano 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)); 2492674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2493984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2494674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 24959162d606SStefano 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]; 2496674ae819SStefano Zampini } 2497674ae819SStefano Zampini } 2498674ae819SStefano Zampini } 2499674ae819SStefano Zampini #else /* on missing GESVD */ 2500e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2501e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2502b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2503674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2504674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25059162d606SStefano 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)); 2506674ae819SStefano Zampini #else 25079162d606SStefano 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)); 2508674ae819SStefano Zampini #endif 2509984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2510674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2511984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2512e310c8b4SStefano Zampini k = temp_constraints; 2513e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2514674ae819SStefano Zampini j = 0; 2515e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 251645a1bb75SStefano Zampini valid_constraints = k-j; 2517911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2518984c4197SStefano Zampini #endif /* on missing GESVD */ 2519674ae819SStefano Zampini } 2520a773dcb8SStefano Zampini } 25219162d606SStefano Zampini /* update pointers information */ 25229162d606SStefano Zampini if (valid_constraints) { 25239162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 25249162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 25259162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 25269162d606SStefano Zampini /* set change_of_basis flag */ 252745a1bb75SStefano Zampini if (boolforchange) { 2528b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 25299162d606SStefano Zampini } 2530b3d85658SStefano Zampini total_counts_cc++; 253145a1bb75SStefano Zampini } 253245a1bb75SStefano Zampini } 2533984c4197SStefano Zampini /* free workspace */ 25348f1c130eSStefano Zampini if (!skip_lapack) { 2535984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2536984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2537984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2538984c4197SStefano Zampini #endif 2539984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2540984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2541984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2542984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2543984c4197SStefano Zampini #endif 2544984c4197SStefano Zampini } 2545984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2546984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2547984c4197SStefano Zampini } 2548984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2549cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2550cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2551cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2552cf5a6209SStefano Zampini } 2553cf5a6209SStefano Zampini if (n_ISForFaces) { 2554cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2555cf5a6209SStefano Zampini } 2556cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2557cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2558cf5a6209SStefano Zampini } 2559cf5a6209SStefano Zampini if (n_ISForEdges) { 2560cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2561cf5a6209SStefano Zampini } 2562cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 256308122e43SStefano Zampini } else { 256408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2565984c4197SStefano Zampini 256608122e43SStefano Zampini total_counts = 0; 256708122e43SStefano Zampini n_vertices = 0; 2568d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2569d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 257008122e43SStefano Zampini } 257108122e43SStefano Zampini max_constraints = 0; 25729162d606SStefano Zampini total_counts_cc = 0; 257308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 257408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 25759162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 257608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 257708122e43SStefano Zampini } 25789162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 25799162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 25809162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 25819162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 258274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 25839162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 25849162d606SStefano Zampini total_counts_cc = 0; 25859162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 25869162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 25879162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 258808122e43SStefano Zampini } 258908122e43SStefano Zampini } 25909162d606SStefano Zampini #if 0 25919162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 25929162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 25939162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 25949162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 25959162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 25969162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 25979162d606SStefano Zampini } 25989162d606SStefano Zampini printf("\n"); 25999162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 26009162d606SStefano Zampini } 26011b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 26028bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 26031b968477SStefano Zampini } 26041b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 26058bec7fa6SStefano 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]); 26061b968477SStefano Zampini } 260708122e43SStefano Zampini #endif 260808122e43SStefano Zampini 26098bec7fa6SStefano Zampini max_size_of_constraint = 0; 26109162d606SStefano 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]); 26119162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 261208122e43SStefano Zampini /* Change of basis */ 2613b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 261408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 261508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 261608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2617b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 261808122e43SStefano Zampini } 261908122e43SStefano Zampini } 262008122e43SStefano Zampini } 262108122e43SStefano Zampini } 2622984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2623*019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2624*019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 262508122e43SStefano Zampini 26269162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 26279162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 26289162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 26299162d606SStefano 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); 263008122e43SStefano Zampini } 2631674ae819SStefano Zampini 2632674ae819SStefano Zampini /* Create constraint matrix */ 2633674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 263416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2635984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2636984c4197SStefano Zampini 2637984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2638a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2639a717540cSStefano Zampini qr_needed = PETSC_FALSE; 264074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2641984c4197SStefano Zampini total_primal_vertices=0; 2642b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 26439162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26449162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2645984c4197SStefano Zampini if (size_of_constraint == 1) { 26469162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2647b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 264864efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 26499162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 26509162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2651a717540cSStefano Zampini } 2652b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 265374d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2654a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2655a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2656a717540cSStefano Zampini } 2657fa434743SStefano Zampini } else { 2658b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2659fa434743SStefano Zampini } 2660a717540cSStefano Zampini } 2661b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2662b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2663674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 266470022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2665b3d85658SStefano Zampini 2666*019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2667*019a44ceSStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc+1,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+1,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 26680e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 26690e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2670984c4197SStefano Zampini 2671984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 267274d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2673785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2674984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 267574d5cdf7SStefano Zampini 2676984c4197SStefano Zampini j = total_primal_vertices; 267774d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2678b3d85658SStefano Zampini cum = total_primal_vertices; 26799162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 26804641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2681b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2682b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2683b3d85658SStefano Zampini cum++; 26849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 268574d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 268674d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 268774d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 268874d5cdf7SStefano Zampini } 26899162d606SStefano Zampini j += constraints_n[i]; 2690674ae819SStefano Zampini } 2691674ae819SStefano Zampini } 2692674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2693674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2694088faed8SStefano Zampini 2695674ae819SStefano Zampini /* set values in constraint matrix */ 2696984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 26970e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2698674ae819SStefano Zampini } 2699984c4197SStefano Zampini total_counts = total_primal_vertices; 27009162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27014641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 27029162d606SStefano Zampini PetscInt *cols; 27039162d606SStefano Zampini 27049162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 27059162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 27069162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27079162d606SStefano Zampini PetscInt row = total_counts+k; 27089162d606SStefano Zampini PetscScalar *vals; 27099162d606SStefano Zampini 27109162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 27119162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 27129162d606SStefano Zampini } 27139162d606SStefano Zampini total_counts += constraints_n[i]; 2714674ae819SStefano Zampini } 2715674ae819SStefano Zampini } 2716674ae819SStefano Zampini /* assembling */ 2717674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2718674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2719088faed8SStefano Zampini 2720984c4197SStefano Zampini /* 272145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2722984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2723984c4197SStefano Zampini */ 2724674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2725674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2726026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2727984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2728984c4197SStefano Zampini /* working stuff for GEQRF */ 272981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2730984c4197SStefano Zampini PetscBLASInt lqr_work; 2731984c4197SStefano Zampini /* working stuff for UNGQR */ 2732984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2733984c4197SStefano Zampini PetscBLASInt lgqr_work; 2734984c4197SStefano Zampini /* working stuff for TRTRS */ 2735984c4197SStefano Zampini PetscScalar *trs_rhs; 27363f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2737984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2738984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2739984c4197SStefano Zampini PetscScalar *start_vals; 2740984c4197SStefano Zampini /* working stuff for values insertion */ 27414641a718SStefano Zampini PetscBT is_primal; 274264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2743906d46d4SStefano Zampini /* matrix sizes */ 2744906d46d4SStefano Zampini PetscInt global_size,local_size; 2745906d46d4SStefano Zampini /* temporary change of basis */ 2746906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2747cf5a6209SStefano Zampini /* extra space for debugging */ 2748cf5a6209SStefano Zampini PetscScalar *dbg_work; 2749984c4197SStefano Zampini 2750906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2751906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 275216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2753bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2754906d46d4SStefano Zampini /* nonzeros for local mat */ 2755bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2756bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 27579162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2758a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 27599162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2760a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 27619162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2762a717540cSStefano Zampini } else { 27639162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 27649162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2765a717540cSStefano Zampini } 2766a717540cSStefano Zampini } 2767a717540cSStefano Zampini } 2768906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2769bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2770a717540cSStefano Zampini /* Set initial identity in the matrix */ 2771bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2772906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2773a717540cSStefano Zampini } 2774a717540cSStefano Zampini 2775a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2776a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2777a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2778a717540cSStefano Zampini } 2779a717540cSStefano Zampini 2780a717540cSStefano Zampini 2781a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2782a717540cSStefano Zampini /* 2783a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2784a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2785a717540cSStefano Zampini 2786a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2787a717540cSStefano Zampini 2788a6b551f4SStefano 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) 2789a6b551f4SStefano Zampini 2790a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2791a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2792a717540cSStefano Zampini | ... | 2793a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2794a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2795a717540cSStefano Zampini 2796a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2797a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2798a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2799a6b551f4SStefano Zampini 2800a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2801a717540cSStefano Zampini */ 2802a717540cSStefano Zampini if (qr_needed) { 2803984c4197SStefano Zampini /* space to store Q */ 2804854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2805984c4197SStefano Zampini /* first we issue queries for optimal work */ 28063f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28073f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 28083f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2809984c4197SStefano Zampini lqr_work = -1; 28103f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2811984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2812984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2813785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2814984c4197SStefano Zampini lgqr_work = -1; 28153f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28163f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 28173f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 28183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 28193f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 28203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2821984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2822984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2823785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2824984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2825785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2826984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2827785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2828a717540cSStefano Zampini /* allocating workspace for check */ 2829a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2830cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2831a717540cSStefano Zampini } 2832a717540cSStefano Zampini } 2833984c4197SStefano Zampini /* array to store whether a node is primal or not */ 28344641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2835473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 28360e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 283739e2fb2aSStefano Zampini if (i != total_primal_vertices) { 283839e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 28394641a718SStefano Zampini } 284039e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 284139e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 284239e2fb2aSStefano Zampini } 284339e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2844984c4197SStefano Zampini 2845a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 28469162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 28479162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 28484641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2849984c4197SStefano Zampini /* get constraint info */ 28509162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2851984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2852984c4197SStefano Zampini 2853984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28549162d606SStefano 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); 2855674ae819SStefano Zampini } 2856984c4197SStefano Zampini 2857fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2858a717540cSStefano Zampini 2859a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2860a717540cSStefano Zampini if (pcbddc->dbg_flag) { 28619162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2862a717540cSStefano Zampini } 2863984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 28649162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2865984c4197SStefano Zampini 2866984c4197SStefano Zampini /* compute QR decomposition of constraints */ 28673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 28683f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 28693f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2870674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 28713f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2872984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2873674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2874984c4197SStefano Zampini 2875984c4197SStefano Zampini /* explictly compute R^-T */ 2876984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2877984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 28783f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 28793f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 28803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 28813f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2882984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 28833f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2884984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2885984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2886984c4197SStefano Zampini 2887a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 28883f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 28893f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 28913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2892984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 28933f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2894984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2895984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2896984c4197SStefano Zampini 2897984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2898984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2899984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 29003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29013f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29043f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 29053f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2906984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29079162d606SStefano 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)); 2908984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 29099162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2910984c4197SStefano Zampini 2911984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 29129162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2913984c4197SStefano Zampini /* insert cols for primal dofs */ 2914984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2915984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 29169162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2917906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2918984c4197SStefano Zampini } 2919984c4197SStefano Zampini /* insert cols for dual dofs */ 2920984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 29219162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2922984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 29239162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2924906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2925984c4197SStefano Zampini j++; 2926674ae819SStefano Zampini } 2927674ae819SStefano Zampini } 2928984c4197SStefano Zampini 2929984c4197SStefano Zampini /* check change of basis */ 2930984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2931984c4197SStefano Zampini PetscInt ii,jj; 2932984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2933c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2934c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2935c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2936c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2937c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2938c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2939984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2940cf5a6209SStefano 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)); 2941984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2942984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2943984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2944cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2945cf5a6209SStefano 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; 2946674ae819SStefano Zampini } 2947674ae819SStefano Zampini } 2948984c4197SStefano Zampini if (!valid_qr) { 294922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2950984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2951984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2952cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2953cf5a6209SStefano 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])); 2954674ae819SStefano Zampini } 2955cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2956cf5a6209SStefano 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])); 2957984c4197SStefano Zampini } 2958984c4197SStefano Zampini } 2959984c4197SStefano Zampini } 2960674ae819SStefano Zampini } else { 296122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2962674ae819SStefano Zampini } 2963674ae819SStefano Zampini } 2964a717540cSStefano Zampini } else { /* simple transformation block */ 2965a717540cSStefano Zampini PetscInt row,col; 2966a6b551f4SStefano Zampini PetscScalar val,norm; 2967a6b551f4SStefano Zampini 2968a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29699162d606SStefano 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)); 2970a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 29719162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 29729162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2973bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 29749162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2975906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 29769162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2977a717540cSStefano Zampini } else { 2978a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 29799162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2980a717540cSStefano Zampini if (row != col) { 29819162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2982a717540cSStefano Zampini } else { 29839162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2984a717540cSStefano Zampini } 2985906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2986a717540cSStefano Zampini } 2987a717540cSStefano Zampini } 2988a717540cSStefano Zampini } 298998a51de6SStefano Zampini if (pcbddc->dbg_flag) { 299022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2991a717540cSStefano Zampini } 2992674ae819SStefano Zampini } 2993984c4197SStefano Zampini } else { 2994984c4197SStefano Zampini if (pcbddc->dbg_flag) { 29959162d606SStefano 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); 2996674ae819SStefano Zampini } 2997674ae819SStefano Zampini } 2998674ae819SStefano Zampini } 2999a717540cSStefano Zampini 3000a717540cSStefano Zampini /* free workspace */ 3001a717540cSStefano Zampini if (qr_needed) { 3002984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3003cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3004984c4197SStefano Zampini } 3005984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3006984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3007984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3008984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3009984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3010674ae819SStefano Zampini } 3011a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3012906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3013906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3014906d46d4SStefano Zampini 3015906d46d4SStefano Zampini /* assembling of global change of variable */ 3016bbb9e6c6SStefano Zampini { 3017bbb9e6c6SStefano Zampini Mat tmat; 301816f15bc4SStefano Zampini PetscInt bs; 301916f15bc4SStefano Zampini 3020906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3021906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3022bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3023bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3024bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3025bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 302616f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 302716f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3028906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3029bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3030bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3031bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3032bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3033bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3034e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3035e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3036bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3037bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3038906d46d4SStefano Zampini } 3039906d46d4SStefano Zampini /* check */ 3040906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3041906d46d4SStefano Zampini PetscReal error; 3042906d46d4SStefano Zampini Vec x,x_change; 3043906d46d4SStefano Zampini 3044906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3045906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3046906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3047906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3048e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3049e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3050bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3051e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3052e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3053906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3054906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3055906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3056906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3057bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3058906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3059906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3060906d46d4SStefano Zampini } 3061b96c3477SStefano Zampini 3062b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3063b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3064b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3065b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3066ac632422SStefano Zampini Mat S_new,tmat; 3067bbb9e6c6SStefano Zampini IS is_all_N; 3068bbb9e6c6SStefano Zampini 3069bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 30706816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3071bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3072ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3073b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3074ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3075ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3076ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3077ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3078ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3079b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3080ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3081ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3082ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3083ac632422SStefano Zampini } 3084b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3085b96c3477SStefano Zampini } 3086b96c3477SStefano Zampini } 3087906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3088906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3089b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3090b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3091b9b85e73SStefano Zampini } 3092906d46d4SStefano Zampini 3093906d46d4SStefano Zampini /* set up change of basis context */ 3094906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3095906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3096906d46d4SStefano Zampini 3097906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3098906d46d4SStefano Zampini PetscInt global_size,local_size; 3099906d46d4SStefano Zampini 3100906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3101906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3102906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3103906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3104906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3105906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3106906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3107906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3108906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3109906d46d4SStefano Zampini } else { 3110906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3111906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3112906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3113906d46d4SStefano Zampini } 3114906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3115906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3116906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3117906d46d4SStefano Zampini } else { 3118906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3119906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3120906d46d4SStefano Zampini } 3121906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3122906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3123906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3124906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3125b9b85e73SStefano Zampini } 3126a717540cSStefano Zampini 3127*019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3128*019a44ceSStefano Zampini if (pcbddc->zerodiag) { 3129*019a44ceSStefano Zampini const PetscInt *idxs; 3130*019a44ceSStefano Zampini PetscInt nz; 3131*019a44ceSStefano Zampini 3132*019a44ceSStefano Zampini ierr = ISGetLocalSize(pcbddc->zerodiag,&nz);CHKERRQ(ierr); 3133*019a44ceSStefano Zampini ierr = ISGetIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3134*019a44ceSStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3135*019a44ceSStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3136*019a44ceSStefano Zampini ierr = ISRestoreIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3137*019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3138*019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3139*019a44ceSStefano Zampini pcbddc->local_primal_size++; 3140*019a44ceSStefano Zampini } 3141*019a44ceSStefano Zampini 3142*019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3143727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3144727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3145aff50787SStefano 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); 3146c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 31470e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3148aff50787SStefano 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); 3149727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3150727cdba6SStefano Zampini } 31510e6343abSStefano Zampini } 31520e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3153727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3154727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3155727cdba6SStefano Zampini 3156a717540cSStefano Zampini /* flush dbg viewer */ 3157b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3158b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3159b8ffe317SStefano Zampini } 3160a717540cSStefano Zampini 3161e310c8b4SStefano Zampini /* free workspace */ 3162a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 31634641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 316408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 31659162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 31669162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 316708122e43SStefano Zampini } else { 31689162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 31699162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 31709162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 317108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 317208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 31739162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 31749162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 317508122e43SStefano Zampini } 3176674ae819SStefano Zampini PetscFunctionReturn(0); 3177674ae819SStefano Zampini } 3178674ae819SStefano Zampini 3179674ae819SStefano Zampini #undef __FUNCT__ 3180674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3181674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3182674ae819SStefano Zampini { 3183674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3184674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3185674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 31867fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3187674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3188674ae819SStefano Zampini 3189674ae819SStefano Zampini PetscFunctionBegin; 31908e61c736SStefano Zampini /* Reset previously computed graph */ 31918e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3192674ae819SStefano Zampini /* Init local Graph struct */ 31937fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 31943bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3195674ae819SStefano Zampini 3196575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3197575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3198575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3199575ad6abSStefano Zampini } 32009577ea80SStefano Zampini 3201674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 32024d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 32034d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 32044d379d7bSStefano Zampini PetscInt nvtxs; 3205e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3206674ae819SStefano Zampini 32074d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 32082fffb893SStefano Zampini 32092fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32102fffb893SStefano Zampini if (flg_row) { 32114d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3212b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 32132fffb893SStefano Zampini } 32142fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32159b28b941SStefano 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 */ 32164d379d7bSStefano Zampini IS is_dummy; 32174d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 32184d379d7bSStefano Zampini PetscInt j,sum; 32194d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 32204d379d7bSStefano Zampini const PetscInt *idxs; 32214d379d7bSStefano Zampini PCBDDCGraph graph; 32224d379d7bSStefano Zampini PetscBT is_on_boundary; 32234d379d7bSStefano Zampini 32244d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 32254d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 32264d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 32274d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 32287fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 32294d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3230e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3231e496cd5dSStefano Zampini if (flg_row) { 32324d379d7bSStefano Zampini graph->xadj = xadj; 32334d379d7bSStefano Zampini graph->adjncy = adjncy; 3234e496cd5dSStefano Zampini } 32354d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 32364d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3237e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32384d379d7bSStefano Zampini 32394d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 32409b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 32414d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32424d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 32434d379d7bSStefano Zampini } 32444d379d7bSStefano Zampini } 32454d379d7bSStefano Zampini 3246e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 32474d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 32484d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 32494d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 32504d379d7bSStefano Zampini } 32514d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 32524d379d7bSStefano Zampini 3253e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 32544d379d7bSStefano Zampini sum = 0; 32554d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32564d379d7bSStefano Zampini PetscInt sizecc = 0; 32574d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 32584d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 32594d379d7bSStefano Zampini sizecc++; 32604d379d7bSStefano Zampini } 32614d379d7bSStefano Zampini } 32624d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 32634d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 32644d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 32654d379d7bSStefano Zampini } 32664d379d7bSStefano Zampini } 32674d379d7bSStefano Zampini sum += sizecc*sizecc; 32684d379d7bSStefano Zampini } 32694d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 32704d379d7bSStefano Zampini sum = 0; 3271e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 32724d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 32734d379d7bSStefano Zampini cxadj[i] = sum; 32744d379d7bSStefano Zampini sum += temp; 32754d379d7bSStefano Zampini } 3276e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 32774d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32784d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 32794d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 32804d379d7bSStefano Zampini PetscInt k,sizecc = 0; 32814d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 32824d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 32834d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 32844d379d7bSStefano Zampini sizecc++; 32854d379d7bSStefano Zampini } 32864d379d7bSStefano Zampini } 32874d379d7bSStefano Zampini } 32884d379d7bSStefano Zampini } 32894d379d7bSStefano Zampini } 32909b28b941SStefano Zampini if (sum) { 3291e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 32924d379d7bSStefano Zampini } else { 32934d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 32944d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 32954d379d7bSStefano Zampini } 32964d379d7bSStefano Zampini graph->xadj = 0; 32974d379d7bSStefano Zampini graph->adjncy = 0; 32984d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 32994d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 33004d379d7bSStefano Zampini } 3301674ae819SStefano Zampini } 33029b28b941SStefano Zampini if (pcbddc->dbg_flag) { 33039b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3304674ae819SStefano Zampini } 3305674ae819SStefano Zampini 330663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3307674ae819SStefano Zampini vertex_size = 1; 330863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 330963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 331095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 331163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3312e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 331363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3314674ae819SStefano Zampini } 331563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 331663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 331763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3318674ae819SStefano Zampini } 331963602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3320674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 332163602bcaSStefano Zampini } else { 332263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 332363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3324854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 332563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 332663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 332763602bcaSStefano Zampini } 332863602bcaSStefano Zampini } 3329674ae819SStefano Zampini } 3330674ae819SStefano Zampini 3331674ae819SStefano Zampini /* Setup of Graph */ 3332785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3333e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3334785d1243SStefano Zampini } 3335785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3336e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3337785d1243SStefano Zampini } 3338302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3339674ae819SStefano Zampini 3340674ae819SStefano Zampini /* Graph's connected components analysis */ 3341674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3342674ae819SStefano Zampini 3343674ae819SStefano Zampini /* print some info to stdout */ 3344674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3345302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3346674ae819SStefano Zampini } 3347fb180af4SStefano Zampini 3348fb180af4SStefano Zampini /* mark topography has done */ 3349fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3350674ae819SStefano Zampini PetscFunctionReturn(0); 3351674ae819SStefano Zampini } 3352674ae819SStefano Zampini 3353dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3354674ae819SStefano Zampini #undef __FUNCT__ 3355674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3356dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3357674ae819SStefano Zampini { 3358dc456d91SStefano Zampini PetscSF sf; 3359dc456d91SStefano Zampini PetscLayout map; 3360dc456d91SStefano Zampini const PetscInt *idxs; 3361dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3362dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3363dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3364dc456d91SStefano Zampini PetscMPIInt commsize; 3365674ae819SStefano Zampini PetscBool first_found; 3366674ae819SStefano Zampini PetscErrorCode ierr; 3367674ae819SStefano Zampini 3368674ae819SStefano Zampini PetscFunctionBegin; 3369dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3370dc456d91SStefano Zampini if (subset_mult) { 3371dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3372dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3373dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3374674ae819SStefano Zampini } 3375dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3376dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3377dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3378dc456d91SStefano Zampini for (i=0;i<n;i++) { 3379dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3380dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3381674ae819SStefano Zampini } 3382dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3383dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3384dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3385dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3386dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3387dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3388dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3389dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3390dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3391dc456d91SStefano Zampini 3392dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3393dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3394dc456d91SStefano Zampini if (subset_mult) { 3395dc456d91SStefano Zampini const PetscInt* idxs_mult; 3396dc456d91SStefano Zampini 3397dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3398dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3399dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3400674ae819SStefano Zampini } else { 3401dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3402674ae819SStefano Zampini } 3403dc456d91SStefano Zampini /* local size of new subset */ 3404dc456d91SStefano Zampini n_n = 0; 3405dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3406dc456d91SStefano Zampini 3407dc456d91SStefano Zampini /* global indexes in layout */ 3408dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3409dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3410dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3411dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3412dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3413dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3414dc456d91SStefano Zampini 3415dc456d91SStefano Zampini /* reduce from leaves to roots */ 3416dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 341764a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 341864a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3419dc456d91SStefano Zampini 3420dc456d91SStefano Zampini /* count indexes in local part of layout */ 3421674ae819SStefano Zampini nlocals = 0; 3422674ae819SStefano Zampini first_index = -1; 3423674ae819SStefano Zampini first_found = PETSC_FALSE; 3424dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3425dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3426674ae819SStefano Zampini first_found = PETSC_TRUE; 3427674ae819SStefano Zampini first_index = i; 3428674ae819SStefano Zampini } 3429dc456d91SStefano Zampini nlocals += root_data[i]; 3430674ae819SStefano Zampini } 3431dc456d91SStefano Zampini 3432dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 34335fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3434dc456d91SStefano Zampini start = 0; 343564a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 34365fa240b1SStefano Zampini #else 343764a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 34385fa240b1SStefano Zampini start = start-nlocals; 34395fa240b1SStefano Zampini #endif 34405fa240b1SStefano Zampini 3441dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3442dc456d91SStefano Zampini *N_n = start + nlocals; 3443dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3444dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3445674ae819SStefano Zampini } 34465fa240b1SStefano Zampini 34475fa240b1SStefano Zampini /* adapt root data with cumulative */ 3448674ae819SStefano Zampini if (first_found) { 3449dc456d91SStefano Zampini PetscInt old_index; 3450dc456d91SStefano Zampini 3451dc456d91SStefano Zampini root_data[first_index] += start; 3452674ae819SStefano Zampini old_index = first_index; 3453dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3454dc456d91SStefano Zampini if (root_data[i]) { 3455dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3456674ae819SStefano Zampini old_index = i; 3457674ae819SStefano Zampini } 3458674ae819SStefano Zampini } 3459674ae819SStefano Zampini } 3460dc456d91SStefano Zampini 3461dc456d91SStefano Zampini /* from roots to leaves */ 3462dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3463dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3464dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3465dc456d91SStefano Zampini 3466dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3467dc456d91SStefano Zampini if (subset_mult) { 3468dc456d91SStefano Zampini const PetscInt* idxs_mult; 3469dc456d91SStefano Zampini PetscInt cum; 3470dc456d91SStefano Zampini 3471dc456d91SStefano Zampini cum = 0; 3472dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3473dc456d91SStefano Zampini for (i=0;i<n;i++) { 3474dc456d91SStefano Zampini PetscInt j; 3475dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3476674ae819SStefano Zampini } 3477dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3478674ae819SStefano Zampini } else { 3479dc456d91SStefano Zampini for (i=0;i<n;i++) { 3480dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3481674ae819SStefano Zampini } 3482674ae819SStefano Zampini } 3483dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3484dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3485674ae819SStefano Zampini PetscFunctionReturn(0); 3486674ae819SStefano Zampini } 34879a7d3425SStefano Zampini 34889a7d3425SStefano Zampini #undef __FUNCT__ 34899a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 34909a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 34919a7d3425SStefano Zampini { 34929a7d3425SStefano Zampini PetscInt i,j; 34939a7d3425SStefano Zampini PetscScalar *alphas; 34949a7d3425SStefano Zampini PetscErrorCode ierr; 34959a7d3425SStefano Zampini 34969a7d3425SStefano Zampini PetscFunctionBegin; 34979a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3498785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 34999a7d3425SStefano Zampini for (i=0;i<n;i++) { 35009a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 35019a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 35029a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 35039a7d3425SStefano Zampini } 35049a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 35059a7d3425SStefano Zampini PetscFunctionReturn(0); 35069a7d3425SStefano Zampini } 35079a7d3425SStefano Zampini 3508e7931f94SStefano Zampini #undef __FUNCT__ 350970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3510b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3511e7931f94SStefano Zampini { 351252e5ac9dSStefano Zampini IS ranks_send_to; 3513e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3514e7931f94SStefano Zampini PetscMPIInt size,rank,color; 351552e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 351652e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 35173837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 35182b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3519e7931f94SStefano Zampini PetscSubcomm subcomm; 352052e5ac9dSStefano Zampini PetscErrorCode ierr; 3521a57a6d2fSStefano Zampini 3522e7931f94SStefano Zampini PetscFunctionBegin; 35232b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 35242b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 35252b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3526e7931f94SStefano Zampini 3527e7931f94SStefano Zampini /* Get info on mapping */ 35283bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 35293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3530e7931f94SStefano Zampini 3531e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3532785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3533e7931f94SStefano Zampini xadj[0] = 0; 3534e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3535785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3536785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3537e7931f94SStefano Zampini 35382b510759SStefano Zampini if (threshold) { 3539d023bfaeSStefano Zampini PetscInt xadj_count = 0; 35402b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3541d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3542d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3543d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3544d023bfaeSStefano Zampini xadj_count++; 3545e7931f94SStefano Zampini } 3546e7931f94SStefano Zampini } 3547d023bfaeSStefano Zampini xadj[1] = xadj_count; 3548c8587f34SStefano Zampini } else { 3549e7931f94SStefano Zampini if (xadj[1]) { 3550e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3551e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3552c8587f34SStefano Zampini } 3553e7931f94SStefano Zampini } 35543bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3555e7931f94SStefano Zampini if (use_square) { 3556e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3557e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3558e7931f94SStefano Zampini } 3559e7931f94SStefano Zampini } 3560e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3561e7931f94SStefano Zampini 35623837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3563e7931f94SStefano Zampini 3564e7931f94SStefano Zampini /* 3565e7931f94SStefano Zampini Restrict work on active processes only. 3566e7931f94SStefano Zampini */ 3567e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3568e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3569e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 35702b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3571d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3572e7931f94SStefano Zampini if (color) { 3573e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3574e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3575e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3576c8587f34SStefano Zampini } else { 357752e5ac9dSStefano Zampini Mat subdomain_adj; 357852e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 357952e5ac9dSStefano Zampini MatPartitioning partitioner; 358052e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 358152e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3582b0c7d250SStefano Zampini PetscBool aggregate; 3583b0c7d250SStefano Zampini 3584306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3585785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3586e7931f94SStefano Zampini prank = rank; 3587306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 35888002ef2cSStefano Zampini /* 3589e7931f94SStefano Zampini for (i=0;i<size;i++) { 3590e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3591c8587f34SStefano Zampini } 35928002ef2cSStefano Zampini */ 3593e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3594e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3595c8587f34SStefano Zampini } 3596e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3597b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3598b0c7d250SStefano Zampini if (aggregate) { 3599b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3600b0c7d250SStefano Zampini PetscMPIInt nrank; 3601b0c7d250SStefano Zampini PetscScalar *vals; 3602b0c7d250SStefano Zampini 3603b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3604b0c7d250SStefano Zampini lrows = 0; 3605b0c7d250SStefano Zampini if (nrank<redprocs) { 3606b0c7d250SStefano Zampini lrows = size/redprocs; 3607b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3608b0c7d250SStefano Zampini } 36095fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3610b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3611b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3612b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3613b0c7d250SStefano Zampini row = nrank; 3614b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3615b0c7d250SStefano Zampini cols = adjncy; 3616b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3617b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3618b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3619b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3620b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 362152e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 362252e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 362352e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3624b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3625b0c7d250SStefano Zampini } else { 3626306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3627b0c7d250SStefano Zampini } 362822b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3629e7931f94SStefano Zampini 3630e7931f94SStefano Zampini /* Partition */ 3631306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3632e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3633e7931f94SStefano Zampini if (use_vwgt) { 36343837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3635e7931f94SStefano Zampini v_wgt[0] = local_size; 3636e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3637c8587f34SStefano Zampini } 363828143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 363928143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3640e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3641e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 364222b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3643e7931f94SStefano Zampini 364452e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 364552e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 364652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 364752e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3648b0c7d250SStefano Zampini if (!redprocs) { 3649b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 365028143c3dSStefano Zampini } else { 3651b0c7d250SStefano Zampini PetscInt idxs[1]; 3652b0c7d250SStefano Zampini PetscMPIInt tag; 3653b0c7d250SStefano Zampini MPI_Request *reqs; 3654b0c7d250SStefano Zampini 3655b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3656b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3657b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3658b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 365928143c3dSStefano Zampini } 3660b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3661b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3662b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3663b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3664e7931f94SStefano Zampini } 366552e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3666e7931f94SStefano Zampini /* clean up */ 3667e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 366852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3669e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3670e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3671e7931f94SStefano Zampini } 3672e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3673e7931f94SStefano Zampini 3674e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3675e7931f94SStefano Zampini i = 1; 3676e7931f94SStefano Zampini if (color) i=0; 3677e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3678e7931f94SStefano Zampini /* get back IS */ 3679e7931f94SStefano Zampini *is_sends = ranks_send_to; 3680e7931f94SStefano Zampini PetscFunctionReturn(0); 3681e7931f94SStefano Zampini } 3682e7931f94SStefano Zampini 3683e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3684e7931f94SStefano Zampini 3685e7931f94SStefano Zampini #undef __FUNCT__ 3686e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 368753a05cb3SStefano 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[]) 3688e7931f94SStefano Zampini { 368970cf5478SStefano Zampini Mat local_mat; 3690e7931f94SStefano Zampini IS is_sends_internal; 36919d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 369228143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 36939d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3694e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3695e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3696e7931f94SStefano Zampini const PetscInt* is_indices; 3697e7931f94SStefano Zampini MatType new_local_type; 3698e7931f94SStefano Zampini /* buffers */ 3699e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 370028143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 37019d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3702e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3703e7931f94SStefano Zampini /* MPI */ 370428143c3dSStefano Zampini MPI_Comm comm,comm_n; 370528143c3dSStefano Zampini PetscSubcomm subcomm; 3706e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 370728143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 370828143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 370928143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 371028143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 371128143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3712e7931f94SStefano Zampini PetscErrorCode ierr; 3713e7931f94SStefano Zampini 3714e7931f94SStefano Zampini PetscFunctionBegin; 371528143c3dSStefano Zampini /* TODO: add missing checks */ 371628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 371728143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 371828143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 371928143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3720e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 372128143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3722e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3723e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3724e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3725e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3726e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 372728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 372870cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 372970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 373028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 373170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 373270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 373370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 373470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 373570cf5478SStefano Zampini } 3736e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3737e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3738e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3739e7931f94SStefano Zampini if (!is_sends) { 374028143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3741b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3742c8587f34SStefano Zampini } else { 3743e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3744e7931f94SStefano Zampini is_sends_internal = is_sends; 3745c8587f34SStefano Zampini } 3746e7931f94SStefano Zampini 3747e7931f94SStefano Zampini /* get comm */ 3748a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3749e7931f94SStefano Zampini 3750e7931f94SStefano Zampini /* compute number of sends */ 3751e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3752e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3753e7931f94SStefano Zampini 3754e7931f94SStefano Zampini /* compute number of receives */ 3755e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3756785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3757e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3758e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3759e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3760e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3761e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3762e7931f94SStefano Zampini 376328143c3dSStefano Zampini /* restrict comm if requested */ 376428143c3dSStefano Zampini subcomm = 0; 376528143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 376628143c3dSStefano Zampini if (restrict_comm) { 3767779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3768779c1cceSStefano Zampini 376928143c3dSStefano Zampini color = 0; 377053a05cb3SStefano Zampini if (restrict_full) { 377153a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 377253a05cb3SStefano Zampini } else { 377353a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 377453a05cb3SStefano Zampini } 377528143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 377628143c3dSStefano Zampini subcommsize = commsize - subcommsize; 377728143c3dSStefano Zampini /* check if reuse has been requested */ 377828143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 377928143c3dSStefano Zampini if (*mat_n) { 378028143c3dSStefano Zampini PetscMPIInt subcommsize2; 378128143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 378228143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 378328143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 378428143c3dSStefano Zampini } else { 378528143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 378628143c3dSStefano Zampini } 378728143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3788779c1cceSStefano Zampini PetscMPIInt rank; 3789779c1cceSStefano Zampini 3790779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 379128143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 379228143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 379328143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3794306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 379528143c3dSStefano Zampini } 379628143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 379728143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 379828143c3dSStefano Zampini } else { 379928143c3dSStefano Zampini comm_n = comm; 380028143c3dSStefano Zampini } 380128143c3dSStefano Zampini 3802e7931f94SStefano Zampini /* prepare send/receive buffers */ 3803785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3804e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3805785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3806e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 380728143c3dSStefano Zampini if (nis) { 3808854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 380928143c3dSStefano Zampini } 3810e7931f94SStefano Zampini 381128143c3dSStefano Zampini /* Get data from local matrices */ 3812e7931f94SStefano Zampini if (!isdense) { 3813a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3814e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3815e7931f94SStefano Zampini /* 3816e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3817e7931f94SStefano Zampini send_buffer_idxs should contain: 3818e7931f94SStefano Zampini - MatType_PRIVATE type 3819e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3820e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3821e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3822e7931f94SStefano Zampini */ 3823e7931f94SStefano Zampini } else { 3824e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 38253bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3826854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3827e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3828e7931f94SStefano Zampini send_buffer_idxs[1] = i; 38293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3830e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 38313bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3832e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3833e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3834e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3835e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3836c8587f34SStefano Zampini } 3837c8587f34SStefano Zampini } 3838e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 383928143c3dSStefano Zampini /* additional is (if any) */ 384028143c3dSStefano Zampini if (nis) { 384128143c3dSStefano Zampini PetscMPIInt psum; 384228143c3dSStefano Zampini PetscInt j; 384328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 384428143c3dSStefano Zampini PetscInt plen; 384528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 384628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 384728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 384828143c3dSStefano Zampini } 3849854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 385028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 385128143c3dSStefano Zampini PetscInt plen; 385228143c3dSStefano Zampini const PetscInt *is_array_idxs; 385328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 385428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 385528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 385628143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 385728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 385828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 385928143c3dSStefano Zampini } 386028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 386128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 386228143c3dSStefano Zampini } 386328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 386428143c3dSStefano Zampini } 386528143c3dSStefano Zampini 3866e7931f94SStefano Zampini buf_size_idxs = 0; 3867e7931f94SStefano Zampini buf_size_vals = 0; 386828143c3dSStefano Zampini buf_size_idxs_is = 0; 3869e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3870e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3871e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 387228143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3873e7931f94SStefano Zampini } 3874785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3875785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 387695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3877e7931f94SStefano Zampini 3878e7931f94SStefano Zampini /* get new tags for clean communications */ 3879e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3880e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 388128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3882e7931f94SStefano Zampini 3883e7931f94SStefano Zampini /* allocate for requests */ 3884785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3885785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 388695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3887785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3888785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 388995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3890e7931f94SStefano Zampini 3891e7931f94SStefano Zampini /* communications */ 3892e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3893e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 389428143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3895e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3896e7931f94SStefano Zampini source_dest = onodes[i]; 3897e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3898e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3899e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3900e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 390128143c3dSStefano Zampini if (nis) { 390228143c3dSStefano 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); 390328143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 390428143c3dSStefano Zampini } 3905e7931f94SStefano Zampini } 3906e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3907e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3909e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 391028143c3dSStefano Zampini if (nis) { 391128143c3dSStefano 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); 391228143c3dSStefano Zampini } 3913e7931f94SStefano Zampini } 3914e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3915e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3916e7931f94SStefano Zampini 3917e7931f94SStefano Zampini /* assemble new l2g map */ 3918e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3919e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39209d30be91SStefano Zampini new_local_rows = 0; 3921e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 39229d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3923e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3924e7931f94SStefano Zampini } 39259d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3926e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39279d30be91SStefano Zampini new_local_rows = 0; 3928e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 39299d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 39309d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3931e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3932e7931f94SStefano Zampini } 39339d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 39349d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3935e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3936e7931f94SStefano Zampini 3937e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3938e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3939e7931f94SStefano 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) */ 3940e7931f94SStefano Zampini if (n_recvs) { 394128143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3942e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3943e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3944e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3945e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3946e7931f94SStefano Zampini break; 3947e7931f94SStefano Zampini } 3948e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3949e7931f94SStefano Zampini } 3950e7931f94SStefano Zampini switch (new_local_type_private) { 395128143c3dSStefano Zampini case MATDENSE_PRIVATE: 395228143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3953e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3954e7931f94SStefano Zampini bs = 1; 395528143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 395628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 395728143c3dSStefano Zampini bs = 1; 395828143c3dSStefano Zampini } 3959e7931f94SStefano Zampini break; 3960e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3961e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3962e7931f94SStefano Zampini bs = 1; 3963e7931f94SStefano Zampini break; 3964e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3965e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3966e7931f94SStefano Zampini break; 3967e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3968e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3969e7931f94SStefano Zampini break; 3970e7931f94SStefano Zampini default: 39719d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3972e7931f94SStefano Zampini break; 3973e7931f94SStefano Zampini } 397428143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 397528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 397628143c3dSStefano Zampini bs = 1; 3977e7931f94SStefano Zampini } 3978e7931f94SStefano Zampini 397970cf5478SStefano Zampini /* create MATIS object if needed */ 398070cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3981e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 3982e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 398370cf5478SStefano Zampini } else { 398470cf5478SStefano Zampini /* it also destroys the local matrices */ 398570cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 398670cf5478SStefano Zampini } 398770cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3988e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 39899d30be91SStefano Zampini 39909d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 39919d30be91SStefano Zampini 39929d30be91SStefano Zampini /* Global to local map of received indices */ 39939d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 39949d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 39959d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 39969d30be91SStefano Zampini 39979d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 39989d30be91SStefano Zampini buf_size_idxs = 0; 39999d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40009d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 40019d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 40029d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 40039d30be91SStefano Zampini } 40049d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 40059d30be91SStefano Zampini 40069d30be91SStefano Zampini /* set preallocation */ 40079d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 40089d30be91SStefano Zampini if (!newisdense) { 40099d30be91SStefano Zampini PetscInt *new_local_nnz=0; 40109d30be91SStefano Zampini 40119d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 40129d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 40139d30be91SStefano Zampini if (n_recvs) { 40149d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 40159d30be91SStefano Zampini } 40169d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40179d30be91SStefano Zampini PetscInt j; 40189d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 40199d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 40209d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 40219d30be91SStefano Zampini } 40229d30be91SStefano Zampini } else { 40239d30be91SStefano Zampini /* TODO */ 40249d30be91SStefano Zampini } 40259d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 40269d30be91SStefano Zampini } 40279d30be91SStefano Zampini if (new_local_nnz) { 40289d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 40299d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 40309d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 40319d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 40329d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 40339d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 40349d30be91SStefano Zampini } else { 40359d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 40369d30be91SStefano Zampini } 40379d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 40389d30be91SStefano Zampini } else { 40399d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 40409d30be91SStefano Zampini } 4041e7931f94SStefano Zampini 4042e7931f94SStefano Zampini /* set values */ 4043e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 40449d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4045e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4046e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4047e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 40489d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4049e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4050e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4051e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 405228143c3dSStefano Zampini } else { 405328143c3dSStefano Zampini /* TODO */ 4054e7931f94SStefano Zampini } 4055e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4056e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4057e7931f94SStefano Zampini } 4058e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4059e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 406070cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 406170cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40629d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 40639d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4064e7931f94SStefano Zampini 4065dfd14d43SStefano Zampini #if 0 406628143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4067e7931f94SStefano Zampini Vec lvec,rvec; 4068e7931f94SStefano Zampini PetscReal infty_error; 4069e7931f94SStefano Zampini 40702a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4071e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4072e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4073e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 407470cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4075e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4076e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4077e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4078e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4079e7931f94SStefano Zampini } 408028143c3dSStefano Zampini #endif 4081e7931f94SStefano Zampini 408228143c3dSStefano Zampini /* assemble new additional is (if any) */ 408328143c3dSStefano Zampini if (nis) { 408428143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 408528143c3dSStefano Zampini 408628143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4087854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 408828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 408928143c3dSStefano Zampini psum = 0; 409028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 409128143c3dSStefano Zampini for (j=0;j<nis;j++) { 409228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 409328143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 409428143c3dSStefano Zampini psum += plen; 409528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 409628143c3dSStefano Zampini } 409728143c3dSStefano Zampini } 4098854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4099854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 410028143c3dSStefano Zampini for (i=1;i<nis;i++) { 410128143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 410228143c3dSStefano Zampini } 410328143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 410428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 410528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 410628143c3dSStefano Zampini for (j=0;j<nis;j++) { 410728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 410828143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 410928143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 411028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 411128143c3dSStefano Zampini } 411228143c3dSStefano Zampini } 411328143c3dSStefano Zampini for (i=0;i<nis;i++) { 411428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 411528143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 411628143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 411728143c3dSStefano Zampini } 411828143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 411928143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 412028143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 412128143c3dSStefano Zampini } 4122e7931f94SStefano Zampini /* free workspace */ 412328143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4124e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4125e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4126e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4127e7931f94SStefano Zampini if (isdense) { 4128e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4129e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4130e7931f94SStefano Zampini } else { 4131e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4132e7931f94SStefano Zampini } 413328143c3dSStefano Zampini if (nis) { 413428143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 413528143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 413628143c3dSStefano Zampini } 4137e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4138e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 413928143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4140e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4141e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 414228143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4143e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4144e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4145e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4146e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4147e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 414828143c3dSStefano Zampini if (nis) { 414928143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 415028143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 415128143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 415228143c3dSStefano Zampini } 415328143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 415428143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 415528143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 415628143c3dSStefano Zampini for (i=0;i<nis;i++) { 415728143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 415828143c3dSStefano Zampini } 415953a05cb3SStefano Zampini *mat_n = NULL; 416028143c3dSStefano Zampini } 4161e7931f94SStefano Zampini PetscFunctionReturn(0); 4162e7931f94SStefano Zampini } 4163a57a6d2fSStefano Zampini 416412edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4165af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 416612edc857SStefano Zampini 4167c8587f34SStefano Zampini #undef __FUNCT__ 4168c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4169c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4170c8587f34SStefano Zampini { 4171c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4172c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 417320a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 41749881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 417520a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 41766e683305SStefano Zampini IS coarse_is,*isarray; 41776e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 41786e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4179f9eb5b7dSStefano Zampini PC pc_temp; 4180c8587f34SStefano Zampini PCType coarse_pc_type; 4181c8587f34SStefano Zampini KSPType coarse_ksp_type; 4182f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 41834f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 41846e683305SStefano Zampini Mat t_coarse_mat_is; 41856e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 41866e683305SStefano Zampini PetscMPIInt all_procs; 418774e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 418868457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 418922bc73bbSStefano Zampini PetscScalar *array; 41909881197aSStefano Zampini PetscErrorCode ierr; 4191fdc09c96SStefano Zampini 4192c8587f34SStefano Zampini PetscFunctionBegin; 4193c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 419468457ee5SStefano 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 */ 4195fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 41965a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4197fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4198f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4199f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4200f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4201fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 420251bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 420351bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4204dc4bcba2SStefano Zampini PC pc; 4205dc4bcba2SStefano Zampini PetscBool isbddc; 4206dc4bcba2SStefano Zampini 4207dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4208dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4209dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4210dc4bcba2SStefano Zampini if (isbddc) { 4211dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4212dc4bcba2SStefano Zampini } 4213727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4214fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4215fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4216fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4217f4ddd8eeSStefano Zampini } 4218fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4219fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4220f4ddd8eeSStefano Zampini } 422170cf5478SStefano Zampini /* reset any subassembling information */ 422270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 42236e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42246e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4225fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4226f4ddd8eeSStefano Zampini } 4227c8587f34SStefano Zampini 42286e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 42292b510759SStefano Zampini im_active = !!(pcis->n); 42302b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 42316e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 42326e683305SStefano Zampini void_procs = all_procs-active_procs; 42336e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 423474e2c79eSStefano Zampini redist = PETSC_FALSE; 423522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 42366e683305SStefano Zampini csin_ml = PETSC_TRUE; 42376e683305SStefano Zampini ncoarse_ml = void_procs; 4238779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4239779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 42406e683305SStefano Zampini csin_ds = PETSC_TRUE; 424118a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 424218a45a71SStefano Zampini redist = PETSC_TRUE; 424318a45a71SStefano Zampini } else { 42446e683305SStefano Zampini csin_ds = PETSC_TRUE; 4245779c1cceSStefano Zampini ncoarse_ds = active_procs; 4246779c1cceSStefano Zampini redist = PETSC_TRUE; 424718a45a71SStefano Zampini } 42486e683305SStefano Zampini } else { 42496e683305SStefano Zampini csin_ml = PETSC_FALSE; 42506e683305SStefano Zampini ncoarse_ml = all_procs; 42516e683305SStefano Zampini if (void_procs) { 42526e683305SStefano Zampini csin_ds = PETSC_TRUE; 42536e683305SStefano Zampini ncoarse_ds = void_procs; 42546e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 42556e683305SStefano Zampini } else { 4256779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 425774e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 425874e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 425974e2c79eSStefano Zampini redist = PETSC_TRUE; 426074e2c79eSStefano Zampini } else { 42616e683305SStefano Zampini csin_ds = PETSC_FALSE; 42626e683305SStefano Zampini ncoarse_ds = all_procs; 42636e683305SStefano Zampini } 42646e683305SStefano Zampini } 426574e2c79eSStefano Zampini } 42666e683305SStefano Zampini 42676e683305SStefano Zampini /* 42686e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 42696e683305SStefano Zampini - we have not exceeded the number of levels requested 42706e683305SStefano Zampini - we can actually subassemble the active processes 42716e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 42726e683305SStefano Zampini */ 42736e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 42746e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 42756e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 42766e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 42776e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4278f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 42792b510759SStefano Zampini } else { 4280f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4281c8587f34SStefano Zampini } 4282c8587f34SStefano Zampini } 42836e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 42846e683305SStefano Zampini if (multilevel_allowed) { 42856e683305SStefano Zampini ncoarse = ncoarse_ml; 42866e683305SStefano Zampini csin = csin_ml; 428758da7f69SStefano Zampini redist = PETSC_FALSE; 42886e683305SStefano Zampini } else { 42896e683305SStefano Zampini ncoarse = ncoarse_ds; 42906e683305SStefano Zampini csin = csin_ds; 42916e683305SStefano Zampini } 4292e7931f94SStefano Zampini 4293abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4294abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4295abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4296abbbba34SStefano Zampini 4297abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 429822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 429922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 430022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 430122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4302b9b85e73SStefano Zampini #if 0 4303b9b85e73SStefano Zampini { 4304b9b85e73SStefano Zampini PetscViewer viewer; 4305b9b85e73SStefano Zampini char filename[256]; 4306b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4307b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4308b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4309b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4310b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4311b9b85e73SStefano Zampini } 4312b9b85e73SStefano Zampini #endif 4313e176bc59SStefano 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); 43146e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 43156e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43166e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4317abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4318abbbba34SStefano Zampini 43196e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 43206e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 43216e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 43226e683305SStefano Zampini const PetscInt *idxs; 43236e683305SStefano Zampini ISLocalToGlobalMapping tmap; 43246e683305SStefano Zampini 43256e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 43260be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 43276e683305SStefano Zampini /* allocate space for temporary storage */ 4328854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4329854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 43306e683305SStefano Zampini /* allocate for IS array */ 43316e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 43326e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 43336e683305SStefano Zampini nis = nisdofs + nisneu; 4334854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 43356e683305SStefano Zampini /* dofs splitting */ 43366e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 43376e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 43386e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 43396e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 43406e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 43416e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 43426e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 43436e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 43446e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 43456e683305SStefano Zampini } 43466e683305SStefano Zampini /* neumann boundaries */ 43476e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 43486e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 43496e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 43506e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 43516e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 43526e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 43536e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 43546e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 43556e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 43566e683305SStefano Zampini } 43576e683305SStefano Zampini /* free memory */ 43586e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 43596e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 43606e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 43616e683305SStefano Zampini } else { 43626e683305SStefano Zampini nis = 0; 43636e683305SStefano Zampini nisdofs = 0; 43646e683305SStefano Zampini nisneu = 0; 43656e683305SStefano Zampini isarray = NULL; 43666e683305SStefano Zampini } 43676e683305SStefano Zampini /* destroy no longer needed map */ 43686e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 43696e683305SStefano Zampini 43706e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 43716e683305SStefano Zampini coarse_mat_is = NULL; 43726e683305SStefano Zampini if (csin) { 43736e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 437474e2c79eSStefano Zampini if (redist) { 437574e2c79eSStefano Zampini PetscMPIInt rank; 4376779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 437774e2c79eSStefano Zampini 437874e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 437958da7f69SStefano Zampini spc = active_procs/ncoarse; 438058da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4381779c1cceSStefano Zampini if (im_active) { 4382779c1cceSStefano Zampini destsize = 1; 438374e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 438474e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 438574e2c79eSStefano Zampini } else { 438674e2c79eSStefano Zampini dest[0] = rank/(spc+1); 438774e2c79eSStefano Zampini } 438874e2c79eSStefano Zampini } else { 4389779c1cceSStefano Zampini destsize = 0; 43906e683305SStefano Zampini } 4391779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4392779c1cceSStefano Zampini } else if (csin_type_simple) { 43936e683305SStefano Zampini PetscMPIInt rank; 43946e683305SStefano Zampini PetscInt issize,isidx; 4395779c1cceSStefano Zampini 43966e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 43976e683305SStefano Zampini if (im_active) { 43986e683305SStefano Zampini issize = 1; 43996e683305SStefano Zampini isidx = (PetscInt)rank; 44006e683305SStefano Zampini } else { 44016e683305SStefano Zampini issize = 0; 44026e683305SStefano Zampini isidx = -1; 44036e683305SStefano Zampini } 44046e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4405779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4406b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 44076e683305SStefano Zampini } 4408779c1cceSStefano Zampini 4409779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4410779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4411779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4412779c1cceSStefano Zampini PetscInt *coarse_candidates; 4413779c1cceSStefano Zampini const PetscInt* tisindices; 4414779c1cceSStefano Zampini 4415779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4416779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4417779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4418779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4419779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4420779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4421779c1cceSStefano Zampini } 4422779c1cceSStefano Zampini } 4423779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4424779c1cceSStefano Zampini 4425779c1cceSStefano Zampini 44266e683305SStefano Zampini if (pcbddc->dbg_flag) { 44276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 44296e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 44306e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4431779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 44326e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 44336e683305SStefano Zampini } 44346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 44356e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44366e683305SStefano Zampini } 44376e683305SStefano Zampini /* shift the pattern on coarse candidates */ 44386e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 44396e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4440854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 44416e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 44426e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 44436e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 44446e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 44456e683305SStefano Zampini } 44466e683305SStefano Zampini if (pcbddc->dbg_flag) { 44476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 44496e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 44506e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44516e683305SStefano Zampini } 4452779c1cceSStefano Zampini } 44536e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 445453a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 445553a05cb3SStefano 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); 445653a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 445753a05cb3SStefano 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); 445853a05cb3SStefano Zampini } 44596e683305SStefano Zampini } else { 44606e683305SStefano Zampini if (pcbddc->dbg_flag) { 44616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44626e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 44636e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44646e683305SStefano Zampini } 44656e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 44666e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 44676e683305SStefano Zampini } 44686e683305SStefano Zampini 44696e683305SStefano Zampini /* create local to global scatters for coarse problem */ 447068457ee5SStefano Zampini if (compute_vecs) { 44716e683305SStefano Zampini PetscInt lrows; 44726e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 44736e683305SStefano Zampini if (coarse_mat_is) { 44746e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 44756e683305SStefano Zampini } else { 44766e683305SStefano Zampini lrows = 0; 44776e683305SStefano Zampini } 44786e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 44796e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 44806e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 44816e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 44826e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 44836e683305SStefano Zampini } 44846e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 44856e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4486c8587f34SStefano Zampini 4487f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4488f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4489f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4490f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4491f9eb5b7dSStefano Zampini } else { 4492f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4493f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4494c8587f34SStefano Zampini } 4495c8587f34SStefano Zampini 44966e683305SStefano Zampini /* print some info if requested */ 44976e683305SStefano Zampini if (pcbddc->dbg_flag) { 44986e683305SStefano Zampini if (!multilevel_allowed) { 44996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45006e683305SStefano Zampini if (multilevel_requested) { 45016e683305SStefano 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); 45026e683305SStefano Zampini } else if (pcbddc->max_levels) { 45036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 45046e683305SStefano Zampini } 45056e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45066e683305SStefano Zampini } 45076e683305SStefano Zampini } 45086e683305SStefano Zampini 4509f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 45106e683305SStefano Zampini if (coarse_mat_is) { 45116e683305SStefano Zampini MatReuse coarse_mat_reuse; 45126a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 45136e683305SStefano Zampini if (pcbddc->dbg_flag) { 45146e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 45156e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 45166e683305SStefano Zampini } 4517f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4518312be037SStefano Zampini char prefix[256],str_level[16]; 4519e604994aSStefano Zampini size_t len; 45206e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4521422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4522c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4523f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 45245f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4525c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 45266e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4527c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4528c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4529e604994aSStefano Zampini /* prefix */ 4530e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4531e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4532e604994aSStefano Zampini if (!pcbddc->current_level) { 4533e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4534e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4535c8587f34SStefano Zampini } else { 4536e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4537312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4538312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 453934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4540312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4541e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4542e604994aSStefano Zampini } 4543e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 45443e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 45453e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 45463e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 45473e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4548f9eb5b7dSStefano Zampini /* allow user customization */ 4549f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 45503e3c6dadSStefano Zampini } 45513e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 455251bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 45533e3c6dadSStefano Zampini if (nisdofs) { 45543e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 45553e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 45563e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 45573e3c6dadSStefano Zampini } 45583e3c6dadSStefano Zampini } 45593e3c6dadSStefano Zampini if (nisneu) { 45603e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 45613e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4562312be037SStefano Zampini } 4563f9eb5b7dSStefano Zampini 4564f9eb5b7dSStefano Zampini /* get some info after set from options */ 4565f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4566f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 45674f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 45686e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4569f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4570f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4571f9eb5b7dSStefano Zampini } 457239f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 45734f3a063dSStefano Zampini if (isredundant) { 45744f3a063dSStefano Zampini KSP inner_ksp; 45754f3a063dSStefano Zampini PC inner_pc; 45764f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 45774f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 45784f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 45794f3a063dSStefano Zampini } 4580f9eb5b7dSStefano Zampini 4581f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4582fa7f1dd8SStefano Zampini if (coarse_reuse) { 458381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4584fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 45856e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4586fa7f1dd8SStefano Zampini } else { 45876e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4588fa7f1dd8SStefano Zampini } 4589c8587f34SStefano Zampini if (isbddc || isnn) { 459022bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 459170cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4592b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 459322b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 45946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 45966e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 45976e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 459822b6e8a2SStefano Zampini } 459970cf5478SStefano Zampini } 460053a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 460170cf5478SStefano Zampini } else { 460222bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 460322bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 460422bc73bbSStefano Zampini } 460522bc73bbSStefano Zampini } else { 46062e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4607c8587f34SStefano Zampini } 4608c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4609c8587f34SStefano Zampini 46103301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 46115a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 46123301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 46133301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 46143301b35fSStefano Zampini } 46153301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 46163301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 46173301b35fSStefano Zampini } 46183301b35fSStefano Zampini if (pc->pmat->spd_set) { 46193301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 46203301b35fSStefano Zampini } 46216e683305SStefano Zampini /* set operators */ 46225f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 46236e683305SStefano Zampini if (pcbddc->dbg_flag) { 46246e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 46256e683305SStefano Zampini } 46266e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 46276e683305SStefano Zampini coarse_mat = 0; 46286e683305SStefano Zampini } 46296e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4630b9b85e73SStefano Zampini #if 0 4631b9b85e73SStefano Zampini { 4632b9b85e73SStefano Zampini PetscViewer viewer; 4633b9b85e73SStefano Zampini char filename[256]; 4634b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4635b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4636b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4637b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4638b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4639b9b85e73SStefano Zampini } 4640b9b85e73SStefano Zampini #endif 4641c8587f34SStefano Zampini 4642c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4643298c0119SStefano Zampini #if 0 4644c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4645c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 464698a51de6SStefano Zampini } 4647298c0119SStefano Zampini #endif 464898a51de6SStefano Zampini 464998a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 465098a51de6SStefano Zampini Vec crhs,csol; 465198a51de6SStefano Zampini PetscBool ispreonly; 465204708bb6SStefano Zampini 465398a51de6SStefano Zampini if (CoarseNullSpace) { 4654c8587f34SStefano Zampini if (isbddc) { 4655c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4656c8587f34SStefano Zampini } else { 46575fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4658c8587f34SStefano Zampini } 4659c8587f34SStefano Zampini } 4660f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4661f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4662f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4663f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 46646e683305SStefano Zampini /* hack */ 4665f347579bSStefano Zampini if (!csol) { 46662a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4667f9eb5b7dSStefano Zampini } 4668f347579bSStefano Zampini if (!crhs) { 46692a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4670f347579bSStefano Zampini } 4671cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4672cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 46736e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4674c8587f34SStefano Zampini KSP check_ksp; 46752b510759SStefano Zampini KSPType check_ksp_type; 4676c8587f34SStefano Zampini PC check_pc; 46776e683305SStefano Zampini Vec check_vec,coarse_vec; 46786a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 46792b510759SStefano Zampini PetscInt its; 46806e683305SStefano Zampini PetscBool compute_eigs; 46816e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 46826e683305SStefano Zampini PetscInt neigs; 46838e185a42SStefano Zampini const char *prefix; 4684c8587f34SStefano Zampini 46852b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 46866e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4687422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 468823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4689f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 46902b510759SStefano Zampini if (ispreonly) { 46912b510759SStefano Zampini check_ksp_type = KSPPREONLY; 46926e683305SStefano Zampini compute_eigs = PETSC_FALSE; 46932b510759SStefano Zampini } else { 4694cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 46956e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4696c8587f34SStefano Zampini } 4697c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 46986e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 46996e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 47006e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4701a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4702a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4703a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4704a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4705c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4706c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4707c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4708c8587f34SStefano Zampini /* create random vec */ 47096e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 47106e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4711c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4712c8587f34SStefano Zampini if (CoarseNullSpace) { 4713c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4714c8587f34SStefano Zampini } 47156e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4716c8587f34SStefano Zampini /* solve coarse problem */ 47176e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4718c8587f34SStefano Zampini if (CoarseNullSpace) { 47196e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4720c8587f34SStefano Zampini } 4721cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 47226e683305SStefano Zampini if (compute_eigs) { 4723854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4724854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 47256e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 47266e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 47276e683305SStefano Zampini lambda_min = eigs_r[0]; 47286e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 47296e683305SStefano Zampini if (lambda_max>lambda_min) { 4730cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4731cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4732cbcc2c2aSStefano Zampini } 4733c8587f34SStefano Zampini } 4734c8587f34SStefano Zampini } 4735cbcc2c2aSStefano Zampini 4736c8587f34SStefano Zampini /* check coarse problem residual error */ 47376e683305SStefano Zampini if (pcbddc->dbg_flag) { 47386e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 47396e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 47406e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4741c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 47426e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 47436e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4744c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4745779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 47466e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 47476e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 47486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 47496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 47506e683305SStefano Zampini if (compute_eigs) { 47516e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4752deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4753c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 47546e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 47556e683305SStefano 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); 47566e683305SStefano Zampini for (i=0;i<neigs;i++) { 47576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4758c8587f34SStefano Zampini } 47596e683305SStefano Zampini } 47606e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 47616e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 47626e683305SStefano Zampini } 4763c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 47646e683305SStefano Zampini if (compute_eigs) { 47656e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 47666e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4767c8587f34SStefano Zampini } 47686e683305SStefano Zampini } 47696e683305SStefano Zampini } 4770cbcc2c2aSStefano Zampini /* print additional info */ 4771cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 47726e683305SStefano Zampini /* waits until all processes reaches this point */ 47736e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4774cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4775cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4776cbcc2c2aSStefano Zampini } 4777cbcc2c2aSStefano Zampini 47782b510759SStefano Zampini /* free memory */ 4779c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4780fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4781c8587f34SStefano Zampini PetscFunctionReturn(0); 4782c8587f34SStefano Zampini } 4783674ae819SStefano Zampini 4784f34684f1SStefano Zampini #undef __FUNCT__ 4785f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4786f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4787f34684f1SStefano Zampini { 4788f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4789f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4790f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4791dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4792dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 479373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4794dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4795f34684f1SStefano Zampini PetscErrorCode ierr; 4796f34684f1SStefano Zampini 4797f34684f1SStefano Zampini PetscFunctionBegin; 4798f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 47990e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 48000e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4801727cdba6SStefano Zampini } 4802dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 48033bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4804dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4805dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4806dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4807dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4808dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4809dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 48100e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 48110e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 48120e6343abSStefano Zampini } 4813dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4814dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4815dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4816dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4817dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4818f34684f1SStefano Zampini 4819f34684f1SStefano Zampini /* check numbering */ 4820f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4821*019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 4822dc456d91SStefano Zampini PetscInt i; 4823b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4824f34684f1SStefano Zampini 4825f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4826f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4827f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 48280fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4829*019a44ceSStefano Zampini /* counter */ 4830*019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4831*019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4832*019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4833*019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4834*019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4835*019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4836*019a44ceSStefano Zampini 4837f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4838f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4839727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4840f34684f1SStefano Zampini } 4841f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4842f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4843f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4844e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4845e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4846e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4847e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4848f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4849*019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4850f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4851*019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 4852*019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 4853*019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 4854b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4855*019a44ceSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a %d processes instead of %d!\n",PetscGlobalRank,i,owned,neigh);CHKERRQ(ierr); 4856f34684f1SStefano Zampini } 4857f34684f1SStefano Zampini } 4858*019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4859b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4860f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4861f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4862f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4863f34684f1SStefano Zampini } 4864f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4865f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4866e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4867e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4868f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4869f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4870b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4871ca8b9ea9SStefano Zampini PetscInt *gidxs; 4872ca8b9ea9SStefano Zampini 4873ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 48743bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4875f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4876f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4877f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4878f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 48794bc2dc4bSStefano 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); 4880f34684f1SStefano Zampini } 4881f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4882ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4883f34684f1SStefano Zampini } 4884f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4885302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4886f34684f1SStefano Zampini } 48878bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4888f34684f1SStefano Zampini /* get back data */ 4889f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4890f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4891674ae819SStefano Zampini PetscFunctionReturn(0); 4892674ae819SStefano Zampini } 4893674ae819SStefano Zampini 4894e456f2a8SStefano Zampini #undef __FUNCT__ 4895e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4896a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4897e456f2a8SStefano Zampini { 4898e456f2a8SStefano Zampini IS localis_t; 4899a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4900e456f2a8SStefano Zampini PetscScalar *vals; 4901e456f2a8SStefano Zampini PetscErrorCode ierr; 4902e456f2a8SStefano Zampini 4903e456f2a8SStefano Zampini PetscFunctionBegin; 4904a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4905e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4906854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4907e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4908e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4909a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4910a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 49111035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4912a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 49131035eff8SStefano Zampini } 4914a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4915e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4916e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4917a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4918a7dc3881SStefano Zampini /* now compute set in local ordering */ 4919a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4920a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4921a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4922a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4923a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4924ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4925e456f2a8SStefano Zampini lsize++; 4926e456f2a8SStefano Zampini } 4927e456f2a8SStefano Zampini } 4928854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4929a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4930ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4931e456f2a8SStefano Zampini idxs[lsize++] = i; 4932e456f2a8SStefano Zampini } 4933e456f2a8SStefano Zampini } 4934a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4935a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4936e456f2a8SStefano Zampini *localis = localis_t; 4937e456f2a8SStefano Zampini PetscFunctionReturn(0); 4938e456f2a8SStefano Zampini } 4939906d46d4SStefano Zampini 4940906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4941906d46d4SStefano Zampini #undef __FUNCT__ 4942906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4943906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4944906d46d4SStefano Zampini { 4945906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4946906d46d4SStefano Zampini PetscErrorCode ierr; 4947906d46d4SStefano Zampini 4948906d46d4SStefano Zampini PetscFunctionBegin; 4949906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4950906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4951906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4952906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4953906d46d4SStefano Zampini PetscFunctionReturn(0); 4954906d46d4SStefano Zampini } 4955906d46d4SStefano Zampini 4956906d46d4SStefano Zampini #undef __FUNCT__ 4957906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4958906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4959906d46d4SStefano Zampini { 4960906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4961906d46d4SStefano Zampini PetscErrorCode ierr; 4962906d46d4SStefano Zampini 4963906d46d4SStefano Zampini PetscFunctionBegin; 4964906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4965906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4966906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4967906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4968906d46d4SStefano Zampini PetscFunctionReturn(0); 4969906d46d4SStefano Zampini } 4970b96c3477SStefano Zampini 4971b96c3477SStefano Zampini #undef __FUNCT__ 4972b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 497308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4974b96c3477SStefano Zampini { 4975a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4976b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4977b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4978a64f4aa4SStefano Zampini Mat S_j; 4979b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4980b96c3477SStefano Zampini PetscBool free_used_adj; 4981b96c3477SStefano Zampini PetscErrorCode ierr; 4982b96c3477SStefano Zampini 4983b96c3477SStefano Zampini PetscFunctionBegin; 4984b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4985b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 498608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4987b96c3477SStefano Zampini used_xadj = NULL; 4988b96c3477SStefano Zampini used_adjncy = NULL; 4989b96c3477SStefano Zampini } else { 499008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 499108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 499208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 499308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4994b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4995b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4996b96c3477SStefano Zampini } else { 49972fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4998b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4999b96c3477SStefano Zampini PetscInt nvtxs; 5000b96c3477SStefano Zampini 50012fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 50022fffb893SStefano Zampini if (flg_row) { 5003b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5004b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5005b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5006b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 50072fffb893SStefano Zampini } else { 50082fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 50092fffb893SStefano Zampini used_xadj = NULL; 50102fffb893SStefano Zampini used_adjncy = NULL; 50112fffb893SStefano Zampini } 50122fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5013b96c3477SStefano Zampini } 5014b96c3477SStefano Zampini } 5015d5574798SStefano Zampini 5016d5574798SStefano Zampini /* setup sub_schurs data */ 5017a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5018a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5019a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5020a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 502106a4e24aSStefano 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); 5022a64f4aa4SStefano Zampini } else { 50236816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 502404708bb6SStefano Zampini PetscBool isseqaij; 50255feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 50265feab87aSStefano Zampini PetscInt n_vertices; 50275feab87aSStefano Zampini 50285feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 50292034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 50305feab87aSStefano Zampini } 503104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 503204708bb6SStefano Zampini if (!isseqaij) { 503304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 503404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 503504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 503604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 503704708bb6SStefano Zampini } else { 503804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 503904708bb6SStefano Zampini } 504004708bb6SStefano Zampini } 504106a4e24aSStefano 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); 5042a64f4aa4SStefano Zampini } 5043a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5044b96c3477SStefano Zampini 5045b96c3477SStefano Zampini /* free adjacency */ 5046b96c3477SStefano Zampini if (free_used_adj) { 5047b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5048b96c3477SStefano Zampini } 5049b96c3477SStefano Zampini PetscFunctionReturn(0); 5050b96c3477SStefano Zampini } 5051b96c3477SStefano Zampini 5052b96c3477SStefano Zampini #undef __FUNCT__ 5053b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 505408122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5055b96c3477SStefano Zampini { 5056b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5057b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5058b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5059b96c3477SStefano Zampini PCBDDCGraph graph; 5060b96c3477SStefano Zampini PetscErrorCode ierr; 5061b96c3477SStefano Zampini 5062b96c3477SStefano Zampini PetscFunctionBegin; 5063b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 506408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 50653301b35fSStefano Zampini IS verticesIS,verticescomm; 50663301b35fSStefano Zampini PetscInt vsize,*idxs; 5067b96c3477SStefano Zampini 5068b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 50693301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 50703301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 50713301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 50723301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 50733301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5074b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 50757fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 50763301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 50773301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5078b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5079b96c3477SStefano Zampini /* 5080b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5081b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5082b96c3477SStefano Zampini } 5083b96c3477SStefano Zampini */ 5084b96c3477SStefano Zampini } else { 5085b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5086b96c3477SStefano Zampini } 5087b96c3477SStefano Zampini 5088b96c3477SStefano Zampini /* sub_schurs init */ 5089a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5090a64f4aa4SStefano Zampini 5091b96c3477SStefano Zampini /* free graph struct */ 509208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5093b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5094b96c3477SStefano Zampini } 5095b96c3477SStefano Zampini PetscFunctionReturn(0); 5096b96c3477SStefano Zampini } 5097