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__ 9efc2fbd9SStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushP0" 10efc2fbd9SStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushP0(PC pc, Vec v, PetscBool pop) 11efc2fbd9SStefano Zampini { 12efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13efc2fbd9SStefano Zampini PetscErrorCode ierr; 14efc2fbd9SStefano Zampini 15efc2fbd9SStefano Zampini PetscFunctionBegin; 16efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 17efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 18efc2fbd9SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,1,NULL,PETSC_OWN_POINTER,&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 19efc2fbd9SStefano Zampini } 20efc2fbd9SStefano Zampini if (pop) { /* use SF to get values */ 21efc2fbd9SStefano Zampini PetscScalar *array; 22efc2fbd9SStefano Zampini 23efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 24efc2fbd9SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 25efc2fbd9SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 26efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 27efc2fbd9SStefano Zampini } else { /* use VecSetValue */ 28efc2fbd9SStefano Zampini ierr = VecSetValue(v,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 29efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 30efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 31efc2fbd9SStefano Zampini } 32efc2fbd9SStefano Zampini PetscFunctionReturn(0); 33efc2fbd9SStefano Zampini } 34efc2fbd9SStefano Zampini 35efc2fbd9SStefano Zampini #undef __FUNCT__ 36c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 37c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 38c263805aSStefano Zampini { 39c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 40c263805aSStefano Zampini PetscErrorCode ierr; 41c263805aSStefano Zampini 42c263805aSStefano Zampini PetscFunctionBegin; 43c263805aSStefano Zampini /* TODO: add error checking 44c263805aSStefano Zampini - avoid nested pop (or push) calls. 45c263805aSStefano Zampini - cannot push before pop. 46c263805aSStefano Zampini */ 47efc2fbd9SStefano Zampini if (!pcbddc->zerodiag) { 48efc2fbd9SStefano Zampini PetscFunctionReturn(0); 49efc2fbd9SStefano Zampini } 50c263805aSStefano Zampini if (pop) { 51c263805aSStefano Zampini const PetscInt *cB0_cols; 52c263805aSStefano Zampini PetscInt cB0_ncol; 53c263805aSStefano Zampini const PetscScalar *cB0_vals; 54c263805aSStefano Zampini 55c263805aSStefano Zampini /* extract B_0 */ 56efc2fbd9SStefano Zampini ierr = MatGetRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 57c263805aSStefano Zampini pcbddc->B0_ncol = cB0_ncol; 58c263805aSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 59c263805aSStefano Zampini ierr = PetscMalloc2(cB0_ncol,&pcbddc->B0_cols,cB0_ncol,&pcbddc->B0_vals);CHKERRQ(ierr); 60c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_cols,cB0_cols,cB0_ncol*sizeof(PetscInt));CHKERRQ(ierr); 61c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_vals,cB0_vals,cB0_ncol*sizeof(PetscScalar));CHKERRQ(ierr); 62efc2fbd9SStefano Zampini ierr = MatRestoreRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 63c263805aSStefano Zampini /* remove rows and cols from local problem */ 64c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 65efc2fbd9SStefano Zampini ierr = MatZeroRowsColumns(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,1.,NULL,NULL);CHKERRQ(ierr); 66c263805aSStefano Zampini } else { /* push */ 67efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,pcbddc->B0_ncol,pcbddc->B0_cols,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 68efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,pcbddc->B0_ncol,pcbddc->B0_cols,1,&pcbddc->benign_p0_lidx,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 69efc2fbd9SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx,pcbddc->benign_p0_lidx,0.0,INSERT_VALUES);CHKERRQ(ierr); 70c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 71c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 72c263805aSStefano Zampini } 73c263805aSStefano Zampini PetscFunctionReturn(0); 74c263805aSStefano Zampini } 75c263805aSStefano Zampini 76c263805aSStefano Zampini #undef __FUNCT__ 77b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 7808122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 79b1b3d7a2SStefano Zampini { 80b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 8208122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 8308122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 8408122e43SStefano Zampini PetscScalar *work,lwork; 8508122e43SStefano Zampini PetscScalar *St,*S,*eigv; 8608122e43SStefano Zampini PetscScalar *Sarray,*Starray; 8708122e43SStefano Zampini PetscReal *eigs,thresh; 881b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 89f6f667cfSStefano Zampini PetscBool allocated_S_St; 9008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9108122e43SStefano Zampini PetscReal *rwork; 9208122e43SStefano Zampini #endif 93b1b3d7a2SStefano Zampini PetscErrorCode ierr; 94b1b3d7a2SStefano Zampini 95b1b3d7a2SStefano Zampini PetscFunctionBegin; 9608122e43SStefano Zampini if (!sub_schurs->use_mumps) { 9708122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 9808122e43SStefano Zampini } 9908122e43SStefano Zampini 10006a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 10106a4e24aSStefano 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); 10206a4e24aSStefano Zampini } 10306a4e24aSStefano Zampini 104fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 105fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 106fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 107fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 108fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 109fd14bc51SStefano Zampini } 110fd14bc51SStefano Zampini 111e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 112e496cd5dSStefano 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); 113e496cd5dSStefano Zampini } 114e496cd5dSStefano Zampini 11508122e43SStefano Zampini /* max size of subsets */ 11608122e43SStefano Zampini mss = 0; 11708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 11808122e43SStefano Zampini PetscInt subset_size; 119862806e4SStefano Zampini 12008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 12108122e43SStefano Zampini mss = PetscMax(mss,subset_size); 12208122e43SStefano Zampini } 12308122e43SStefano Zampini 12408122e43SStefano Zampini /* min/max and threshold */ 12508122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 126f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 12708122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 128f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 129f6f667cfSStefano Zampini if (nmin) { 130f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 131f6f667cfSStefano Zampini } 13208122e43SStefano Zampini 13308122e43SStefano Zampini /* allocate lapack workspace */ 13408122e43SStefano Zampini cum = cum2 = 0; 13508122e43SStefano Zampini maxneigs = 0; 13608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 13708122e43SStefano Zampini PetscInt n,subset_size; 138f6f667cfSStefano Zampini 13908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 14008122e43SStefano Zampini n = PetscMin(subset_size,nmax); 1419162d606SStefano Zampini cum += subset_size; 1429162d606SStefano Zampini cum2 += subset_size*n; 14308122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 14408122e43SStefano Zampini } 14508122e43SStefano Zampini if (mss) { 1469ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 14708122e43SStefano Zampini PetscBLASInt B_itype = 1; 14808122e43SStefano Zampini PetscBLASInt B_N = mss; 1494c6709b3SStefano Zampini PetscReal zero = 0.0; 1504c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 15108122e43SStefano Zampini 15208122e43SStefano Zampini B_lwork = -1; 15308122e43SStefano Zampini S = NULL; 15408122e43SStefano Zampini St = NULL; 155a58a30b4SStefano Zampini eigs = NULL; 156a58a30b4SStefano Zampini eigv = NULL; 157a58a30b4SStefano Zampini B_iwork = NULL; 158a58a30b4SStefano Zampini B_ifail = NULL; 159d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 160d1710679SStefano Zampini rwork = NULL; 161d1710679SStefano Zampini #endif 1628bec7fa6SStefano Zampini thresh = 1.0; 16308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 16408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 16508122e43SStefano 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)); 16608122e43SStefano Zampini #else 16708122e43SStefano 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)); 16808122e43SStefano Zampini #endif 16908122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 17008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 17108122e43SStefano Zampini } else { 17208122e43SStefano Zampini /* TODO */ 17308122e43SStefano Zampini } 17408122e43SStefano Zampini } else { 17508122e43SStefano Zampini lwork = 0; 17608122e43SStefano Zampini } 17708122e43SStefano Zampini 17808122e43SStefano Zampini nv = 0; 179d62866d3SStefano 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) */ 180d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 18108122e43SStefano Zampini } 1824c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 183f6f667cfSStefano Zampini if (allocated_S_St) { 184f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 185f6f667cfSStefano Zampini } 186f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 18708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 18808122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 18908122e43SStefano Zampini #endif 1909162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1919162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1929162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 19308122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1949162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 19508122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 19608122e43SStefano Zampini 19708122e43SStefano Zampini maxneigs = 0; 19808122e43SStefano Zampini cum = cum2 = cumarray = 0; 1999162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 2009162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 201d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 20208122e43SStefano Zampini const PetscInt *idxs; 20308122e43SStefano Zampini 204d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 20508122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 20608122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 20708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 20808122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 2099162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 2109162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 21108122e43SStefano Zampini } 21208122e43SStefano Zampini cum2 = cum; 213d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 21408122e43SStefano Zampini } 21508122e43SStefano Zampini 21608122e43SStefano Zampini if (mss) { /* multilevel */ 21708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 21808122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 21908122e43SStefano Zampini } 22008122e43SStefano Zampini 22108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 22208122e43SStefano Zampini 22308122e43SStefano Zampini const PetscInt *idxs; 224f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 225862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 22608122e43SStefano Zampini PetscBLASInt B_N; 227aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 22808122e43SStefano Zampini 229862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 230f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 231f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 2329ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 233aff50787SStefano Zampini PetscInt j,k; 234aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 235aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 236aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 23708122e43SStefano Zampini } 23808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 239aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 240aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 241aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 242aff50787SStefano Zampini } 24308122e43SStefano Zampini } 24408122e43SStefano Zampini } else { 24508122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24608122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24708122e43SStefano Zampini } 2488bec7fa6SStefano Zampini } else { 249f6f667cfSStefano Zampini S = Sarray + cumarray; 250f6f667cfSStefano Zampini St = Starray + cumarray; 2518bec7fa6SStefano Zampini } 25208122e43SStefano Zampini 253f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 254aff50787SStefano Zampini /* see if we can save some work */ 255aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 256aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 257aff50787SStefano Zampini } 258aff50787SStefano Zampini 259aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 260aff50787SStefano Zampini B_neigs = 0; 261aff50787SStefano Zampini } else { 262aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 263f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 264f6f667cfSStefano Zampini 2659ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 26608122e43SStefano Zampini PetscBLASInt B_itype = 1; 267f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 2684c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 2699552c7c7SStefano Zampini PetscInt nmin_s; 27008122e43SStefano Zampini 271f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 272fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2738bec7fa6SStefano 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]]); 274fd14bc51SStefano Zampini } 27508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 277f6f667cfSStefano 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)); 27808122e43SStefano Zampini #else 279f6f667cfSStefano 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)); 28008122e43SStefano Zampini #endif 28108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 28208122e43SStefano Zampini if (B_ierr) { 28308122e43SStefano Zampini if (B_ierr < 0 ) { 28408122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 28508122e43SStefano Zampini } else if (B_ierr <= B_N) { 28608122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 28708122e43SStefano Zampini } else { 2889552c7c7SStefano 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); 28908122e43SStefano Zampini } 29008122e43SStefano Zampini } 29108122e43SStefano Zampini 29208122e43SStefano Zampini if (B_neigs > nmax) { 293fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 294fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 295fd14bc51SStefano Zampini } 296f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 29708122e43SStefano Zampini B_neigs = nmax; 29808122e43SStefano Zampini } 29908122e43SStefano Zampini 3009552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 3019552c7c7SStefano Zampini if (B_neigs < nmin_s) { 30208122e43SStefano Zampini PetscBLASInt B_neigs2; 30308122e43SStefano Zampini 304f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 305f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 306fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 307fd14bc51SStefano 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); 308fd14bc51SStefano Zampini } 3099ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 31008122e43SStefano Zampini PetscInt j; 31108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 31208122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 31308122e43SStefano Zampini } 31408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 31508122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 31608122e43SStefano Zampini } 31708122e43SStefano Zampini } else { 31808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 31908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 32008122e43SStefano Zampini } 32108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 32208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 323f6f667cfSStefano 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)); 32408122e43SStefano Zampini #else 325f6f667cfSStefano 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)); 32608122e43SStefano Zampini #endif 32708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 32808122e43SStefano Zampini B_neigs += B_neigs2; 32908122e43SStefano Zampini } 33008122e43SStefano Zampini if (B_ierr) { 33108122e43SStefano Zampini if (B_ierr < 0 ) { 33208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 33308122e43SStefano Zampini } else if (B_ierr <= B_N) { 33408122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 33508122e43SStefano Zampini } else { 3369552c7c7SStefano 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); 33708122e43SStefano Zampini } 33808122e43SStefano Zampini } 339fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 340ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 34108122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 34208122e43SStefano Zampini if (eigs[j] == 0.0) { 343ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 34408122e43SStefano Zampini } else { 345ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 346fd14bc51SStefano Zampini } 34708122e43SStefano Zampini } 34808122e43SStefano Zampini } 34908122e43SStefano Zampini } else { 35008122e43SStefano Zampini /* TODO */ 35108122e43SStefano Zampini } 352aff50787SStefano Zampini } 3538bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 3548bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 3559162d606SStefano Zampini if (B_neigs) { 3569162d606SStefano 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); 357fd14bc51SStefano Zampini 358fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 3599552c7c7SStefano Zampini PetscInt ii; 3609552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 361ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 3629552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 363ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 364ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 365ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 366ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 367ac47001eSStefano Zampini #else 368ac47001eSStefano 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); 369ac47001eSStefano Zampini #endif 3709552c7c7SStefano Zampini } 3719552c7c7SStefano Zampini } 372fd14bc51SStefano Zampini } 37308122e43SStefano Zampini #if 0 3749162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 37508122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 37608122e43SStefano Zampini PetscScalar norm; 37708122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3789162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3799162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 38008122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 38108122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 38208122e43SStefano Zampini } else { 38308122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 38408122e43SStefano Zampini } 3859162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 386b1b3d7a2SStefano Zampini } 387b1b3d7a2SStefano Zampini #endif 3889162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3899162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3909162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3919162d606SStefano Zampini cum++; 39208122e43SStefano Zampini } 39308122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 39408122e43SStefano Zampini /* shift for next computation */ 39508122e43SStefano Zampini cumarray += subset_size*subset_size; 39608122e43SStefano Zampini } 397fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 398fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 399fd14bc51SStefano Zampini } 40008122e43SStefano Zampini 40108122e43SStefano Zampini if (mss) { 40208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 40308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 404f6f667cfSStefano Zampini /* destroy matrices (junk) */ 405f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 406f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 40708122e43SStefano Zampini } 408f6f667cfSStefano Zampini if (allocated_S_St) { 409f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 410f6f667cfSStefano Zampini } 411f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 41208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 41308122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 41408122e43SStefano Zampini #endif 41508122e43SStefano Zampini if (pcbddc->dbg_flag) { 4161b968477SStefano Zampini PetscInt maxneigs_r; 41708122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4189b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 41908122e43SStefano Zampini } 42008122e43SStefano Zampini PetscFunctionReturn(0); 42108122e43SStefano Zampini } 422b1b3d7a2SStefano Zampini 423674ae819SStefano Zampini #undef __FUNCT__ 424c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 425c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 426c8587f34SStefano Zampini { 427c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4288629588bSStefano Zampini PetscScalar *coarse_submat_vals; 429c8587f34SStefano Zampini PetscErrorCode ierr; 430c8587f34SStefano Zampini 431c8587f34SStefano Zampini PetscFunctionBegin; 432f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 4335e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 434c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 435c8587f34SStefano Zampini 436684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 4370fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 438684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 439c8587f34SStefano Zampini 440c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 441b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 442c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 443c8587f34SStefano Zampini } 444c8587f34SStefano Zampini 4458629588bSStefano Zampini /* 4468629588bSStefano Zampini Setup local correction and local part of coarse basis. 4478629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 4488629588bSStefano Zampini */ 44947f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 4508629588bSStefano Zampini 4518629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 4528629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 4538629588bSStefano Zampini 4548629588bSStefano Zampini /* free */ 4558629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 456c8587f34SStefano Zampini PetscFunctionReturn(0); 457c8587f34SStefano Zampini } 458c8587f34SStefano Zampini 459c8587f34SStefano Zampini #undef __FUNCT__ 460674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 461674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 462674ae819SStefano Zampini { 463674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 464674ae819SStefano Zampini PetscErrorCode ierr; 465674ae819SStefano Zampini 466674ae819SStefano Zampini PetscFunctionBegin; 467674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 468674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 469674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 470674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 471785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 472674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 473f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 474f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 475785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 47663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 47763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 478674ae819SStefano Zampini PetscFunctionReturn(0); 479674ae819SStefano Zampini } 480674ae819SStefano Zampini 481674ae819SStefano Zampini #undef __FUNCT__ 482674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 483674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 484674ae819SStefano Zampini { 485674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 486674ae819SStefano Zampini PetscErrorCode ierr; 487674ae819SStefano Zampini 488674ae819SStefano Zampini PetscFunctionBegin; 489b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 490674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 491674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 492674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 493b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 494674ae819SStefano Zampini PetscFunctionReturn(0); 495674ae819SStefano Zampini } 496674ae819SStefano Zampini 497674ae819SStefano Zampini #undef __FUNCT__ 498674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 499674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 500674ae819SStefano Zampini { 501674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 50206656605SStefano Zampini PetscScalar *array; 503674ae819SStefano Zampini PetscErrorCode ierr; 504674ae819SStefano Zampini 505674ae819SStefano Zampini PetscFunctionBegin; 506674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 50758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 50806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 50906656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 51058da7f69SStefano Zampini } 511674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 512674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 51315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 51415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 515674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 516674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 517674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 51806656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 519674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 520674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5218ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 522674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 523674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 524674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 525f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 526f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 527f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 528f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 529727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5300e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 531f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 53270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5336e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 534f5fd1fbfSStefano Zampini ierr = ISDestroy(&pcbddc->zerodiag);CHKERRQ(ierr); 53581d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 536*8b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 537*8b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 53881d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 539674ae819SStefano Zampini PetscFunctionReturn(0); 540674ae819SStefano Zampini } 541674ae819SStefano Zampini 542674ae819SStefano Zampini #undef __FUNCT__ 543f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 544f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 5456bfb1811SStefano Zampini { 5466bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5476bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 5486bfb1811SStefano Zampini VecType impVecType; 549019a44ceSStefano Zampini PetscInt n_constraints,n_R,old_size,n_benign; 5506bfb1811SStefano Zampini PetscErrorCode ierr; 5516bfb1811SStefano Zampini 5526bfb1811SStefano Zampini PetscFunctionBegin; 553f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 554019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 555f4ddd8eeSStefano Zampini } 556e7b262bdSStefano Zampini /* get sizes */ 557019a44ceSStefano Zampini n_benign = (PetscInt)(!!pcbddc->zerodiag); 558019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - pcbddc->n_vertices; 559019a44ceSStefano Zampini if (n_benign && n_constraints) { 560019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot handle benign susbspace trick without change of basis"); 561019a44ceSStefano Zampini } 562b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 5636bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 564e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 565e7b262bdSStefano Zampini /* R nodes */ 566e7b262bdSStefano Zampini old_size = -1; 567e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 568e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 569e7b262bdSStefano Zampini } 570e7b262bdSStefano Zampini if (n_R != old_size) { 571e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 572e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5736bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5746bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5756bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5766bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 577e7b262bdSStefano Zampini } 578e7b262bdSStefano Zampini /* local primal dofs */ 579e7b262bdSStefano Zampini old_size = -1; 580e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 581e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 582e7b262bdSStefano Zampini } 583e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 584e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 58583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 586e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5876bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 588e7b262bdSStefano Zampini } 589e7b262bdSStefano Zampini /* local explicit constraints */ 590e7b262bdSStefano Zampini old_size = -1; 591e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 592e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 593e7b262bdSStefano Zampini } 594e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 595e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 59683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 59783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 59883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 59983b7ccabSStefano Zampini } 6006bfb1811SStefano Zampini PetscFunctionReturn(0); 6016bfb1811SStefano Zampini } 6026bfb1811SStefano Zampini 6036bfb1811SStefano Zampini #undef __FUNCT__ 60447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 60547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 60688ebb749SStefano Zampini { 60725084f0cSStefano Zampini PetscErrorCode ierr; 60825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 60988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 61088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 611d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 61225084f0cSStefano Zampini /* submatrices of local problem */ 61380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 614019a44ceSStefano Zampini /* submatrices of benign trick */ 615019a44ceSStefano Zampini Mat B0_R,B0_V; 61606656605SStefano Zampini /* submatrices of local coarse problem */ 61706656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 61825084f0cSStefano Zampini /* working matrices */ 61906656605SStefano Zampini Mat C_CR; 62025084f0cSStefano Zampini /* additional working stuff */ 62106656605SStefano Zampini PC pc_R; 62206656605SStefano Zampini Mat F; 623019a44ceSStefano Zampini Vec vec_R_benign; 62406656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 62506656605SStefano Zampini 62625084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 62706656605SStefano Zampini PetscScalar *work; 62806656605SStefano Zampini PetscInt *idx_V_B; 629019a44ceSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign; 63006656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 631b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 63245a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 63388ebb749SStefano Zampini MatType impMatType; 63425084f0cSStefano Zampini /* some shortcuts to scalars */ 63506656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 63688ebb749SStefano Zampini 63788ebb749SStefano Zampini PetscFunctionBegin; 638019a44ceSStefano Zampini n_benign = (PetscInt)(!!pcbddc->zerodiag); 639b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 640019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - n_vertices; 64188ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 642b371cd4fSStefano Zampini n_B = pcis->n_B; 643b371cd4fSStefano Zampini n_D = pcis->n - n_B; 64488ebb749SStefano Zampini n_R = pcis->n - n_vertices; 64588ebb749SStefano Zampini 64688ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 64788ebb749SStefano Zampini impMatType = MATSEQDENSE; 64888ebb749SStefano Zampini 64988ebb749SStefano Zampini /* vertices in boundary numbering */ 650785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 6510e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 65288ebb749SStefano Zampini if (i != n_vertices) { 65322d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 65488ebb749SStefano Zampini } 65588ebb749SStefano Zampini 65606656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 657019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 658019a44ceSStefano Zampini if (!n_benign) { 65906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 66006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 66106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 66206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 66306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 66406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 66506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 66606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 667019a44ceSStefano Zampini } else { /* these wrappers are not needed with the benign trick */ 668019a44ceSStefano Zampini S_VV = S_CV = S_VC = S_CC = NULL; 669019a44ceSStefano Zampini } 67006656605SStefano Zampini 67106656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 67206656605SStefano Zampini /* allocate workspace */ 67306656605SStefano Zampini n = 0; 67406656605SStefano Zampini if (n_constraints) { 67506656605SStefano Zampini n += n_R*n_constraints; 67606656605SStefano Zampini } 67706656605SStefano Zampini if (n_vertices) { 67806656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 67980677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 68006656605SStefano Zampini } 6813301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 68206656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 68306656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 68406656605SStefano Zampini } 68506656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 68606656605SStefano Zampini 68706656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 68806656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 68906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 69006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 69106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 69206656605SStefano Zampini if (isLU || isILU || isCHOL) { 69306656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 694d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 695d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 696d62866d3SStefano Zampini MatFactorType type; 697d62866d3SStefano Zampini 6986816873aSStefano Zampini F = reuse_mumps->F; 6996816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 700d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 70106656605SStefano Zampini } else { 70206656605SStefano Zampini F = NULL; 70306656605SStefano Zampini } 70406656605SStefano Zampini 70588ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 70688ebb749SStefano Zampini if (n_constraints) { 70706656605SStefano Zampini Mat M1,M2,M3; 70880677318SStefano Zampini Mat auxmat; 70906656605SStefano Zampini IS is_aux; 71080677318SStefano Zampini PetscScalar *array,*array2; 71106656605SStefano Zampini 712f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 71380677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 71488ebb749SStefano Zampini 71525084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 71625084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 7178ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 71880677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 71988ebb749SStefano Zampini 72080677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 72180677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 72206656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 72388ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 72406656605SStefano Zampini const PetscScalar *row_cmat_values; 72506656605SStefano Zampini const PetscInt *row_cmat_indices; 72606656605SStefano Zampini PetscInt size_of_constraint,j; 72788ebb749SStefano Zampini 72806656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 72906656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 73006656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 73106656605SStefano Zampini } 73206656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 73306656605SStefano Zampini } 73480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 73506656605SStefano Zampini if (F) { 73606656605SStefano Zampini Mat B; 73706656605SStefano Zampini 73806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 73980677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 74006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 74106656605SStefano Zampini } else { 74280677318SStefano Zampini PetscScalar *marr; 74380677318SStefano Zampini 74480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 74506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 74606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 74780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 74806656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 74906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 75006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 75106656605SStefano Zampini } 75280677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 75306656605SStefano Zampini } 75480677318SStefano Zampini if (!pcbddc->switch_static) { 75580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 75680677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 75780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 75880677318SStefano Zampini for (i=0;i<n_constraints;i++) { 75980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 76080677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 76180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 76280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 76380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 76480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 76580677318SStefano Zampini } 76680677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 76780677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 76880677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 76980677318SStefano Zampini } else { 77080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 77180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 77225084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 77380677318SStefano Zampini } 77480677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 77580677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 77680677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 77706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 77806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 77980677318SStefano Zampini if (isCHOL) { 78080677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 78180677318SStefano Zampini } else { 78225084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 78380677318SStefano Zampini } 78480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 78506656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 78625084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 78725084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 78825084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 78980677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 79080677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 79180677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 79206656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 79306656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 794f4ddd8eeSStefano Zampini } 79588ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 79688ebb749SStefano Zampini if (n_vertices) { 79706656605SStefano Zampini IS is_aux; 7983a50541eSStefano Zampini 7996816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 8006816873aSStefano Zampini IS tis; 8016816873aSStefano Zampini 8026816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 8036816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 8046816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 8056816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 8066816873aSStefano Zampini } else { 8073a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 8086816873aSStefano Zampini } 8099577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 8109577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 81104708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 812019a44ceSStefano Zampini if (n_benign) { 813019a44ceSStefano Zampini Mat B0; 814019a44ceSStefano Zampini IS dummy; 815019a44ceSStefano Zampini PetscInt ii[2]; 816019a44ceSStefano Zampini 817019a44ceSStefano Zampini ii[0] = 0; 818019a44ceSStefano Zampini ii[1] = pcbddc->B0_ncol; 819019a44ceSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 820019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 821019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 822019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 823019a44ceSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 824019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 825019a44ceSStefano Zampini } 82625084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 82788ebb749SStefano Zampini } 82888ebb749SStefano Zampini 82988ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 830f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 83106656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 83206656605SStefano Zampini if (pcbddc->coarse_phi_D) { 83306656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 83406656605SStefano Zampini } 835f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 83606656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 83706656605SStefano Zampini PetscScalar *marray; 83806656605SStefano Zampini 83906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 84006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 841f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 842f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 843f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 844f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 845f4ddd8eeSStefano Zampini } 846f4ddd8eeSStefano Zampini } 84706656605SStefano Zampini 848f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 84906656605SStefano Zampini PetscScalar *marray; 85088ebb749SStefano Zampini 85106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8528eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 85306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 85488ebb749SStefano Zampini } 8553301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 85606656605SStefano Zampini n *= 2; 85788ebb749SStefano Zampini } 85806656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 86006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8618eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 86306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 86488ebb749SStefano Zampini } 8653301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 86606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 8678eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 86906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 87088ebb749SStefano Zampini } 87188ebb749SStefano Zampini } else { 872c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 873c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8741b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 875c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 876c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 877c0553b1fSStefano Zampini } 87888ebb749SStefano Zampini } 87906656605SStefano Zampini } 880019a44ceSStefano Zampini 88106656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 88206656605SStefano Zampini /* vertices */ 88306656605SStefano Zampini if (n_vertices) { 88416f15bc4SStefano Zampini 88504708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 88604708bb6SStefano Zampini 88716f15bc4SStefano Zampini if (n_R) { 88806656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 889019a44ceSStefano Zampini Mat S_V_benign; 89006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 89116f15bc4SStefano Zampini PetscScalar *x,*y; 89204708bb6SStefano Zampini PetscBool isseqaij; 89306656605SStefano Zampini 89421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 89506656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 89604708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 8976816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 89806656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 89906656605SStefano Zampini } else { 90006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 90106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 90206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 90306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 90406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 90706656605SStefano Zampini } 90806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 90906656605SStefano Zampini } 910019a44ceSStefano Zampini if (n_benign) { 911019a44ceSStefano Zampini const PetscScalar *vals; 912019a44ceSStefano Zampini const PetscInt *idxs; 913019a44ceSStefano Zampini PetscInt n; 914019a44ceSStefano Zampini 915019a44ceSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign);CHKERRQ(ierr); 916019a44ceSStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 917019a44ceSStefano Zampini ierr = MatGetRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 918019a44ceSStefano Zampini ierr = VecSetValues(pcbddc->vec1_R,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 919019a44ceSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_R);CHKERRQ(ierr); 920019a44ceSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_R);CHKERRQ(ierr); 921019a44ceSStefano Zampini ierr = MatRestoreRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 922019a44ceSStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,vec_R_benign);CHKERRQ(ierr); 923019a44ceSStefano Zampini ierr = VecScale(vec_R_benign,m_one);CHKERRQ(ierr); 924019a44ceSStefano Zampini } 92580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 92606656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 92706656605SStefano Zampini if (n_constraints) { 92806656605SStefano Zampini Mat B; 92980677318SStefano Zampini 930b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 93180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 93280677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 93380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 93480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 93580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 93680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 93780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 93880677318SStefano Zampini } 93980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 94080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 94180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 94380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 94406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 94506656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 94606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 94706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94806656605SStefano Zampini } 94904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 95004708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 95104708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 95204708bb6SStefano Zampini } 95306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 954019a44ceSStefano Zampini if (n_benign) { 955019a44ceSStefano Zampini ierr = MatMatMult(B0_R,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_V_benign);CHKERRQ(ierr); 956019a44ceSStefano Zampini } 95780677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 96006656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 96106656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 96206656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 96306656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 964019a44ceSStefano Zampini if (!n_benign) { 96506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 966019a44ceSStefano Zampini } else { 967019a44ceSStefano Zampini const PetscScalar *vals; 968019a44ceSStefano Zampini PetscScalar val; 969019a44ceSStefano Zampini const PetscInt *idxs; 970019a44ceSStefano Zampini PetscInt n,j; 971019a44ceSStefano Zampini 972019a44ceSStefano Zampini /* Use explicit loops to populate the local coarse matrix */ 973019a44ceSStefano Zampini ierr = MatDenseGetArray(S_VVt,&x);CHKERRQ(ierr); 974019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 975019a44ceSStefano Zampini for (i=0;i<n_vertices;i++) { 976019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+i] = x[j*n_vertices+i]; 977019a44ceSStefano Zampini } 978019a44ceSStefano Zampini } 979019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&x);CHKERRQ(ierr); 980019a44ceSStefano Zampini ierr = MatDenseGetArray(S_V_benign,&x);CHKERRQ(ierr); 981019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 982019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 983019a44ceSStefano Zampini coarse_submat_vals[n_vertices*pcbddc->local_primal_size+j] = x[j]+vals[j]; 984019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices] = x[j]+vals[j]; 985019a44ceSStefano Zampini } 986019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 987019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_V_benign,&x);CHKERRQ(ierr); 988019a44ceSStefano Zampini ierr = MatDestroy(&S_V_benign);CHKERRQ(ierr); 989019a44ceSStefano Zampini ierr = VecDot(vec_R_benign,vec_R_benign,&val);CHKERRQ(ierr); 990019a44ceSStefano Zampini coarse_submat_vals[pcbddc->local_primal_size*pcbddc->local_primal_size-1] = -val; 991019a44ceSStefano Zampini } 99206656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 99316f15bc4SStefano Zampini } else { 994019a44ceSStefano Zampini /* TODO benign case also here */ 995019a44ceSStefano Zampini if (n_benign) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not yet implemented"); 99616f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 99716f15bc4SStefano Zampini } 99821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 99906656605SStefano Zampini /* coarse basis functions */ 100006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 100116f15bc4SStefano Zampini PetscScalar *y; 100216f15bc4SStefano Zampini 100306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 100406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 100506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 100606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 100906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 101106656605SStefano Zampini 101206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 101306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 101406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 101506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 101806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 101906656605SStefano Zampini } 102006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 102106656605SStefano Zampini } 1022019a44ceSStefano Zampini if (n_benign) { 1023019a44ceSStefano Zampini PetscScalar *y; 1024019a44ceSStefano Zampini 1025019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1026019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*n_vertices);CHKERRQ(ierr); 1027019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1028019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1029019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1030019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 1031019a44ceSStefano Zampini 1032019a44ceSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1033019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1034019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 1035019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1036019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1037019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1038019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1039019a44ceSStefano Zampini } 1040019a44ceSStefano Zampini } 104104708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 104204708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 104306656605SStefano Zampini } 104406656605SStefano Zampini 104506656605SStefano Zampini if (n_constraints) { 104606656605SStefano Zampini Mat B; 104706656605SStefano Zampini 104806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 104906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 105080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 105106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 105206656605SStefano Zampini if (n_vertices) { 105380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 105480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 105580677318SStefano Zampini } else { 105680677318SStefano Zampini Mat S_VCt; 105780677318SStefano Zampini 105880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 105980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 106080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 106180677318SStefano Zampini } 106206656605SStefano Zampini } 106306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 106406656605SStefano Zampini /* coarse basis functions */ 106506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 106606656605SStefano Zampini PetscScalar *y; 106706656605SStefano Zampini 106806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 106906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 107006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 107106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 107406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 107506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 107606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 107706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 107806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 108206656605SStefano Zampini } 108306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 108406656605SStefano Zampini } 108506656605SStefano Zampini } 108680677318SStefano Zampini if (n_constraints) { 108780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 108880677318SStefano Zampini } 108906656605SStefano Zampini 1090019a44ceSStefano Zampini if (n_benign) { 1091019a44ceSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1092019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1093019a44ceSStefano Zampini ierr = VecDestroy(&vec_R_benign);CHKERRQ(ierr); 1094019a44ceSStefano Zampini } 1095019a44ceSStefano Zampini 109606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 10973301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 109806656605SStefano Zampini 109906656605SStefano Zampini if (n_constraints) { 110016f15bc4SStefano Zampini Mat S_CCT,B_C; 110106656605SStefano Zampini 110280677318SStefano Zampini /* this is a lazy thing */ 110380677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 110406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 110506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 110606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 110716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 110806656605SStefano Zampini if (n_vertices) { 110916f15bc4SStefano Zampini Mat B_V,S_VCT; 111006656605SStefano Zampini 111106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 111206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 111306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 111406656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 111516f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 111606656605SStefano Zampini } 111706656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 111880677318SStefano Zampini } else { /* if there are no constraints, reset work */ 111980677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 112006656605SStefano Zampini } 112116f15bc4SStefano Zampini if (n_vertices && n_R) { 112206656605SStefano Zampini Mat A_VRT; 112380677318SStefano Zampini PetscScalar *marray; 112406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 112506656605SStefano Zampini 112680677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 112780677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 112880677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 112906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 113080677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 113180677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 113216f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 113306656605SStefano Zampini } 113406656605SStefano Zampini 113506656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 113606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 113706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 113806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 113906656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 114006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 114106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 114206656605SStefano Zampini } 114306656605SStefano Zampini } else { 114406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 114506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 114706656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 114906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 115006656605SStefano Zampini } 115106656605SStefano Zampini } 115206656605SStefano Zampini /* coarse basis functions */ 115306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 115406656605SStefano Zampini PetscScalar *y; 115506656605SStefano Zampini 115606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 115706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 115806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 115906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 116006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 116106656605SStefano Zampini if (i<n_vertices) { 116206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 116306656605SStefano Zampini } 116406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 116506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 116606656605SStefano Zampini 116706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 116806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 116906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 117006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 117306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 117406656605SStefano Zampini } 117506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 117606656605SStefano Zampini } 117706656605SStefano Zampini } 1178d62866d3SStefano Zampini /* free memory */ 117988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 118006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 118106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 118206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 118306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1184d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1185d62866d3SStefano Zampini if (n_vertices) { 1186d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1187d62866d3SStefano Zampini } 1188d62866d3SStefano Zampini if (n_constraints) { 1189d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1190d62866d3SStefano Zampini } 119188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 119288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 119388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1194019a44ceSStefano Zampini //if (pcbddc->dbg_flag) { 1195019a44ceSStefano Zampini if (0) { 119688ebb749SStefano Zampini Mat coarse_sub_mat; 119725084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 119888ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 119988ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 120088ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 12018bec7fa6SStefano Zampini Mat C_B,CPHI; 12028bec7fa6SStefano Zampini IS is_dummy; 12038bec7fa6SStefano Zampini Vec mones; 120488ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 120588ebb749SStefano Zampini PetscReal real_value; 120688ebb749SStefano Zampini 120788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 120888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 120988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 121088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 121188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 121288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1213c0553b1fSStefano Zampini if (unsymmetric_check) { 121488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 121588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 121688ebb749SStefano Zampini } 121788ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 121888ebb749SStefano Zampini 121925084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 12203301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 122125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1222c0553b1fSStefano Zampini if (unsymmetric_check) { 122388ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 122488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 122588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 122688ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 122788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 122888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 122988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 123188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 123488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123588ebb749SStefano Zampini } else { 123688ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 123788ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 123888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 124088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 124388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124488ebb749SStefano Zampini } 124588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 124688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 124788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 124888ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 124981d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 12508bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 12510fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 125206656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 12538bec7fa6SStefano Zampini 12548bec7fa6SStefano Zampini /* check constraints */ 12558bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 12568bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 12578bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 12588bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 12598bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 12608bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 12618bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1262bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1263c0553b1fSStefano Zampini if (unsymmetric_check) { 1264bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1265bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1266bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1267bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1268bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 126988ebb749SStefano Zampini } 12708bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 12718bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 12728bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 12738bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 12748bec7fa6SStefano Zampini 127525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 127688ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 127788ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 127888ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 127988ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 128088ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 128188ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 128288ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 128388ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 128488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 128588ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1286c0553b1fSStefano Zampini if (unsymmetric_check) { 128788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 128888ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 128988ebb749SStefano Zampini } 129088ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 129188ebb749SStefano Zampini } 12928629588bSStefano Zampini /* get back data */ 12938629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 129488ebb749SStefano Zampini PetscFunctionReturn(0); 129588ebb749SStefano Zampini } 129688ebb749SStefano Zampini 129788ebb749SStefano Zampini #undef __FUNCT__ 1298d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1299d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1300aa0d41d4SStefano Zampini { 1301d65f70fdSStefano Zampini Mat *work_mat; 1302d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1303d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1304d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1305aa0d41d4SStefano Zampini PetscErrorCode ierr; 1306aa0d41d4SStefano Zampini 1307aa0d41d4SStefano Zampini PetscFunctionBegin; 1308d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1309d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1310d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1311d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1312aa0d41d4SStefano Zampini 1313d65f70fdSStefano Zampini if (!rsorted) { 1314906d46d4SStefano Zampini const PetscInt *idxs; 1315906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1316aa0d41d4SStefano Zampini 1317d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1318d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1319d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1320d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1321aa0d41d4SStefano Zampini } 1322d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1323d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1324d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1325d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1326aa0d41d4SStefano Zampini } 1327d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1328d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1329d65f70fdSStefano Zampini } else { 1330d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1331d65f70fdSStefano Zampini isrow_s = isrow; 1332aa0d41d4SStefano Zampini } 1333906d46d4SStefano Zampini 1334d65f70fdSStefano Zampini if (!csorted) { 1335d65f70fdSStefano Zampini if (isrow == iscol) { 1336d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1337d65f70fdSStefano Zampini iscol_s = isrow_s; 1338d65f70fdSStefano Zampini } else { 1339d65f70fdSStefano Zampini const PetscInt *idxs; 1340d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1341906d46d4SStefano Zampini 1342d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1343d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1344d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1345d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1346d65f70fdSStefano Zampini } 1347d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1348d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1349d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1350d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1351d65f70fdSStefano Zampini } 1352d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1353d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1354d65f70fdSStefano Zampini } 1355d65f70fdSStefano Zampini } else { 1356d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1357d65f70fdSStefano Zampini iscol_s = iscol; 1358d65f70fdSStefano Zampini } 1359d65f70fdSStefano Zampini 1360d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1361d65f70fdSStefano Zampini 1362d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1363906d46d4SStefano Zampini Mat new_mat; 1364d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1365906d46d4SStefano Zampini 1366d65f70fdSStefano Zampini if (!rsorted) { 1367d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1368d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1369d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1370d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1371906d46d4SStefano Zampini } 1372d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1373d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1374d65f70fdSStefano Zampini } else { 1375d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1376906d46d4SStefano Zampini } 1377d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1378d65f70fdSStefano Zampini 1379d65f70fdSStefano Zampini if (!csorted) { 1380d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1381d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1382d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1383d65f70fdSStefano Zampini } else { 1384d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1385d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1386d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1387d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1388d65f70fdSStefano Zampini } 1389d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1390d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1391d65f70fdSStefano Zampini } 1392d65f70fdSStefano Zampini } else { 1393d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1394d65f70fdSStefano Zampini } 1395d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1396d65f70fdSStefano Zampini 1397d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1398d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1399d65f70fdSStefano Zampini work_mat[0] = new_mat; 1400d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1401d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1402d65f70fdSStefano Zampini } 1403d65f70fdSStefano Zampini 1404d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1405d65f70fdSStefano Zampini *B = work_mat[0]; 1406d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1407d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1408d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1409d65f70fdSStefano Zampini PetscFunctionReturn(0); 1410d65f70fdSStefano Zampini } 1411d65f70fdSStefano Zampini 1412d65f70fdSStefano Zampini #undef __FUNCT__ 14135e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 14145e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1415aa0d41d4SStefano Zampini { 1416aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 14173bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 14185e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1419d65f70fdSStefano Zampini Mat new_mat; 14205e8657edSStefano Zampini IS is_local,is_global; 1421d65f70fdSStefano Zampini PetscInt local_size; 1422d65f70fdSStefano Zampini PetscBool isseqaij; 1423aa0d41d4SStefano Zampini PetscErrorCode ierr; 1424aa0d41d4SStefano Zampini 1425aa0d41d4SStefano Zampini PetscFunctionBegin; 1426aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 14275e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 14285e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 14293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1430aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1431d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1432aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1433906d46d4SStefano Zampini 1434906d46d4SStefano Zampini /* check */ 1435906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1436906d46d4SStefano Zampini Vec x,x_change; 1437906d46d4SStefano Zampini PetscReal error; 1438906d46d4SStefano Zampini 14395e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1440906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 14415e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1442e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1443e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1444d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1445e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1446e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1447906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1448906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1449906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1450906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1451906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1452906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1453906d46d4SStefano Zampini } 1454906d46d4SStefano Zampini 145522d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 14569b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 145722d5777bSStefano Zampini if (isseqaij) { 1458d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1459aa0d41d4SStefano Zampini } else { 1460aa0d41d4SStefano Zampini Mat work_mat; 1461aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1462d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1463aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1464aa0d41d4SStefano Zampini } 14653301b35fSStefano Zampini if (matis->A->symmetric_set) { 14663301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1467e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 14683301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1469e496cd5dSStefano Zampini #endif 14703301b35fSStefano Zampini } 147145a1bb75SStefano Zampini /* 147245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1473d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 147445a1bb75SStefano Zampini */ 1475d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1476aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1477aa0d41d4SStefano Zampini } 1478aa0d41d4SStefano Zampini 1479aa0d41d4SStefano Zampini #undef __FUNCT__ 1480a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 14818ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1482a64d13efSStefano Zampini { 1483a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1484a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1485d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 148653892102SStefano Zampini PetscInt *idx_R_local=NULL; 14873a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 14883a50541eSStefano Zampini PetscInt vbs,bs; 14896816873aSStefano Zampini PetscBT bitmask=NULL; 1490a64d13efSStefano Zampini PetscErrorCode ierr; 1491a64d13efSStefano Zampini 1492a64d13efSStefano Zampini PetscFunctionBegin; 1493b23d619eSStefano Zampini /* 1494b23d619eSStefano Zampini No need to setup local scatters if 1495b23d619eSStefano Zampini - primal space is unchanged 1496b23d619eSStefano Zampini AND 1497b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1498b23d619eSStefano Zampini AND 1499b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1500b23d619eSStefano Zampini */ 1501b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1502f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1503f4ddd8eeSStefano Zampini } 1504f4ddd8eeSStefano Zampini /* destroy old objects */ 1505f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1506f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1507f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1508a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1509b371cd4fSStefano Zampini n_B = pcis->n_B; 1510b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1511b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15123a50541eSStefano Zampini 1513a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 15146816873aSStefano Zampini 151553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 15166816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1517854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1518a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1519a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 15200e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1521a64d13efSStefano Zampini } 1522a64d13efSStefano Zampini 1523a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 15244641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 15256816873aSStefano Zampini idx_R_local[n_R++] = i; 1526a64d13efSStefano Zampini } 1527a64d13efSStefano Zampini } 152853892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 15296816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15306816873aSStefano Zampini 153153892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 153253892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 15336816873aSStefano Zampini } 15343a50541eSStefano Zampini 15353a50541eSStefano Zampini /* Block code */ 15363a50541eSStefano Zampini vbs = 1; 15373a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 15383a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 15393a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 15403a50541eSStefano Zampini PetscInt *vary; 1541d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1542785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 15433a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1544d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1545d3df7717SStefano 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 */ 15460e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1547d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 15483a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 15493a50541eSStefano Zampini is_blocked = PETSC_FALSE; 15503a50541eSStefano Zampini break; 15513a50541eSStefano Zampini } 15523a50541eSStefano Zampini } 1553d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1554d3df7717SStefano Zampini } else { 1555d3df7717SStefano Zampini /* Verify directly the R set */ 1556d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1557d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1558d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1559d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1560d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1561d3df7717SStefano Zampini break; 1562d3df7717SStefano Zampini } 1563d3df7717SStefano Zampini } 1564d3df7717SStefano Zampini } 1565d3df7717SStefano Zampini } 15663a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 15673a50541eSStefano Zampini vbs = bs; 15683a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 15693a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 15703a50541eSStefano Zampini } 15713a50541eSStefano Zampini } 15723a50541eSStefano Zampini } 15733a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 157453892102SStefano Zampini if (sub_schurs->reuse_mumps) { 157553892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 157653892102SStefano Zampini 157753892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 157853892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 157953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 158053892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 158153892102SStefano Zampini } else { 15823a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 158353892102SStefano Zampini } 1584a64d13efSStefano Zampini 1585a64d13efSStefano Zampini /* print some info if requested */ 1586a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1587019a44ceSStefano Zampini PetscInt benign; 1588019a44ceSStefano Zampini 1589019a44ceSStefano Zampini benign = (PetscInt)(!!pcbddc->zerodiag); 1590a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1591a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 15920fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1593a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1594a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1595019a44ceSStefano 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); 1596a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1597a64d13efSStefano Zampini } 1598a64d13efSStefano Zampini 1599a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 16006816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16016816873aSStefano Zampini IS is_aux1,is_aux2; 16026816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 16036816873aSStefano Zampini 16043a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1605854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1606854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1607a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 16084641a718SStefano Zampini for (i=0; i<n_D; i++) { 16094641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 16104641a718SStefano Zampini } 1611a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1612a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16134641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 16144641a718SStefano Zampini aux_array1[j++] = i; 1615a64d13efSStefano Zampini } 1616a64d13efSStefano Zampini } 1617a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1618a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1619a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 16204641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 16214641a718SStefano Zampini aux_array2[j++] = i; 1622a64d13efSStefano Zampini } 1623a64d13efSStefano Zampini } 1624a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1625a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1626a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1627a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1628a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1629a64d13efSStefano Zampini 16308eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1631785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1632a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16334641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 16344641a718SStefano Zampini aux_array1[j++] = i; 1635a64d13efSStefano Zampini } 1636a64d13efSStefano Zampini } 1637a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1638a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1639a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1640a64d13efSStefano Zampini } 16414641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 16423a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1643d62866d3SStefano Zampini } else { 164453892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16456816873aSStefano Zampini IS tis; 16466816873aSStefano Zampini PetscInt schur_size; 16476816873aSStefano Zampini 164853892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 16496816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 165053892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 16516816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16526816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 16536816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 16546816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 16556816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1656d62866d3SStefano Zampini } 1657d62866d3SStefano Zampini } 1658a64d13efSStefano Zampini PetscFunctionReturn(0); 1659a64d13efSStefano Zampini } 1660a64d13efSStefano Zampini 1661304d26faSStefano Zampini 1662304d26faSStefano Zampini #undef __FUNCT__ 1663304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1664684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1665304d26faSStefano Zampini { 1666304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1667304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1668304d26faSStefano Zampini PC pc_temp; 1669304d26faSStefano Zampini Mat A_RR; 1670f4ddd8eeSStefano Zampini MatReuse reuse; 1671304d26faSStefano Zampini PetscScalar m_one = -1.0; 1672304d26faSStefano Zampini PetscReal value; 167304708bb6SStefano Zampini PetscInt n_D,n_R; 16749577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1675304d26faSStefano Zampini PetscErrorCode ierr; 1676e604994aSStefano Zampini /* prefixes stuff */ 1677312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1678e604994aSStefano Zampini size_t len; 1679304d26faSStefano Zampini 1680304d26faSStefano Zampini PetscFunctionBegin; 1681304d26faSStefano Zampini 1682e604994aSStefano Zampini /* compute prefixes */ 1683e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1684e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1685e604994aSStefano Zampini if (!pcbddc->current_level) { 1686e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1687e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1688e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1689e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1690e604994aSStefano Zampini } else { 1691e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1692312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1693e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1694e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1695312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1696312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 169734d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 169834d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1699e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1700e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1701e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1702e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1703e604994aSStefano Zampini } 1704e604994aSStefano Zampini 1705304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1706684f6988SStefano Zampini if (dirichlet) { 1707d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 17083301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 17093301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1710964fefecSStefano Zampini } 1711ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1712964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1713304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1714304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1715304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1716304d26faSStefano Zampini /* default */ 1717304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1718e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 17199577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1720304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 17219577ea80SStefano Zampini if (issbaij) { 17229577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 17239577ea80SStefano Zampini } else { 1724304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 17259577ea80SStefano Zampini } 1726304d26faSStefano Zampini /* Allow user's customization */ 1727304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1728304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1729304d26faSStefano Zampini } 1730d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1731d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1732d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1733d62866d3SStefano Zampini 1734d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1735d5574798SStefano Zampini } 1736304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1737304d26faSStefano Zampini if (!n_D) { 1738304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1739304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1740304d26faSStefano Zampini } 1741304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1742304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1743304d26faSStefano Zampini /* set ksp_D into pcis data */ 1744304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1745304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1746304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1747684f6988SStefano Zampini } 1748304d26faSStefano Zampini 1749304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1750684f6988SStefano Zampini A_RR = 0; 1751684f6988SStefano Zampini if (neumann) { 1752d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 175304708bb6SStefano Zampini PetscInt ibs,mbs; 175404708bb6SStefano Zampini PetscBool issbaij; 175504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1756f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 17578ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1758f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1759f4ddd8eeSStefano Zampini PetscInt nn_R; 176081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1761f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1762f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1763f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1764f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1765f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1766f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1767f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1768727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1769f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1770f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1771f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1772f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1773f4ddd8eeSStefano Zampini } 1774f4ddd8eeSStefano Zampini } 1775f4ddd8eeSStefano Zampini /* last check */ 1776d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1777f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1778f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1779f4ddd8eeSStefano Zampini } 1780f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1781f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1782f4ddd8eeSStefano Zampini } 1783f4ddd8eeSStefano Zampini /* extract A_RR */ 1784af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1785af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 178604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 178704708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 178804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 178904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 179004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1791af732b37SStefano Zampini } else { 179204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 17936816873aSStefano Zampini } 179404708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 179504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 179604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 179704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 179804708bb6SStefano Zampini } else { 179904708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 180004708bb6SStefano Zampini } 180104708bb6SStefano Zampini } 180204708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1803f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 18043301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 18053301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 18066816873aSStefano Zampini } 18076816873aSStefano Zampini } else { 18086816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 18096816873aSStefano Zampini 18106816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 18116816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 18126816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1813af732b37SStefano Zampini } 1814f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1815304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1816304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1817304d26faSStefano Zampini /* default */ 1818304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1819e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1820304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 18219577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 18229577ea80SStefano Zampini if (issbaij) { 18239577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 18249577ea80SStefano Zampini } else { 1825304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 18269577ea80SStefano Zampini } 1827304d26faSStefano Zampini /* Allow user's customization */ 1828304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1829304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1830304d26faSStefano Zampini } 1831d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1832304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1833304d26faSStefano Zampini if (!n_R) { 1834304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1835304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1836304d26faSStefano Zampini } 1837d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1838d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1839d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1840d62866d3SStefano Zampini 1841d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1842d62866d3SStefano Zampini } 1843304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1844304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1845684f6988SStefano Zampini } 18466816873aSStefano Zampini /* free Neumann problem's matrix */ 18476816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1848304d26faSStefano Zampini 1849304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 18500fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1851684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1852684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1853684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1854684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1855684f6988SStefano Zampini } 1856684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 18570fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 18580fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 18590fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 18600fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 18610fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1862304d26faSStefano Zampini /* need to be adapted? */ 1863b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1864b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1865b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1866304d26faSStefano Zampini /* print info */ 1867304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1868e604994aSStefano 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); 1869304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1870304d26faSStefano Zampini } 1871b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1872298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1873304d26faSStefano Zampini } 1874684f6988SStefano Zampini } 1875684f6988SStefano Zampini if (neumann) { /* Neumann */ 18766816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 18770fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 18780fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 18790fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 18800fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 18810fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1882304d26faSStefano Zampini /* need to be adapted? */ 1883b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1884b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1885304d26faSStefano Zampini /* print info */ 1886304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1887e604994aSStefano 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); 1888304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1889304d26faSStefano Zampini } 1890b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1891298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1892304d26faSStefano Zampini } 18930fccc4e9SStefano Zampini } 1894684f6988SStefano Zampini } 1895304d26faSStefano Zampini PetscFunctionReturn(0); 1896304d26faSStefano Zampini } 1897304d26faSStefano Zampini 1898304d26faSStefano Zampini #undef __FUNCT__ 1899ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 190080677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1901674ae819SStefano Zampini { 1902674ae819SStefano Zampini PetscErrorCode ierr; 1903674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1904be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1905674ae819SStefano Zampini 1906674ae819SStefano Zampini PetscFunctionBegin; 1907be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 190880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 190920c7b377SStefano Zampini } 191080677318SStefano Zampini if (!pcbddc->switch_static) { 191180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 191280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 191380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 191420c7b377SStefano Zampini } 1915be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 191680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 191780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 191820c7b377SStefano Zampini } else { 1919be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1920be83ff47SStefano Zampini 192153892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192253892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192320c7b377SStefano Zampini } 1924be83ff47SStefano Zampini } else { 192580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 193080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 193180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 193280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 193380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1934674ae819SStefano Zampini } 1935674ae819SStefano Zampini } 1936be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 193780677318SStefano Zampini if (applytranspose) { 193880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 193980677318SStefano Zampini } else { 194080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 194180677318SStefano Zampini } 194253892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1943be83ff47SStefano Zampini } else { 1944be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1945be83ff47SStefano Zampini 1946be83ff47SStefano Zampini if (applytranspose) { 194753892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1948be83ff47SStefano Zampini } else { 194953892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1950be83ff47SStefano Zampini } 195153892102SStefano Zampini #endif 1952be83ff47SStefano Zampini } 195380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 195480677318SStefano Zampini if (!pcbddc->switch_static) { 1955be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 195680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1958be83ff47SStefano Zampini } else { 1959be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1960be83ff47SStefano Zampini 196153892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 196253892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1963be83ff47SStefano Zampini } 196480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 196580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 196680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 196780677318SStefano Zampini } 196880677318SStefano Zampini } else { 196980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 197480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 197580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 197680677318SStefano Zampini } 197780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1981674ae819SStefano Zampini } 1982674ae819SStefano Zampini PetscFunctionReturn(0); 1983674ae819SStefano Zampini } 1984674ae819SStefano Zampini 1985dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1986674ae819SStefano Zampini #undef __FUNCT__ 1987674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1988dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1989674ae819SStefano Zampini { 1990674ae819SStefano Zampini PetscErrorCode ierr; 1991674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1992674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1993674ae819SStefano Zampini const PetscScalar zero = 0.0; 1994674ae819SStefano Zampini 1995674ae819SStefano Zampini PetscFunctionBegin; 1996dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1997dc359a40SStefano Zampini if (applytranspose) { 1998674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 19998eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2000dc359a40SStefano Zampini } else { 2001674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2002674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 200315aaf578SStefano Zampini } 2004efc2fbd9SStefano Zampini 2005efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 2006efc2fbd9SStefano Zampini if (pcbddc->benign_change) { 2007efc2fbd9SStefano Zampini PetscScalar *array; 2008efc2fbd9SStefano Zampini 2009efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2010efc2fbd9SStefano Zampini array[pcbddc->local_primal_size-1] += pcbddc->benign_p0; 2011efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2012efc2fbd9SStefano Zampini } 2013efc2fbd9SStefano Zampini 201412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 201512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 201612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201812edc857SStefano Zampini 20199f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 202012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 202112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 2022964fefecSStefano Zampini Vec rhs,sol; 2023964fefecSStefano Zampini 2024964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2025964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 202612edc857SStefano Zampini if (applytranspose) { 2027964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 202812edc857SStefano Zampini } else { 2029964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 203012edc857SStefano Zampini } 203112edc857SStefano Zampini } 2032674ae819SStefano Zampini 2033674ae819SStefano Zampini /* Local solution on R nodes */ 203480677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 203580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 20369f00e9b4SStefano Zampini } 2037674ae819SStefano Zampini 20389f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 20399f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 204012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2041674ae819SStefano Zampini 2042674ae819SStefano Zampini /* Sum contributions from two levels */ 2043dc359a40SStefano Zampini if (applytranspose) { 2044dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2045dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2046dc359a40SStefano Zampini } else { 2047674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 20488eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2049dc359a40SStefano Zampini } 2050efc2fbd9SStefano Zampini /* store p0 */ 2051efc2fbd9SStefano Zampini if (pcbddc->benign_change) { 2052efc2fbd9SStefano Zampini PetscScalar *array; 2053efc2fbd9SStefano Zampini 2054efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2055efc2fbd9SStefano Zampini pcbddc->benign_p0 = array[pcbddc->local_primal_size-1]; 2056efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2057efc2fbd9SStefano Zampini } 2058674ae819SStefano Zampini PetscFunctionReturn(0); 2059674ae819SStefano Zampini } 2060674ae819SStefano Zampini 2061674ae819SStefano Zampini #undef __FUNCT__ 2062674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 206312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2064674ae819SStefano Zampini { 2065674ae819SStefano Zampini PetscErrorCode ierr; 2066674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 206758da7f69SStefano Zampini PetscScalar *array; 206812edc857SStefano Zampini Vec from,to; 2069674ae819SStefano Zampini 2070674ae819SStefano Zampini PetscFunctionBegin; 207112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 207212edc857SStefano Zampini from = pcbddc->coarse_vec; 207312edc857SStefano Zampini to = pcbddc->vec1_P; 207412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 207512edc857SStefano Zampini Vec tvec; 207658da7f69SStefano Zampini 207758da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 207858da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 207912edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 208058da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 208158da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 208258da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 208312edc857SStefano Zampini } 208412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 208512edc857SStefano Zampini from = pcbddc->vec1_P; 208612edc857SStefano Zampini to = pcbddc->coarse_vec; 208712edc857SStefano Zampini } 208812edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2089674ae819SStefano Zampini PetscFunctionReturn(0); 2090674ae819SStefano Zampini } 2091674ae819SStefano Zampini 2092674ae819SStefano Zampini #undef __FUNCT__ 2093674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 209412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2095674ae819SStefano Zampini { 2096674ae819SStefano Zampini PetscErrorCode ierr; 2097674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 209858da7f69SStefano Zampini PetscScalar *array; 209912edc857SStefano Zampini Vec from,to; 2100674ae819SStefano Zampini 2101674ae819SStefano Zampini PetscFunctionBegin; 210212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 210312edc857SStefano Zampini from = pcbddc->coarse_vec; 210412edc857SStefano Zampini to = pcbddc->vec1_P; 210512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 210612edc857SStefano Zampini from = pcbddc->vec1_P; 210712edc857SStefano Zampini to = pcbddc->coarse_vec; 210812edc857SStefano Zampini } 210912edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 211012edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 211112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 211212edc857SStefano Zampini Vec tvec; 211358da7f69SStefano Zampini 211412edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 211558da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 211658da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 211758da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 211858da7f69SStefano Zampini } 211958da7f69SStefano Zampini } else { 212058da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 212158da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 212212edc857SStefano Zampini } 212312edc857SStefano Zampini } 2124674ae819SStefano Zampini PetscFunctionReturn(0); 2125674ae819SStefano Zampini } 2126674ae819SStefano Zampini 2127984c4197SStefano Zampini /* uncomment for testing purposes */ 2128984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2129674ae819SStefano Zampini #undef __FUNCT__ 2130674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2131674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2132674ae819SStefano Zampini { 2133674ae819SStefano Zampini PetscErrorCode ierr; 2134674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2135674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2136674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2137984c4197SStefano Zampini /* one and zero */ 2138984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2139984c4197SStefano Zampini /* space to store constraints and their local indices */ 21409162d606SStefano Zampini PetscScalar *constraints_data; 21419162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 21429162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 21439162d606SStefano Zampini PetscInt *constraints_n; 2144984c4197SStefano Zampini /* iterators */ 2145b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2146984c4197SStefano Zampini /* BLAS integers */ 2147e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2148e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2149c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2150727cdba6SStefano Zampini /* reuse */ 21510e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 21520e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2153984c4197SStefano Zampini /* change of basis */ 2154b3d85658SStefano Zampini PetscBool qr_needed; 21559162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2156984c4197SStefano Zampini /* auxiliary stuff */ 215764efe560SStefano Zampini PetscInt *nnz,*is_indices; 21588a0068c3SStefano Zampini PetscInt ncc; 2159984c4197SStefano Zampini /* some quantities */ 216045a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2161a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2162984c4197SStefano Zampini 2163674ae819SStefano Zampini PetscFunctionBegin; 21648e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 21658e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 21668e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2167088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2168088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 21690e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 21700e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 21710e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21720e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21730e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2174088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2175cf5a6209SStefano Zampini 2176cf5a6209SStefano Zampini /* print some info */ 2177cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2178cf5a6209SStefano Zampini IS vertices; 2179cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2180cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2181cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2182cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2183cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2184cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2185cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2186fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2187fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2188cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2189cf5a6209SStefano Zampini } 2190cf5a6209SStefano Zampini 2191cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 21929162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2193cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2194cf5a6209SStefano Zampini const Vec *nearnullvecs; 2195cf5a6209SStefano Zampini Vec *localnearnullsp; 2196cf5a6209SStefano Zampini PetscScalar *array; 2197cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2198cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2199674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2200b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2201674ae819SStefano Zampini PetscScalar *work; 2202674ae819SStefano Zampini PetscReal *singular_vals; 2203674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2204674ae819SStefano Zampini PetscReal *rwork; 2205674ae819SStefano Zampini #endif 2206674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2207674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2208674ae819SStefano Zampini #else 2209964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2210964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2211674ae819SStefano Zampini #endif 2212674ae819SStefano Zampini 2213674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2214d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2215d06fc5fdSStefano Zampini /* free unneeded index sets */ 2216d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2217d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2218674ae819SStefano Zampini } 2219d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2220d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2221d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2222d06fc5fdSStefano Zampini } 2223d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2224d06fc5fdSStefano Zampini n_ISForEdges = 0; 2225d06fc5fdSStefano Zampini } 2226d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2227d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2228d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2229d06fc5fdSStefano Zampini } 2230d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2231d06fc5fdSStefano Zampini n_ISForFaces = 0; 2232d06fc5fdSStefano Zampini } 223370022509SStefano Zampini 223470022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 223570022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 223670022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 223770022509SStefano Zampini if (pcbddc->NullSpace) { 223870022509SStefano Zampini PetscBool tbool[2],gbool[2]; 223970022509SStefano Zampini 224070022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2241b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2242d06fc5fdSStefano Zampini if (!ISForEdges) { 2243d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2244d06fc5fdSStefano Zampini } 2245b8ffe317SStefano Zampini } 2246d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2247d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2248d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2249d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2250d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 225198a51de6SStefano Zampini } 225270022509SStefano Zampini #endif 225308122e43SStefano Zampini 2254674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2255674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2256674ae819SStefano Zampini if (nearnullsp) { 2257674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2258f4ddd8eeSStefano Zampini /* remove any stored info */ 2259f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2260f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2261f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2262f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2263f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2264473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2265f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2266f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2267f4ddd8eeSStefano Zampini } 2268984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2269984c4197SStefano Zampini nnsp_size = 0; 2270674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2271674ae819SStefano Zampini } 2272984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2273984c4197SStefano Zampini max_constraints = nnsp_size; 2274984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2275984c4197SStefano Zampini 2276674ae819SStefano Zampini /* 2277674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 22789162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 22799162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 22809162d606SStefano Zampini There can be multiple constraints per connected component 2281674ae819SStefano Zampini */ 2282674ae819SStefano Zampini n_vertices = 0; 2283674ae819SStefano Zampini if (ISForVertices) { 2284674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2285674ae819SStefano Zampini } 22869162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 22879162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 22889162d606SStefano Zampini 22899162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 22909162d606SStefano Zampini total_counts *= max_constraints; 2291674ae819SStefano Zampini total_counts += n_vertices; 22924641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 22939162d606SStefano Zampini 2294674ae819SStefano Zampini total_counts = 0; 2295674ae819SStefano Zampini max_size_of_constraint = 0; 2296674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 22979162d606SStefano Zampini IS used_is; 2298674ae819SStefano Zampini if (i<n_ISForEdges) { 22999162d606SStefano Zampini used_is = ISForEdges[i]; 2300674ae819SStefano Zampini } else { 23019162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2302674ae819SStefano Zampini } 23039162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2304674ae819SStefano Zampini total_counts += j; 2305674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2306674ae819SStefano Zampini } 23079162d606SStefano 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); 23089162d606SStefano Zampini 2309984c4197SStefano Zampini /* get local part of global near null space vectors */ 2310785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2311984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2312984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2313e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2314e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2315984c4197SStefano Zampini } 2316674ae819SStefano Zampini 2317242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2318242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2319a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2320242a89d7SStefano Zampini 2321984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2322a773dcb8SStefano Zampini if (!skip_lapack) { 2323674ae819SStefano Zampini PetscScalar temp_work; 2324911cabfeSStefano Zampini 2325674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2326984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2327785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2328785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2329785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2330674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2331785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2332674ae819SStefano Zampini #endif 2333674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2334c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2335c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2336674ae819SStefano Zampini lwork = -1; 2337674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2338674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2339c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2340674ae819SStefano Zampini #else 2341c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2342674ae819SStefano Zampini #endif 2343674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2344984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2345674ae819SStefano Zampini #else /* on missing GESVD */ 2346674ae819SStefano Zampini /* SVD */ 2347674ae819SStefano Zampini PetscInt max_n,min_n; 2348674ae819SStefano Zampini max_n = max_size_of_constraint; 2349984c4197SStefano Zampini min_n = max_constraints; 2350984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2351674ae819SStefano Zampini min_n = max_size_of_constraint; 2352984c4197SStefano Zampini max_n = max_constraints; 2353674ae819SStefano Zampini } 2354785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2355674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2356785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2357674ae819SStefano Zampini #endif 2358674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2359674ae819SStefano Zampini lwork = -1; 2360e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2361e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2362b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2363674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2364674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23659162d606SStefano 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)); 2366674ae819SStefano Zampini #else 23679162d606SStefano 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)); 2368674ae819SStefano Zampini #endif 2369674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2370984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2371984c4197SStefano Zampini #endif /* on missing GESVD */ 2372674ae819SStefano Zampini /* Allocate optimal workspace */ 2373674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2374854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2375674ae819SStefano Zampini } 2376674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2377674ae819SStefano Zampini total_counts = 0; 23789162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 23799162d606SStefano Zampini constraints_data_ptr[0] = 0; 2380674ae819SStefano Zampini /* vertices */ 23819162d606SStefano Zampini if (n_vertices) { 2382674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 23839162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 23849162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2385674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 23869162d606SStefano Zampini constraints_n[total_counts] = 1; 23879162d606SStefano Zampini constraints_data[total_counts] = 1.0; 23889162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 23899162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2390674ae819SStefano Zampini total_counts++; 2391674ae819SStefano Zampini } 2392674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2393984c4197SStefano Zampini PetscBool used_vertex; 2394674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2395674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2396674ae819SStefano Zampini k = 0; 2397674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2398984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2399984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 24009162d606SStefano Zampini constraints_n[total_counts] = 1; 24019162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 24029162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24039162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24049162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2405674ae819SStefano Zampini total_counts++; 2406674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2407674ae819SStefano Zampini } 2408984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2409674ae819SStefano Zampini k++; 2410674ae819SStefano Zampini } 2411674ae819SStefano Zampini } 2412674ae819SStefano Zampini } 2413674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2414674ae819SStefano Zampini n_vertices = total_counts; 2415674ae819SStefano Zampini } 2416984c4197SStefano Zampini 2417674ae819SStefano Zampini /* edges and faces */ 24189162d606SStefano Zampini total_counts_cc = total_counts; 2419911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 24209162d606SStefano Zampini IS used_is; 24219162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 24229162d606SStefano Zampini 2423911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 24249162d606SStefano Zampini used_is = ISForEdges[ncc]; 2425984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2426674ae819SStefano Zampini } else { 24279162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2428984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2429674ae819SStefano Zampini } 2430674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 24319162d606SStefano Zampini 24329162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 24339162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2434984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2435984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2436674ae819SStefano Zampini if (nnsp_has_cnst) { 24375b08dc53SStefano Zampini PetscScalar quad_value; 24389162d606SStefano Zampini 24399162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24409162d606SStefano Zampini idxs_copied = PETSC_TRUE; 24419162d606SStefano Zampini 2442a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2443674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2444a773dcb8SStefano Zampini } else { 2445a773dcb8SStefano Zampini quad_value = 1.0; 2446a773dcb8SStefano Zampini } 2447674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24489162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2449674ae819SStefano Zampini } 24509162d606SStefano Zampini temp_constraints++; 2451674ae819SStefano Zampini total_counts++; 2452674ae819SStefano Zampini } 2453674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2454984c4197SStefano Zampini PetscReal real_value; 24559162d606SStefano Zampini PetscScalar *ptr_to_data; 24569162d606SStefano Zampini 2457984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 24589162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2459674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24609162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2461674ae819SStefano Zampini } 2462984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2463984c4197SStefano Zampini /* check if array is null on the connected component */ 2464e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24659162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 24665b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2467674ae819SStefano Zampini temp_constraints++; 2468674ae819SStefano Zampini total_counts++; 24699162d606SStefano Zampini if (!idxs_copied) { 24709162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24719162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2472674ae819SStefano Zampini } 2473674ae819SStefano Zampini } 24749162d606SStefano Zampini } 24759162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 247645a1bb75SStefano Zampini valid_constraints = temp_constraints; 2477eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2478a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 24799162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 24809162d606SStefano Zampini 24819162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2482a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24839162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2484a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 24859162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2486a773dcb8SStefano Zampini } else { /* perform SVD */ 2487984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 24889162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2489674ae819SStefano Zampini 2490674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2491984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2492984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2493984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2494984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2495984c4197SStefano Zampini from that computed using LAPACKgesvd 2496984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2497984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2498984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2499674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2500e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2501984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2502674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2503674ae819SStefano Zampini for (k=0;k<j+1;k++) { 25049162d606SStefano 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)); 2505674ae819SStefano Zampini } 2506674ae819SStefano Zampini } 2507e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2508e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2509e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2510674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2511c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2512674ae819SStefano Zampini #else 2513c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2514674ae819SStefano Zampini #endif 2515674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2516984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2517984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2518674ae819SStefano Zampini j = 0; 2519984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2520674ae819SStefano Zampini total_counts = total_counts-j; 252145a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2522e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2523c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2524c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2525c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2526c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2527c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2528c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2529674ae819SStefano Zampini if (j<temp_constraints) { 2530984c4197SStefano Zampini PetscInt ii; 2531984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2532674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25339162d606SStefano 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)); 2534674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2535984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2536674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 25379162d606SStefano 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]; 2538674ae819SStefano Zampini } 2539674ae819SStefano Zampini } 2540674ae819SStefano Zampini } 2541674ae819SStefano Zampini #else /* on missing GESVD */ 2542e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2543e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2544b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2545674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2546674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25479162d606SStefano 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)); 2548674ae819SStefano Zampini #else 25499162d606SStefano 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)); 2550674ae819SStefano Zampini #endif 2551984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2552674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2553984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2554e310c8b4SStefano Zampini k = temp_constraints; 2555e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2556674ae819SStefano Zampini j = 0; 2557e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 255845a1bb75SStefano Zampini valid_constraints = k-j; 2559911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2560984c4197SStefano Zampini #endif /* on missing GESVD */ 2561674ae819SStefano Zampini } 2562a773dcb8SStefano Zampini } 25639162d606SStefano Zampini /* update pointers information */ 25649162d606SStefano Zampini if (valid_constraints) { 25659162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 25669162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 25679162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 25689162d606SStefano Zampini /* set change_of_basis flag */ 256945a1bb75SStefano Zampini if (boolforchange) { 2570b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 25719162d606SStefano Zampini } 2572b3d85658SStefano Zampini total_counts_cc++; 257345a1bb75SStefano Zampini } 257445a1bb75SStefano Zampini } 2575984c4197SStefano Zampini /* free workspace */ 25768f1c130eSStefano Zampini if (!skip_lapack) { 2577984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2578984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2579984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2580984c4197SStefano Zampini #endif 2581984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2582984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2583984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2584984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2585984c4197SStefano Zampini #endif 2586984c4197SStefano Zampini } 2587984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2588984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2589984c4197SStefano Zampini } 2590984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2591cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2592cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2593cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2594cf5a6209SStefano Zampini } 2595cf5a6209SStefano Zampini if (n_ISForFaces) { 2596cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2597cf5a6209SStefano Zampini } 2598cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2599cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2600cf5a6209SStefano Zampini } 2601cf5a6209SStefano Zampini if (n_ISForEdges) { 2602cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2603cf5a6209SStefano Zampini } 2604cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 260508122e43SStefano Zampini } else { 260608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2607984c4197SStefano Zampini 260808122e43SStefano Zampini total_counts = 0; 260908122e43SStefano Zampini n_vertices = 0; 2610d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2611d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 261208122e43SStefano Zampini } 261308122e43SStefano Zampini max_constraints = 0; 26149162d606SStefano Zampini total_counts_cc = 0; 261508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 261608122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 26179162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 261808122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 261908122e43SStefano Zampini } 26209162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 26219162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 26229162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 26239162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 262474d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 26259162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 26269162d606SStefano Zampini total_counts_cc = 0; 26279162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 26289162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 26299162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 263008122e43SStefano Zampini } 263108122e43SStefano Zampini } 26329162d606SStefano Zampini #if 0 26339162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 26349162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26359162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 26369162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 26379162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 26389162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 26399162d606SStefano Zampini } 26409162d606SStefano Zampini printf("\n"); 26419162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 26429162d606SStefano Zampini } 26431b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 26448bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 26451b968477SStefano Zampini } 26461b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 26478bec7fa6SStefano 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]); 26481b968477SStefano Zampini } 264908122e43SStefano Zampini #endif 265008122e43SStefano Zampini 26518bec7fa6SStefano Zampini max_size_of_constraint = 0; 26529162d606SStefano 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]); 26539162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 265408122e43SStefano Zampini /* Change of basis */ 2655b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 265608122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 265708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 265808122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2659b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 266008122e43SStefano Zampini } 266108122e43SStefano Zampini } 266208122e43SStefano Zampini } 266308122e43SStefano Zampini } 2664984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2665019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2666019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 266708122e43SStefano Zampini 26689162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 26699162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 26709162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 26719162d606SStefano 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); 267208122e43SStefano Zampini } 2673674ae819SStefano Zampini 2674674ae819SStefano Zampini /* Create constraint matrix */ 2675674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 267616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2677984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2678984c4197SStefano Zampini 2679984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2680a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2681a717540cSStefano Zampini qr_needed = PETSC_FALSE; 268274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2683984c4197SStefano Zampini total_primal_vertices=0; 2684b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 26859162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26869162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2687984c4197SStefano Zampini if (size_of_constraint == 1) { 26889162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2689b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 269064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 26919162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 26929162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2693a717540cSStefano Zampini } 2694b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 269574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2696a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2697a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2698a717540cSStefano Zampini } 2699fa434743SStefano Zampini } else { 2700b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2701fa434743SStefano Zampini } 2702a717540cSStefano Zampini } 2703b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2704b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2705674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 270670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2707b3d85658SStefano Zampini 2708019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2709019a44ceSStefano 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); 27100e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 27110e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2712984c4197SStefano Zampini 2713984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 271474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2715785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2716984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 271774d5cdf7SStefano Zampini 2718984c4197SStefano Zampini j = total_primal_vertices; 271974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2720b3d85658SStefano Zampini cum = total_primal_vertices; 27219162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27224641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2723b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2724b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2725b3d85658SStefano Zampini cum++; 27269162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 272774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 272874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 272974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 273074d5cdf7SStefano Zampini } 27319162d606SStefano Zampini j += constraints_n[i]; 2732674ae819SStefano Zampini } 2733674ae819SStefano Zampini } 2734674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2735674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2736088faed8SStefano Zampini 2737674ae819SStefano Zampini /* set values in constraint matrix */ 2738984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 27390e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2740674ae819SStefano Zampini } 2741984c4197SStefano Zampini total_counts = total_primal_vertices; 27429162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27434641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 27449162d606SStefano Zampini PetscInt *cols; 27459162d606SStefano Zampini 27469162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 27479162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 27489162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27499162d606SStefano Zampini PetscInt row = total_counts+k; 27509162d606SStefano Zampini PetscScalar *vals; 27519162d606SStefano Zampini 27529162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 27539162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 27549162d606SStefano Zampini } 27559162d606SStefano Zampini total_counts += constraints_n[i]; 2756674ae819SStefano Zampini } 2757674ae819SStefano Zampini } 2758674ae819SStefano Zampini /* assembling */ 2759674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2760674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2761088faed8SStefano Zampini 2762984c4197SStefano Zampini /* 276345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2764984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2765984c4197SStefano Zampini */ 2766674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2767674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2768026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2769984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2770984c4197SStefano Zampini /* working stuff for GEQRF */ 277181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2772984c4197SStefano Zampini PetscBLASInt lqr_work; 2773984c4197SStefano Zampini /* working stuff for UNGQR */ 2774984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2775984c4197SStefano Zampini PetscBLASInt lgqr_work; 2776984c4197SStefano Zampini /* working stuff for TRTRS */ 2777984c4197SStefano Zampini PetscScalar *trs_rhs; 27783f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2779984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2780984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2781984c4197SStefano Zampini PetscScalar *start_vals; 2782984c4197SStefano Zampini /* working stuff for values insertion */ 27834641a718SStefano Zampini PetscBT is_primal; 278464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2785906d46d4SStefano Zampini /* matrix sizes */ 2786906d46d4SStefano Zampini PetscInt global_size,local_size; 2787906d46d4SStefano Zampini /* temporary change of basis */ 2788906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2789cf5a6209SStefano Zampini /* extra space for debugging */ 2790cf5a6209SStefano Zampini PetscScalar *dbg_work; 2791984c4197SStefano Zampini 2792906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2793906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 279416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2795bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2796906d46d4SStefano Zampini /* nonzeros for local mat */ 2797bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2798bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 27999162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2800a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 28019162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2802a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 28039162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2804a717540cSStefano Zampini } else { 28059162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 28069162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2807a717540cSStefano Zampini } 2808a717540cSStefano Zampini } 2809a717540cSStefano Zampini } 2810906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2811bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2812a717540cSStefano Zampini /* Set initial identity in the matrix */ 2813bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2814906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2815a717540cSStefano Zampini } 2816a717540cSStefano Zampini 2817a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2818a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2819a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2820a717540cSStefano Zampini } 2821a717540cSStefano Zampini 2822a717540cSStefano Zampini 2823a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2824a717540cSStefano Zampini /* 2825a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2826a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2827a717540cSStefano Zampini 2828a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2829a717540cSStefano Zampini 2830a6b551f4SStefano 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) 2831a6b551f4SStefano Zampini 2832a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2833a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2834a717540cSStefano Zampini | ... | 2835a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2836a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2837a717540cSStefano Zampini 2838a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2839a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2840a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2841a6b551f4SStefano Zampini 2842a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2843a717540cSStefano Zampini */ 2844a717540cSStefano Zampini if (qr_needed) { 2845984c4197SStefano Zampini /* space to store Q */ 2846854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2847984c4197SStefano Zampini /* first we issue queries for optimal work */ 28483f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28493f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 28503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2851984c4197SStefano Zampini lqr_work = -1; 28523f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2853984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2854984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2855785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2856984c4197SStefano Zampini lgqr_work = -1; 28573f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28583f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 28593f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 28603f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 28613f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 28623f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2863984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2864984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2865785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2866984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2867785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2868984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2869785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2870a717540cSStefano Zampini /* allocating workspace for check */ 2871a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2872cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2873a717540cSStefano Zampini } 2874a717540cSStefano Zampini } 2875984c4197SStefano Zampini /* array to store whether a node is primal or not */ 28764641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2877473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 28780e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 287939e2fb2aSStefano Zampini if (i != total_primal_vertices) { 288039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 28814641a718SStefano Zampini } 288239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 288339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 288439e2fb2aSStefano Zampini } 288539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2886984c4197SStefano Zampini 2887a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 28889162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 28899162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 28904641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2891984c4197SStefano Zampini /* get constraint info */ 28929162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2893984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2894984c4197SStefano Zampini 2895984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28969162d606SStefano 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); 2897674ae819SStefano Zampini } 2898984c4197SStefano Zampini 2899fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2900a717540cSStefano Zampini 2901a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2902a717540cSStefano Zampini if (pcbddc->dbg_flag) { 29039162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2904a717540cSStefano Zampini } 2905984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 29069162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2907984c4197SStefano Zampini 2908984c4197SStefano Zampini /* compute QR decomposition of constraints */ 29093f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29103f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2912674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29133f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2914984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2915674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2916984c4197SStefano Zampini 2917984c4197SStefano Zampini /* explictly compute R^-T */ 2918984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2919984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 29203f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29213f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 29223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2924984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29253f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2926984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2927984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2928984c4197SStefano Zampini 2929a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 29303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29323f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29333f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2934984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29353f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2936984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2937984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2938984c4197SStefano Zampini 2939984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2940984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2941984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 29423f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29443f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 29473f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2948984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29499162d606SStefano 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)); 2950984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 29519162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2952984c4197SStefano Zampini 2953984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 29549162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2955984c4197SStefano Zampini /* insert cols for primal dofs */ 2956984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2957984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 29589162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2959906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2960984c4197SStefano Zampini } 2961984c4197SStefano Zampini /* insert cols for dual dofs */ 2962984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 29639162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2964984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 29659162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2966906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2967984c4197SStefano Zampini j++; 2968674ae819SStefano Zampini } 2969674ae819SStefano Zampini } 2970984c4197SStefano Zampini 2971984c4197SStefano Zampini /* check change of basis */ 2972984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2973984c4197SStefano Zampini PetscInt ii,jj; 2974984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2975c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2976c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2977c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2978c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2979c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2980c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2981984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2982cf5a6209SStefano 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)); 2983984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2984984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2985984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2986cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2987cf5a6209SStefano 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; 2988674ae819SStefano Zampini } 2989674ae819SStefano Zampini } 2990984c4197SStefano Zampini if (!valid_qr) { 299122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2992984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2993984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2994cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2995cf5a6209SStefano 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])); 2996674ae819SStefano Zampini } 2997cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2998cf5a6209SStefano 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])); 2999984c4197SStefano Zampini } 3000984c4197SStefano Zampini } 3001984c4197SStefano Zampini } 3002674ae819SStefano Zampini } else { 300322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3004674ae819SStefano Zampini } 3005674ae819SStefano Zampini } 3006a717540cSStefano Zampini } else { /* simple transformation block */ 3007a717540cSStefano Zampini PetscInt row,col; 3008a6b551f4SStefano Zampini PetscScalar val,norm; 3009a6b551f4SStefano Zampini 3010a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30119162d606SStefano 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)); 3012a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 30139162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 30149162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3015bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 30169162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3017906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 30189162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3019a717540cSStefano Zampini } else { 3020a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 30219162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3022a717540cSStefano Zampini if (row != col) { 30239162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3024a717540cSStefano Zampini } else { 30259162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3026a717540cSStefano Zampini } 3027906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3028a717540cSStefano Zampini } 3029a717540cSStefano Zampini } 3030a717540cSStefano Zampini } 303198a51de6SStefano Zampini if (pcbddc->dbg_flag) { 303222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3033a717540cSStefano Zampini } 3034674ae819SStefano Zampini } 3035984c4197SStefano Zampini } else { 3036984c4197SStefano Zampini if (pcbddc->dbg_flag) { 30379162d606SStefano 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); 3038674ae819SStefano Zampini } 3039674ae819SStefano Zampini } 3040674ae819SStefano Zampini } 3041a717540cSStefano Zampini 3042a717540cSStefano Zampini /* free workspace */ 3043a717540cSStefano Zampini if (qr_needed) { 3044984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3045cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3046984c4197SStefano Zampini } 3047984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3048984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3049984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3050984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3051984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3052674ae819SStefano Zampini } 3053a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3054906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3055906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3056906d46d4SStefano Zampini 3057906d46d4SStefano Zampini /* assembling of global change of variable */ 3058bbb9e6c6SStefano Zampini { 3059bbb9e6c6SStefano Zampini Mat tmat; 306016f15bc4SStefano Zampini PetscInt bs; 306116f15bc4SStefano Zampini 3062906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3063906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3064bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3065bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3066bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3067bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 306816f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 306916f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3070906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3071bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3072bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3073bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3074bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3075bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3076e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3077e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3078bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3079bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3080906d46d4SStefano Zampini } 3081906d46d4SStefano Zampini /* check */ 3082906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3083906d46d4SStefano Zampini PetscReal error; 3084906d46d4SStefano Zampini Vec x,x_change; 3085906d46d4SStefano Zampini 3086906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3087906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3088906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3089906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3090e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3091e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3092bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3093e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3094e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3095906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3096906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3097906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3098906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3099bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3100906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3101906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3102906d46d4SStefano Zampini } 3103b96c3477SStefano Zampini 3104b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3105b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3106b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3107b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3108ac632422SStefano Zampini Mat S_new,tmat; 3109bbb9e6c6SStefano Zampini IS is_all_N; 3110bbb9e6c6SStefano Zampini 3111bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 31126816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3113bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3114ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3115b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3116ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3117ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3118ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3119ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3120ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3121b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3122ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3123ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3124ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3125ac632422SStefano Zampini } 3126b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3127b96c3477SStefano Zampini } 3128b96c3477SStefano Zampini } 3129906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3130906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3131b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3132b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3133b9b85e73SStefano Zampini } 3134906d46d4SStefano Zampini 3135906d46d4SStefano Zampini /* set up change of basis context */ 3136906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3137906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3138906d46d4SStefano Zampini 3139906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3140906d46d4SStefano Zampini PetscInt global_size,local_size; 3141906d46d4SStefano Zampini 3142906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3143906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3144906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3145906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3146906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3147906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3148906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3149906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3150906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3151906d46d4SStefano Zampini } else { 3152906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3153906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3154906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3155906d46d4SStefano Zampini } 3156906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3157906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3158906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3159906d46d4SStefano Zampini } else { 3160906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3161906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3162906d46d4SStefano Zampini } 3163906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3164906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3165906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3166906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3167b9b85e73SStefano Zampini } 3168a717540cSStefano Zampini 3169019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3170019a44ceSStefano Zampini if (pcbddc->zerodiag) { 3171019a44ceSStefano Zampini const PetscInt *idxs; 3172019a44ceSStefano Zampini PetscInt nz; 3173019a44ceSStefano Zampini 3174019a44ceSStefano Zampini ierr = ISGetLocalSize(pcbddc->zerodiag,&nz);CHKERRQ(ierr); 3175019a44ceSStefano Zampini ierr = ISGetIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3176019a44ceSStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3177019a44ceSStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3178019a44ceSStefano Zampini ierr = ISRestoreIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3179019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3180019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3181019a44ceSStefano Zampini pcbddc->local_primal_size++; 3182019a44ceSStefano Zampini } 3183019a44ceSStefano Zampini 3184019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3185727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3186727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3187aff50787SStefano 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); 3188c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 31890e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3190aff50787SStefano 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); 3191727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3192727cdba6SStefano Zampini } 31930e6343abSStefano Zampini } 31940e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3195727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3196727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3197727cdba6SStefano Zampini 3198a717540cSStefano Zampini /* flush dbg viewer */ 3199b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3200b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3201b8ffe317SStefano Zampini } 3202a717540cSStefano Zampini 3203e310c8b4SStefano Zampini /* free workspace */ 3204a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 32054641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 320608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 32079162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 32089162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 320908122e43SStefano Zampini } else { 32109162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 32119162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 32129162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 321308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 321408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 32159162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 32169162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 321708122e43SStefano Zampini } 3218674ae819SStefano Zampini PetscFunctionReturn(0); 3219674ae819SStefano Zampini } 3220674ae819SStefano Zampini 3221674ae819SStefano Zampini #undef __FUNCT__ 3222674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3223674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3224674ae819SStefano Zampini { 3225674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3226674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3227674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 32287fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3229674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3230674ae819SStefano Zampini 3231674ae819SStefano Zampini PetscFunctionBegin; 32328e61c736SStefano Zampini /* Reset previously computed graph */ 32338e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3234674ae819SStefano Zampini /* Init local Graph struct */ 32357fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 32363bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3237674ae819SStefano Zampini 3238575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3239575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3240575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3241575ad6abSStefano Zampini } 32429577ea80SStefano Zampini 3243674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 32444d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 32454d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 32464d379d7bSStefano Zampini PetscInt nvtxs; 3247e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3248674ae819SStefano Zampini 32494d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 32502fffb893SStefano Zampini 32512fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32522fffb893SStefano Zampini if (flg_row) { 32534d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3254b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 32552fffb893SStefano Zampini } 32562fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32579b28b941SStefano 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 */ 32584d379d7bSStefano Zampini IS is_dummy; 32594d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 32604d379d7bSStefano Zampini PetscInt j,sum; 32614d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 32624d379d7bSStefano Zampini const PetscInt *idxs; 32634d379d7bSStefano Zampini PCBDDCGraph graph; 32644d379d7bSStefano Zampini PetscBT is_on_boundary; 32654d379d7bSStefano Zampini 32664d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 32674d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 32684d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 32694d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 32707fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 32714d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3272e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3273e496cd5dSStefano Zampini if (flg_row) { 32744d379d7bSStefano Zampini graph->xadj = xadj; 32754d379d7bSStefano Zampini graph->adjncy = adjncy; 3276e496cd5dSStefano Zampini } 32774d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 32784d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3279e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32804d379d7bSStefano Zampini 32814d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 32829b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 32834d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32844d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 32854d379d7bSStefano Zampini } 32864d379d7bSStefano Zampini } 32874d379d7bSStefano Zampini 3288e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 32894d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 32904d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 32914d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 32924d379d7bSStefano Zampini } 32934d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 32944d379d7bSStefano Zampini 3295e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 32964d379d7bSStefano Zampini sum = 0; 32974d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32984d379d7bSStefano Zampini PetscInt sizecc = 0; 32994d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33004d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33014d379d7bSStefano Zampini sizecc++; 33024d379d7bSStefano Zampini } 33034d379d7bSStefano Zampini } 33044d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33054d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33064d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 33074d379d7bSStefano Zampini } 33084d379d7bSStefano Zampini } 33094d379d7bSStefano Zampini sum += sizecc*sizecc; 33104d379d7bSStefano Zampini } 33114d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 33124d379d7bSStefano Zampini sum = 0; 3313e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 33144d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 33154d379d7bSStefano Zampini cxadj[i] = sum; 33164d379d7bSStefano Zampini sum += temp; 33174d379d7bSStefano Zampini } 3318e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 33194d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33204d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33214d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33224d379d7bSStefano Zampini PetscInt k,sizecc = 0; 33234d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 33244d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 33254d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 33264d379d7bSStefano Zampini sizecc++; 33274d379d7bSStefano Zampini } 33284d379d7bSStefano Zampini } 33294d379d7bSStefano Zampini } 33304d379d7bSStefano Zampini } 33314d379d7bSStefano Zampini } 33329b28b941SStefano Zampini if (sum) { 3333e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 33344d379d7bSStefano Zampini } else { 33354d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 33364d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 33374d379d7bSStefano Zampini } 33384d379d7bSStefano Zampini graph->xadj = 0; 33394d379d7bSStefano Zampini graph->adjncy = 0; 33404d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 33414d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 33424d379d7bSStefano Zampini } 3343674ae819SStefano Zampini } 33449b28b941SStefano Zampini if (pcbddc->dbg_flag) { 33459b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3346674ae819SStefano Zampini } 3347674ae819SStefano Zampini 334863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3349674ae819SStefano Zampini vertex_size = 1; 335063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 335163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 335295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 335363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3354e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 335563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3356674ae819SStefano Zampini } 335763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 335863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 335963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3360674ae819SStefano Zampini } 336163602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3362674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 336363602bcaSStefano Zampini } else { 336463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 336563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3366854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 336763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 336863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 336963602bcaSStefano Zampini } 337063602bcaSStefano Zampini } 3371674ae819SStefano Zampini } 3372674ae819SStefano Zampini 3373674ae819SStefano Zampini /* Setup of Graph */ 3374785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3375e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3376785d1243SStefano Zampini } 3377785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3378e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3379785d1243SStefano Zampini } 3380302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3381674ae819SStefano Zampini 3382674ae819SStefano Zampini /* Graph's connected components analysis */ 3383674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3384674ae819SStefano Zampini 3385674ae819SStefano Zampini /* print some info to stdout */ 3386674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3387302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3388674ae819SStefano Zampini } 3389fb180af4SStefano Zampini 3390fb180af4SStefano Zampini /* mark topography has done */ 3391fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3392674ae819SStefano Zampini PetscFunctionReturn(0); 3393674ae819SStefano Zampini } 3394674ae819SStefano Zampini 3395dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3396674ae819SStefano Zampini #undef __FUNCT__ 3397674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3398dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3399674ae819SStefano Zampini { 3400dc456d91SStefano Zampini PetscSF sf; 3401dc456d91SStefano Zampini PetscLayout map; 3402dc456d91SStefano Zampini const PetscInt *idxs; 3403dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3404dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3405dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3406dc456d91SStefano Zampini PetscMPIInt commsize; 3407674ae819SStefano Zampini PetscBool first_found; 3408674ae819SStefano Zampini PetscErrorCode ierr; 3409674ae819SStefano Zampini 3410674ae819SStefano Zampini PetscFunctionBegin; 3411dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3412dc456d91SStefano Zampini if (subset_mult) { 3413dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3414dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3415dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3416674ae819SStefano Zampini } 3417dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3418dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3419dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3420dc456d91SStefano Zampini for (i=0;i<n;i++) { 3421dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3422dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3423674ae819SStefano Zampini } 3424dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3425dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3426dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3427dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3428dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3429dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3430dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3431dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3432dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3433dc456d91SStefano Zampini 3434dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3435dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3436dc456d91SStefano Zampini if (subset_mult) { 3437dc456d91SStefano Zampini const PetscInt* idxs_mult; 3438dc456d91SStefano Zampini 3439dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3440dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3441dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3442674ae819SStefano Zampini } else { 3443dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3444674ae819SStefano Zampini } 3445dc456d91SStefano Zampini /* local size of new subset */ 3446dc456d91SStefano Zampini n_n = 0; 3447dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3448dc456d91SStefano Zampini 3449dc456d91SStefano Zampini /* global indexes in layout */ 3450dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3451dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3452dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3453dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3454dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3455dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3456dc456d91SStefano Zampini 3457dc456d91SStefano Zampini /* reduce from leaves to roots */ 3458dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 345964a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 346064a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3461dc456d91SStefano Zampini 3462dc456d91SStefano Zampini /* count indexes in local part of layout */ 3463674ae819SStefano Zampini nlocals = 0; 3464674ae819SStefano Zampini first_index = -1; 3465674ae819SStefano Zampini first_found = PETSC_FALSE; 3466dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3467dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3468674ae819SStefano Zampini first_found = PETSC_TRUE; 3469674ae819SStefano Zampini first_index = i; 3470674ae819SStefano Zampini } 3471dc456d91SStefano Zampini nlocals += root_data[i]; 3472674ae819SStefano Zampini } 3473dc456d91SStefano Zampini 3474dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 34755fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3476dc456d91SStefano Zampini start = 0; 347764a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 34785fa240b1SStefano Zampini #else 347964a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 34805fa240b1SStefano Zampini start = start-nlocals; 34815fa240b1SStefano Zampini #endif 34825fa240b1SStefano Zampini 3483dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3484dc456d91SStefano Zampini *N_n = start + nlocals; 3485dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3486dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3487674ae819SStefano Zampini } 34885fa240b1SStefano Zampini 34895fa240b1SStefano Zampini /* adapt root data with cumulative */ 3490674ae819SStefano Zampini if (first_found) { 3491dc456d91SStefano Zampini PetscInt old_index; 3492dc456d91SStefano Zampini 3493dc456d91SStefano Zampini root_data[first_index] += start; 3494674ae819SStefano Zampini old_index = first_index; 3495dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3496dc456d91SStefano Zampini if (root_data[i]) { 3497dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3498674ae819SStefano Zampini old_index = i; 3499674ae819SStefano Zampini } 3500674ae819SStefano Zampini } 3501674ae819SStefano Zampini } 3502dc456d91SStefano Zampini 3503dc456d91SStefano Zampini /* from roots to leaves */ 3504dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3505dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3506dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3507dc456d91SStefano Zampini 3508dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3509dc456d91SStefano Zampini if (subset_mult) { 3510dc456d91SStefano Zampini const PetscInt* idxs_mult; 3511dc456d91SStefano Zampini PetscInt cum; 3512dc456d91SStefano Zampini 3513dc456d91SStefano Zampini cum = 0; 3514dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3515dc456d91SStefano Zampini for (i=0;i<n;i++) { 3516dc456d91SStefano Zampini PetscInt j; 3517dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3518674ae819SStefano Zampini } 3519dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3520674ae819SStefano Zampini } else { 3521dc456d91SStefano Zampini for (i=0;i<n;i++) { 3522dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3523674ae819SStefano Zampini } 3524674ae819SStefano Zampini } 3525dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3526dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3527674ae819SStefano Zampini PetscFunctionReturn(0); 3528674ae819SStefano Zampini } 35299a7d3425SStefano Zampini 35309a7d3425SStefano Zampini #undef __FUNCT__ 35319a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 35329a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 35339a7d3425SStefano Zampini { 35349a7d3425SStefano Zampini PetscInt i,j; 35359a7d3425SStefano Zampini PetscScalar *alphas; 35369a7d3425SStefano Zampini PetscErrorCode ierr; 35379a7d3425SStefano Zampini 35389a7d3425SStefano Zampini PetscFunctionBegin; 35399a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3540785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 35419a7d3425SStefano Zampini for (i=0;i<n;i++) { 35429a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 35439a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 35449a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 35459a7d3425SStefano Zampini } 35469a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 35479a7d3425SStefano Zampini PetscFunctionReturn(0); 35489a7d3425SStefano Zampini } 35499a7d3425SStefano Zampini 3550e7931f94SStefano Zampini #undef __FUNCT__ 355170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3552b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3553e7931f94SStefano Zampini { 355452e5ac9dSStefano Zampini IS ranks_send_to; 3555e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3556e7931f94SStefano Zampini PetscMPIInt size,rank,color; 355752e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 355852e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 35593837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 35602b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3561e7931f94SStefano Zampini PetscSubcomm subcomm; 356252e5ac9dSStefano Zampini PetscErrorCode ierr; 3563a57a6d2fSStefano Zampini 3564e7931f94SStefano Zampini PetscFunctionBegin; 35652b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 35662b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 35672b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3568e7931f94SStefano Zampini 3569e7931f94SStefano Zampini /* Get info on mapping */ 35703bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 35713bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3572e7931f94SStefano Zampini 3573e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3574785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3575e7931f94SStefano Zampini xadj[0] = 0; 3576e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3577785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3578785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3579e7931f94SStefano Zampini 35802b510759SStefano Zampini if (threshold) { 3581d023bfaeSStefano Zampini PetscInt xadj_count = 0; 35822b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3583d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3584d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3585d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3586d023bfaeSStefano Zampini xadj_count++; 3587e7931f94SStefano Zampini } 3588e7931f94SStefano Zampini } 3589d023bfaeSStefano Zampini xadj[1] = xadj_count; 3590c8587f34SStefano Zampini } else { 3591e7931f94SStefano Zampini if (xadj[1]) { 3592e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3593e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3594c8587f34SStefano Zampini } 3595e7931f94SStefano Zampini } 35963bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3597e7931f94SStefano Zampini if (use_square) { 3598e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3599e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3600e7931f94SStefano Zampini } 3601e7931f94SStefano Zampini } 3602e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3603e7931f94SStefano Zampini 36043837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3605e7931f94SStefano Zampini 3606e7931f94SStefano Zampini /* 3607e7931f94SStefano Zampini Restrict work on active processes only. 3608e7931f94SStefano Zampini */ 3609e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3610e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3611e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 36122b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3613d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3614e7931f94SStefano Zampini if (color) { 3615e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3616e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3617e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3618c8587f34SStefano Zampini } else { 361952e5ac9dSStefano Zampini Mat subdomain_adj; 362052e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 362152e5ac9dSStefano Zampini MatPartitioning partitioner; 362252e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 362352e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3624b0c7d250SStefano Zampini PetscBool aggregate; 3625b0c7d250SStefano Zampini 3626306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3627785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3628e7931f94SStefano Zampini prank = rank; 3629306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 36308002ef2cSStefano Zampini /* 3631e7931f94SStefano Zampini for (i=0;i<size;i++) { 3632e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3633c8587f34SStefano Zampini } 36348002ef2cSStefano Zampini */ 3635e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3636e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3637c8587f34SStefano Zampini } 3638e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3639b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3640b0c7d250SStefano Zampini if (aggregate) { 3641b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3642b0c7d250SStefano Zampini PetscMPIInt nrank; 3643b0c7d250SStefano Zampini PetscScalar *vals; 3644b0c7d250SStefano Zampini 3645b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3646b0c7d250SStefano Zampini lrows = 0; 3647b0c7d250SStefano Zampini if (nrank<redprocs) { 3648b0c7d250SStefano Zampini lrows = size/redprocs; 3649b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3650b0c7d250SStefano Zampini } 36515fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3652b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3653b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3654b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3655b0c7d250SStefano Zampini row = nrank; 3656b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3657b0c7d250SStefano Zampini cols = adjncy; 3658b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3659b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3660b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3661b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3662b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 366352e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 366452e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 366552e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3666b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3667b0c7d250SStefano Zampini } else { 3668306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3669b0c7d250SStefano Zampini } 367022b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3671e7931f94SStefano Zampini 3672e7931f94SStefano Zampini /* Partition */ 3673306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3674e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3675e7931f94SStefano Zampini if (use_vwgt) { 36763837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3677e7931f94SStefano Zampini v_wgt[0] = local_size; 3678e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3679c8587f34SStefano Zampini } 368028143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 368128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3682e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3683e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 368422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3685e7931f94SStefano Zampini 368652e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 368752e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 368852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 368952e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3690b0c7d250SStefano Zampini if (!redprocs) { 3691b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 369228143c3dSStefano Zampini } else { 3693b0c7d250SStefano Zampini PetscInt idxs[1]; 3694b0c7d250SStefano Zampini PetscMPIInt tag; 3695b0c7d250SStefano Zampini MPI_Request *reqs; 3696b0c7d250SStefano Zampini 3697b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3698b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3699b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3700b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 370128143c3dSStefano Zampini } 3702b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3703b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3704b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3705b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3706e7931f94SStefano Zampini } 370752e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3708e7931f94SStefano Zampini /* clean up */ 3709e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 371052e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3711e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3712e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3713e7931f94SStefano Zampini } 3714e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3715e7931f94SStefano Zampini 3716e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3717e7931f94SStefano Zampini i = 1; 3718e7931f94SStefano Zampini if (color) i=0; 3719e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3720e7931f94SStefano Zampini /* get back IS */ 3721e7931f94SStefano Zampini *is_sends = ranks_send_to; 3722e7931f94SStefano Zampini PetscFunctionReturn(0); 3723e7931f94SStefano Zampini } 3724e7931f94SStefano Zampini 3725e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3726e7931f94SStefano Zampini 3727e7931f94SStefano Zampini #undef __FUNCT__ 3728e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 372953a05cb3SStefano 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[]) 3730e7931f94SStefano Zampini { 373170cf5478SStefano Zampini Mat local_mat; 3732e7931f94SStefano Zampini IS is_sends_internal; 37339d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 373428143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 37359d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3736e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3737e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3738e7931f94SStefano Zampini const PetscInt* is_indices; 3739e7931f94SStefano Zampini MatType new_local_type; 3740e7931f94SStefano Zampini /* buffers */ 3741e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 374228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 37439d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3744e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3745e7931f94SStefano Zampini /* MPI */ 374628143c3dSStefano Zampini MPI_Comm comm,comm_n; 374728143c3dSStefano Zampini PetscSubcomm subcomm; 3748e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 374928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 375028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 375128143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 375228143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 375328143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3754e7931f94SStefano Zampini PetscErrorCode ierr; 3755e7931f94SStefano Zampini 3756e7931f94SStefano Zampini PetscFunctionBegin; 375728143c3dSStefano Zampini /* TODO: add missing checks */ 375828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 375928143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 376028143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 376128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3762e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 376328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3764e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3765e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3766e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3767e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3768e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 376928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 377070cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 377170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 377228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 377370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 377470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 377570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 377670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 377770cf5478SStefano Zampini } 3778e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3779e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3780e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3781e7931f94SStefano Zampini if (!is_sends) { 378228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3783b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3784c8587f34SStefano Zampini } else { 3785e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3786e7931f94SStefano Zampini is_sends_internal = is_sends; 3787c8587f34SStefano Zampini } 3788e7931f94SStefano Zampini 3789e7931f94SStefano Zampini /* get comm */ 3790a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3791e7931f94SStefano Zampini 3792e7931f94SStefano Zampini /* compute number of sends */ 3793e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3794e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3795e7931f94SStefano Zampini 3796e7931f94SStefano Zampini /* compute number of receives */ 3797e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3798785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3799e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3800e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3801e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3802e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3803e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3804e7931f94SStefano Zampini 380528143c3dSStefano Zampini /* restrict comm if requested */ 380628143c3dSStefano Zampini subcomm = 0; 380728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 380828143c3dSStefano Zampini if (restrict_comm) { 3809779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3810779c1cceSStefano Zampini 381128143c3dSStefano Zampini color = 0; 381253a05cb3SStefano Zampini if (restrict_full) { 381353a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 381453a05cb3SStefano Zampini } else { 381553a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 381653a05cb3SStefano Zampini } 381728143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 381828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 381928143c3dSStefano Zampini /* check if reuse has been requested */ 382028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 382128143c3dSStefano Zampini if (*mat_n) { 382228143c3dSStefano Zampini PetscMPIInt subcommsize2; 382328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 382428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 382528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 382628143c3dSStefano Zampini } else { 382728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 382828143c3dSStefano Zampini } 382928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3830779c1cceSStefano Zampini PetscMPIInt rank; 3831779c1cceSStefano Zampini 3832779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 383328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 383428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 383528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3836306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 383728143c3dSStefano Zampini } 383828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 383928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 384028143c3dSStefano Zampini } else { 384128143c3dSStefano Zampini comm_n = comm; 384228143c3dSStefano Zampini } 384328143c3dSStefano Zampini 3844e7931f94SStefano Zampini /* prepare send/receive buffers */ 3845785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3846e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3847785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3848e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 384928143c3dSStefano Zampini if (nis) { 3850854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 385128143c3dSStefano Zampini } 3852e7931f94SStefano Zampini 385328143c3dSStefano Zampini /* Get data from local matrices */ 3854e7931f94SStefano Zampini if (!isdense) { 3855a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3856e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3857e7931f94SStefano Zampini /* 3858e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3859e7931f94SStefano Zampini send_buffer_idxs should contain: 3860e7931f94SStefano Zampini - MatType_PRIVATE type 3861e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3862e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3863e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3864e7931f94SStefano Zampini */ 3865e7931f94SStefano Zampini } else { 3866e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 38673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3868854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3869e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3870e7931f94SStefano Zampini send_buffer_idxs[1] = i; 38713bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3872e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 38733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3874e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3875e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3876e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3877e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3878c8587f34SStefano Zampini } 3879c8587f34SStefano Zampini } 3880e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 388128143c3dSStefano Zampini /* additional is (if any) */ 388228143c3dSStefano Zampini if (nis) { 388328143c3dSStefano Zampini PetscMPIInt psum; 388428143c3dSStefano Zampini PetscInt j; 388528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 388628143c3dSStefano Zampini PetscInt plen; 388728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 388828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 388928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 389028143c3dSStefano Zampini } 3891854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 389228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 389328143c3dSStefano Zampini PetscInt plen; 389428143c3dSStefano Zampini const PetscInt *is_array_idxs; 389528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 389628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 389728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 389828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 389928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 390028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 390128143c3dSStefano Zampini } 390228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 390328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 390428143c3dSStefano Zampini } 390528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 390628143c3dSStefano Zampini } 390728143c3dSStefano Zampini 3908e7931f94SStefano Zampini buf_size_idxs = 0; 3909e7931f94SStefano Zampini buf_size_vals = 0; 391028143c3dSStefano Zampini buf_size_idxs_is = 0; 3911e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3912e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3913e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 391428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3915e7931f94SStefano Zampini } 3916785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3917785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 391895ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3919e7931f94SStefano Zampini 3920e7931f94SStefano Zampini /* get new tags for clean communications */ 3921e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3922e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 392328143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3924e7931f94SStefano Zampini 3925e7931f94SStefano Zampini /* allocate for requests */ 3926785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3927785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 392895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3929785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3930785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 393195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3932e7931f94SStefano Zampini 3933e7931f94SStefano Zampini /* communications */ 3934e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3935e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 393628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3937e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3938e7931f94SStefano Zampini source_dest = onodes[i]; 3939e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3940e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3941e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3942e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 394328143c3dSStefano Zampini if (nis) { 394428143c3dSStefano 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); 394528143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 394628143c3dSStefano Zampini } 3947e7931f94SStefano Zampini } 3948e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3949e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3950e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3951e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 395228143c3dSStefano Zampini if (nis) { 395328143c3dSStefano 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); 395428143c3dSStefano Zampini } 3955e7931f94SStefano Zampini } 3956e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3957e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3958e7931f94SStefano Zampini 3959e7931f94SStefano Zampini /* assemble new l2g map */ 3960e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3961e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39629d30be91SStefano Zampini new_local_rows = 0; 3963e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 39649d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3965e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3966e7931f94SStefano Zampini } 39679d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3968e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39699d30be91SStefano Zampini new_local_rows = 0; 3970e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 39719d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 39729d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3973e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3974e7931f94SStefano Zampini } 39759d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 39769d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3977e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3978e7931f94SStefano Zampini 3979e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3980e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3981e7931f94SStefano 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) */ 3982e7931f94SStefano Zampini if (n_recvs) { 398328143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3984e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3985e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3986e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3987e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3988e7931f94SStefano Zampini break; 3989e7931f94SStefano Zampini } 3990e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3991e7931f94SStefano Zampini } 3992e7931f94SStefano Zampini switch (new_local_type_private) { 399328143c3dSStefano Zampini case MATDENSE_PRIVATE: 399428143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3995e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3996e7931f94SStefano Zampini bs = 1; 399728143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 399828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 399928143c3dSStefano Zampini bs = 1; 400028143c3dSStefano Zampini } 4001e7931f94SStefano Zampini break; 4002e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4003e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4004e7931f94SStefano Zampini bs = 1; 4005e7931f94SStefano Zampini break; 4006e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4007e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4008e7931f94SStefano Zampini break; 4009e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4010e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4011e7931f94SStefano Zampini break; 4012e7931f94SStefano Zampini default: 40139d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4014e7931f94SStefano Zampini break; 4015e7931f94SStefano Zampini } 401628143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 401728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 401828143c3dSStefano Zampini bs = 1; 4019e7931f94SStefano Zampini } 4020e7931f94SStefano Zampini 402170cf5478SStefano Zampini /* create MATIS object if needed */ 402270cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4023e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4024e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 402570cf5478SStefano Zampini } else { 402670cf5478SStefano Zampini /* it also destroys the local matrices */ 402770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 402870cf5478SStefano Zampini } 402970cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4030e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 40319d30be91SStefano Zampini 40329d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 40339d30be91SStefano Zampini 40349d30be91SStefano Zampini /* Global to local map of received indices */ 40359d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 40369d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 40379d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 40389d30be91SStefano Zampini 40399d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 40409d30be91SStefano Zampini buf_size_idxs = 0; 40419d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40429d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 40439d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 40449d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 40459d30be91SStefano Zampini } 40469d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 40479d30be91SStefano Zampini 40489d30be91SStefano Zampini /* set preallocation */ 40499d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 40509d30be91SStefano Zampini if (!newisdense) { 40519d30be91SStefano Zampini PetscInt *new_local_nnz=0; 40529d30be91SStefano Zampini 40539d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 40549d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 40559d30be91SStefano Zampini if (n_recvs) { 40569d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 40579d30be91SStefano Zampini } 40589d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40599d30be91SStefano Zampini PetscInt j; 40609d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 40619d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 40629d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 40639d30be91SStefano Zampini } 40649d30be91SStefano Zampini } else { 40659d30be91SStefano Zampini /* TODO */ 40669d30be91SStefano Zampini } 40679d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 40689d30be91SStefano Zampini } 40699d30be91SStefano Zampini if (new_local_nnz) { 40709d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 40719d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 40729d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 40739d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 40749d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 40759d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 40769d30be91SStefano Zampini } else { 40779d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 40789d30be91SStefano Zampini } 40799d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 40809d30be91SStefano Zampini } else { 40819d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 40829d30be91SStefano Zampini } 4083e7931f94SStefano Zampini 4084e7931f94SStefano Zampini /* set values */ 4085e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 40869d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4087e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4088e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4089e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 40909d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4091e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4092e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4093e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 409428143c3dSStefano Zampini } else { 409528143c3dSStefano Zampini /* TODO */ 4096e7931f94SStefano Zampini } 4097e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4098e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4099e7931f94SStefano Zampini } 4100e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4101e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 410270cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 410370cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41049d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 41059d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4106e7931f94SStefano Zampini 4107dfd14d43SStefano Zampini #if 0 410828143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4109e7931f94SStefano Zampini Vec lvec,rvec; 4110e7931f94SStefano Zampini PetscReal infty_error; 4111e7931f94SStefano Zampini 41122a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4113e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4114e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4115e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 411670cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4117e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4118e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4119e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4120e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4121e7931f94SStefano Zampini } 412228143c3dSStefano Zampini #endif 4123e7931f94SStefano Zampini 412428143c3dSStefano Zampini /* assemble new additional is (if any) */ 412528143c3dSStefano Zampini if (nis) { 412628143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 412728143c3dSStefano Zampini 412828143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4129854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 413028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 413128143c3dSStefano Zampini psum = 0; 413228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 413328143c3dSStefano Zampini for (j=0;j<nis;j++) { 413428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 413528143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 413628143c3dSStefano Zampini psum += plen; 413728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 413828143c3dSStefano Zampini } 413928143c3dSStefano Zampini } 4140854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4141854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 414228143c3dSStefano Zampini for (i=1;i<nis;i++) { 414328143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 414428143c3dSStefano Zampini } 414528143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 414628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 414728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 414828143c3dSStefano Zampini for (j=0;j<nis;j++) { 414928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 415028143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 415128143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 415228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 415328143c3dSStefano Zampini } 415428143c3dSStefano Zampini } 415528143c3dSStefano Zampini for (i=0;i<nis;i++) { 415628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 415728143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 415828143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 415928143c3dSStefano Zampini } 416028143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 416128143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 416228143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 416328143c3dSStefano Zampini } 4164e7931f94SStefano Zampini /* free workspace */ 416528143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4166e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4167e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4168e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4169e7931f94SStefano Zampini if (isdense) { 4170e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4171e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4172e7931f94SStefano Zampini } else { 4173e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4174e7931f94SStefano Zampini } 417528143c3dSStefano Zampini if (nis) { 417628143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 417728143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 417828143c3dSStefano Zampini } 4179e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4180e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 418128143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4182e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4183e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 418428143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4185e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4186e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4187e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4188e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4189e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 419028143c3dSStefano Zampini if (nis) { 419128143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 419228143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 419328143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 419428143c3dSStefano Zampini } 419528143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 419628143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 419728143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 419828143c3dSStefano Zampini for (i=0;i<nis;i++) { 419928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 420028143c3dSStefano Zampini } 420153a05cb3SStefano Zampini *mat_n = NULL; 420228143c3dSStefano Zampini } 4203e7931f94SStefano Zampini PetscFunctionReturn(0); 4204e7931f94SStefano Zampini } 4205a57a6d2fSStefano Zampini 420612edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4207af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 420812edc857SStefano Zampini 4209c8587f34SStefano Zampini #undef __FUNCT__ 4210c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4211c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4212c8587f34SStefano Zampini { 4213c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4214c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 421520a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 42169881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 421720a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 42186e683305SStefano Zampini IS coarse_is,*isarray; 42196e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 42206e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4221f9eb5b7dSStefano Zampini PC pc_temp; 4222c8587f34SStefano Zampini PCType coarse_pc_type; 4223c8587f34SStefano Zampini KSPType coarse_ksp_type; 4224f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 42254f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 42266e683305SStefano Zampini Mat t_coarse_mat_is; 42276e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 42286e683305SStefano Zampini PetscMPIInt all_procs; 422974e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 423068457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 423122bc73bbSStefano Zampini PetscScalar *array; 42329881197aSStefano Zampini PetscErrorCode ierr; 4233fdc09c96SStefano Zampini 4234c8587f34SStefano Zampini PetscFunctionBegin; 4235c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 423668457ee5SStefano 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 */ 4237fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 42385a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4239fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4240f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4241f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4242f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4243fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 424451bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 424551bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4246dc4bcba2SStefano Zampini PC pc; 4247dc4bcba2SStefano Zampini PetscBool isbddc; 4248dc4bcba2SStefano Zampini 4249dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4250dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4251dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4252dc4bcba2SStefano Zampini if (isbddc) { 4253dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4254dc4bcba2SStefano Zampini } 4255727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4256fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4257fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4258fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4259f4ddd8eeSStefano Zampini } 4260fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4261fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4262f4ddd8eeSStefano Zampini } 426370cf5478SStefano Zampini /* reset any subassembling information */ 426470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 42656e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42666e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4267fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4268f4ddd8eeSStefano Zampini } 4269c8587f34SStefano Zampini 42706e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 42712b510759SStefano Zampini im_active = !!(pcis->n); 42722b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 42736e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 42746e683305SStefano Zampini void_procs = all_procs-active_procs; 42756e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 427674e2c79eSStefano Zampini redist = PETSC_FALSE; 427722bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 42786e683305SStefano Zampini csin_ml = PETSC_TRUE; 42796e683305SStefano Zampini ncoarse_ml = void_procs; 4280779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4281779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 42826e683305SStefano Zampini csin_ds = PETSC_TRUE; 428318a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 428418a45a71SStefano Zampini redist = PETSC_TRUE; 428518a45a71SStefano Zampini } else { 42866e683305SStefano Zampini csin_ds = PETSC_TRUE; 4287779c1cceSStefano Zampini ncoarse_ds = active_procs; 4288779c1cceSStefano Zampini redist = PETSC_TRUE; 428918a45a71SStefano Zampini } 42906e683305SStefano Zampini } else { 42916e683305SStefano Zampini csin_ml = PETSC_FALSE; 42926e683305SStefano Zampini ncoarse_ml = all_procs; 42936e683305SStefano Zampini if (void_procs) { 42946e683305SStefano Zampini csin_ds = PETSC_TRUE; 42956e683305SStefano Zampini ncoarse_ds = void_procs; 42966e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 42976e683305SStefano Zampini } else { 4298779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 429974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 430074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 430174e2c79eSStefano Zampini redist = PETSC_TRUE; 430274e2c79eSStefano Zampini } else { 43036e683305SStefano Zampini csin_ds = PETSC_FALSE; 43046e683305SStefano Zampini ncoarse_ds = all_procs; 43056e683305SStefano Zampini } 43066e683305SStefano Zampini } 430774e2c79eSStefano Zampini } 43086e683305SStefano Zampini 43096e683305SStefano Zampini /* 43106e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 43116e683305SStefano Zampini - we have not exceeded the number of levels requested 43126e683305SStefano Zampini - we can actually subassemble the active processes 43136e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 43146e683305SStefano Zampini */ 43156e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 43166e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 43176e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 43186e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 43196e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4320f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 43212b510759SStefano Zampini } else { 4322f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4323c8587f34SStefano Zampini } 4324c8587f34SStefano Zampini } 43256e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 43266e683305SStefano Zampini if (multilevel_allowed) { 43276e683305SStefano Zampini ncoarse = ncoarse_ml; 43286e683305SStefano Zampini csin = csin_ml; 432958da7f69SStefano Zampini redist = PETSC_FALSE; 43306e683305SStefano Zampini } else { 43316e683305SStefano Zampini ncoarse = ncoarse_ds; 43326e683305SStefano Zampini csin = csin_ds; 43336e683305SStefano Zampini } 4334e7931f94SStefano Zampini 4335abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4336abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4337abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4338abbbba34SStefano Zampini 4339abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 434022bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 434122bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 434222bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 434322bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4344b9b85e73SStefano Zampini #if 0 4345b9b85e73SStefano Zampini { 4346b9b85e73SStefano Zampini PetscViewer viewer; 4347b9b85e73SStefano Zampini char filename[256]; 4348b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4349b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4350b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4351b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4352b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4353b9b85e73SStefano Zampini } 4354b9b85e73SStefano Zampini #endif 4355e176bc59SStefano 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); 43566e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 43576e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43586e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4359abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4360abbbba34SStefano Zampini 43616e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 43626e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 43636e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 43646e683305SStefano Zampini const PetscInt *idxs; 43656e683305SStefano Zampini ISLocalToGlobalMapping tmap; 43666e683305SStefano Zampini 43676e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 43680be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 43696e683305SStefano Zampini /* allocate space for temporary storage */ 4370854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4371854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 43726e683305SStefano Zampini /* allocate for IS array */ 43736e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 43746e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 43756e683305SStefano Zampini nis = nisdofs + nisneu; 4376854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 43776e683305SStefano Zampini /* dofs splitting */ 43786e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 43796e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 43806e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 43816e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 43826e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 43836e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 43846e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 43856e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 43866e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 43876e683305SStefano Zampini } 43886e683305SStefano Zampini /* neumann boundaries */ 43896e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 43906e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 43916e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 43926e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 43936e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 43946e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 43956e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 43966e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 43976e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 43986e683305SStefano Zampini } 43996e683305SStefano Zampini /* free memory */ 44006e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 44016e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 44026e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 44036e683305SStefano Zampini } else { 44046e683305SStefano Zampini nis = 0; 44056e683305SStefano Zampini nisdofs = 0; 44066e683305SStefano Zampini nisneu = 0; 44076e683305SStefano Zampini isarray = NULL; 44086e683305SStefano Zampini } 44096e683305SStefano Zampini /* destroy no longer needed map */ 44106e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 44116e683305SStefano Zampini 44126e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 44136e683305SStefano Zampini coarse_mat_is = NULL; 44146e683305SStefano Zampini if (csin) { 44156e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 441674e2c79eSStefano Zampini if (redist) { 441774e2c79eSStefano Zampini PetscMPIInt rank; 4418779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 441974e2c79eSStefano Zampini 442074e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 442158da7f69SStefano Zampini spc = active_procs/ncoarse; 442258da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4423779c1cceSStefano Zampini if (im_active) { 4424779c1cceSStefano Zampini destsize = 1; 442574e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 442674e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 442774e2c79eSStefano Zampini } else { 442874e2c79eSStefano Zampini dest[0] = rank/(spc+1); 442974e2c79eSStefano Zampini } 443074e2c79eSStefano Zampini } else { 4431779c1cceSStefano Zampini destsize = 0; 44326e683305SStefano Zampini } 4433779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4434779c1cceSStefano Zampini } else if (csin_type_simple) { 44356e683305SStefano Zampini PetscMPIInt rank; 44366e683305SStefano Zampini PetscInt issize,isidx; 4437779c1cceSStefano Zampini 44386e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 44396e683305SStefano Zampini if (im_active) { 44406e683305SStefano Zampini issize = 1; 44416e683305SStefano Zampini isidx = (PetscInt)rank; 44426e683305SStefano Zampini } else { 44436e683305SStefano Zampini issize = 0; 44446e683305SStefano Zampini isidx = -1; 44456e683305SStefano Zampini } 44466e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4447779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4448b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 44496e683305SStefano Zampini } 4450779c1cceSStefano Zampini 4451779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4452779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4453779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4454779c1cceSStefano Zampini PetscInt *coarse_candidates; 4455779c1cceSStefano Zampini const PetscInt* tisindices; 4456779c1cceSStefano Zampini 4457779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4458779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4459779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4460779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4461779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4462779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4463779c1cceSStefano Zampini } 4464779c1cceSStefano Zampini } 4465779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4466779c1cceSStefano Zampini 4467779c1cceSStefano Zampini 44686e683305SStefano Zampini if (pcbddc->dbg_flag) { 44696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 44716e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 44726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4473779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 44746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 44756e683305SStefano Zampini } 44766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 44776e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44786e683305SStefano Zampini } 44796e683305SStefano Zampini /* shift the pattern on coarse candidates */ 44806e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 44816e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4482854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 44836e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 44846e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 44856e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 44866e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 44876e683305SStefano Zampini } 44886e683305SStefano Zampini if (pcbddc->dbg_flag) { 44896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 44916e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 44926e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44936e683305SStefano Zampini } 4494779c1cceSStefano Zampini } 44956e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 449653a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 449753a05cb3SStefano 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); 449853a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 449953a05cb3SStefano 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); 450053a05cb3SStefano Zampini } 45016e683305SStefano Zampini } else { 45026e683305SStefano Zampini if (pcbddc->dbg_flag) { 45036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 45056e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45066e683305SStefano Zampini } 45076e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 45086e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 45096e683305SStefano Zampini } 45106e683305SStefano Zampini 45116e683305SStefano Zampini /* create local to global scatters for coarse problem */ 451268457ee5SStefano Zampini if (compute_vecs) { 45136e683305SStefano Zampini PetscInt lrows; 45146e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 45156e683305SStefano Zampini if (coarse_mat_is) { 45166e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 45176e683305SStefano Zampini } else { 45186e683305SStefano Zampini lrows = 0; 45196e683305SStefano Zampini } 45206e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 45216e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 45226e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 45236e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45246e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45256e683305SStefano Zampini } 45266e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 45276e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4528c8587f34SStefano Zampini 4529f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4530f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4531f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4532f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4533f9eb5b7dSStefano Zampini } else { 4534f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4535f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4536c8587f34SStefano Zampini } 4537c8587f34SStefano Zampini 45386e683305SStefano Zampini /* print some info if requested */ 45396e683305SStefano Zampini if (pcbddc->dbg_flag) { 45406e683305SStefano Zampini if (!multilevel_allowed) { 45416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45426e683305SStefano Zampini if (multilevel_requested) { 45436e683305SStefano 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); 45446e683305SStefano Zampini } else if (pcbddc->max_levels) { 45456e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 45466e683305SStefano Zampini } 45476e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45486e683305SStefano Zampini } 45496e683305SStefano Zampini } 45506e683305SStefano Zampini 4551f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 45526e683305SStefano Zampini if (coarse_mat_is) { 45536e683305SStefano Zampini MatReuse coarse_mat_reuse; 45546a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 45556e683305SStefano Zampini if (pcbddc->dbg_flag) { 45566e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 45576e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 45586e683305SStefano Zampini } 4559f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4560312be037SStefano Zampini char prefix[256],str_level[16]; 4561e604994aSStefano Zampini size_t len; 45626e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4563422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4564c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4565f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 45665f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4567c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 45686e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4569c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4570c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4571e604994aSStefano Zampini /* prefix */ 4572e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4573e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4574e604994aSStefano Zampini if (!pcbddc->current_level) { 4575e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4576e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4577c8587f34SStefano Zampini } else { 4578e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4579312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4580312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 458134d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4582312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4583e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4584e604994aSStefano Zampini } 4585e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 45863e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 45873e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 45883e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 45893e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4590f9eb5b7dSStefano Zampini /* allow user customization */ 4591f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 45923e3c6dadSStefano Zampini } 45933e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 459451bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 45953e3c6dadSStefano Zampini if (nisdofs) { 45963e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 45973e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 45983e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 45993e3c6dadSStefano Zampini } 46003e3c6dadSStefano Zampini } 46013e3c6dadSStefano Zampini if (nisneu) { 46023e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 46033e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4604312be037SStefano Zampini } 4605f9eb5b7dSStefano Zampini 4606f9eb5b7dSStefano Zampini /* get some info after set from options */ 4607f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4608f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 46094f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 46106e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4611f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4612f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4613f9eb5b7dSStefano Zampini } 461439f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 46154f3a063dSStefano Zampini if (isredundant) { 46164f3a063dSStefano Zampini KSP inner_ksp; 46174f3a063dSStefano Zampini PC inner_pc; 46184f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 46194f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 46204f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 46214f3a063dSStefano Zampini } 4622f9eb5b7dSStefano Zampini 4623f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4624fa7f1dd8SStefano Zampini if (coarse_reuse) { 462581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4626fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 46276e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4628fa7f1dd8SStefano Zampini } else { 46296e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4630fa7f1dd8SStefano Zampini } 4631c8587f34SStefano Zampini if (isbddc || isnn) { 463222bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 463370cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4634b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 463522b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 46366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 46386e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 46396e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 464022b6e8a2SStefano Zampini } 464170cf5478SStefano Zampini } 464253a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 464370cf5478SStefano Zampini } else { 464422bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 464522bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 464622bc73bbSStefano Zampini } 464722bc73bbSStefano Zampini } else { 46482e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4649c8587f34SStefano Zampini } 4650c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4651c8587f34SStefano Zampini 46523301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 46535a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 46543301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 46553301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 46563301b35fSStefano Zampini } 46573301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 46583301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 46593301b35fSStefano Zampini } 46603301b35fSStefano Zampini if (pc->pmat->spd_set) { 46613301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 46623301b35fSStefano Zampini } 46636e683305SStefano Zampini /* set operators */ 46645f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 46656e683305SStefano Zampini if (pcbddc->dbg_flag) { 46666e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 46676e683305SStefano Zampini } 46686e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 46696e683305SStefano Zampini coarse_mat = 0; 46706e683305SStefano Zampini } 46716e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4672b9b85e73SStefano Zampini #if 0 4673b9b85e73SStefano Zampini { 4674b9b85e73SStefano Zampini PetscViewer viewer; 4675b9b85e73SStefano Zampini char filename[256]; 4676b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4677b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4678b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4679b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4680b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4681b9b85e73SStefano Zampini } 4682b9b85e73SStefano Zampini #endif 4683c8587f34SStefano Zampini 4684c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4685298c0119SStefano Zampini #if 0 4686c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4687c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 468898a51de6SStefano Zampini } 4689298c0119SStefano Zampini #endif 469098a51de6SStefano Zampini 469198a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 469298a51de6SStefano Zampini Vec crhs,csol; 469398a51de6SStefano Zampini PetscBool ispreonly; 469404708bb6SStefano Zampini 469598a51de6SStefano Zampini if (CoarseNullSpace) { 4696c8587f34SStefano Zampini if (isbddc) { 4697c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4698c8587f34SStefano Zampini } else { 46995fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4700c8587f34SStefano Zampini } 4701c8587f34SStefano Zampini } 4702f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4703f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4704f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4705f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 47066e683305SStefano Zampini /* hack */ 4707f347579bSStefano Zampini if (!csol) { 47082a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4709f9eb5b7dSStefano Zampini } 4710f347579bSStefano Zampini if (!crhs) { 47112a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4712f347579bSStefano Zampini } 4713cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4714cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 47156e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4716c8587f34SStefano Zampini KSP check_ksp; 47172b510759SStefano Zampini KSPType check_ksp_type; 4718c8587f34SStefano Zampini PC check_pc; 47196e683305SStefano Zampini Vec check_vec,coarse_vec; 47206a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 47212b510759SStefano Zampini PetscInt its; 47226e683305SStefano Zampini PetscBool compute_eigs; 47236e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 47246e683305SStefano Zampini PetscInt neigs; 47258e185a42SStefano Zampini const char *prefix; 4726c8587f34SStefano Zampini 47272b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 47286e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4729422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 473023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4731f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 47322b510759SStefano Zampini if (ispreonly) { 47332b510759SStefano Zampini check_ksp_type = KSPPREONLY; 47346e683305SStefano Zampini compute_eigs = PETSC_FALSE; 47352b510759SStefano Zampini } else { 4736cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 47376e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4738c8587f34SStefano Zampini } 4739c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 47406e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 47416e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 47426e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4743a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4744a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4745a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4746a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4747c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4748c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4749c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4750c8587f34SStefano Zampini /* create random vec */ 47516e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 47526e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4753c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4754c8587f34SStefano Zampini if (CoarseNullSpace) { 4755c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4756c8587f34SStefano Zampini } 47576e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4758c8587f34SStefano Zampini /* solve coarse problem */ 47596e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4760c8587f34SStefano Zampini if (CoarseNullSpace) { 47616e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4762c8587f34SStefano Zampini } 4763cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 47646e683305SStefano Zampini if (compute_eigs) { 4765854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4766854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 47676e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 47686e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 47696e683305SStefano Zampini lambda_min = eigs_r[0]; 47706e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 47716e683305SStefano Zampini if (lambda_max>lambda_min) { 4772cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4773cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4774cbcc2c2aSStefano Zampini } 4775c8587f34SStefano Zampini } 4776c8587f34SStefano Zampini } 4777cbcc2c2aSStefano Zampini 4778c8587f34SStefano Zampini /* check coarse problem residual error */ 47796e683305SStefano Zampini if (pcbddc->dbg_flag) { 47806e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 47816e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 47826e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4783c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 47846e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 47856e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4786c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4787779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 47886e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 47896e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 47906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 47916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 47926e683305SStefano Zampini if (compute_eigs) { 47936e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4794deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4795c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 47966e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 47976e683305SStefano 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); 47986e683305SStefano Zampini for (i=0;i<neigs;i++) { 47996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4800c8587f34SStefano Zampini } 48016e683305SStefano Zampini } 48026e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 48036e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48046e683305SStefano Zampini } 4805c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 48066e683305SStefano Zampini if (compute_eigs) { 48076e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 48086e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4809c8587f34SStefano Zampini } 48106e683305SStefano Zampini } 48116e683305SStefano Zampini } 4812cbcc2c2aSStefano Zampini /* print additional info */ 4813cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 48146e683305SStefano Zampini /* waits until all processes reaches this point */ 48156e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4816cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4817cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4818cbcc2c2aSStefano Zampini } 4819cbcc2c2aSStefano Zampini 48202b510759SStefano Zampini /* free memory */ 4821c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4822fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4823c8587f34SStefano Zampini PetscFunctionReturn(0); 4824c8587f34SStefano Zampini } 4825674ae819SStefano Zampini 4826f34684f1SStefano Zampini #undef __FUNCT__ 4827f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4828f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4829f34684f1SStefano Zampini { 4830f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4831f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4832f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4833dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4834dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 483573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4836dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4837f34684f1SStefano Zampini PetscErrorCode ierr; 4838f34684f1SStefano Zampini 4839f34684f1SStefano Zampini PetscFunctionBegin; 4840f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 48410e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 48420e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4843727cdba6SStefano Zampini } 4844dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 48453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4846dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4847dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4848dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4849dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4850dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4851dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 48520e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 48530e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 48540e6343abSStefano Zampini } 4855dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4856dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4857dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4858dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4859dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4860f34684f1SStefano Zampini 4861f34684f1SStefano Zampini /* check numbering */ 4862f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4863019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 4864dc456d91SStefano Zampini PetscInt i; 4865b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4866f34684f1SStefano Zampini 4867f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4868f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4869f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 48700fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4871019a44ceSStefano Zampini /* counter */ 4872019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4873019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4874019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4875019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4876019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4877019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4878019a44ceSStefano Zampini 4879f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4880f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4881727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4882f34684f1SStefano Zampini } 4883f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4884f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4885f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4886e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4887e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4888e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4889e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4890f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4891019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4892f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4893019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 4894019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 4895019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 4896b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4897019a44ceSStefano 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); 4898f34684f1SStefano Zampini } 4899f34684f1SStefano Zampini } 4900019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4901b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4902f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4903f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4904f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4905f34684f1SStefano Zampini } 4906f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4907f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4908e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4909e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4910f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4911f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4912b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4913ca8b9ea9SStefano Zampini PetscInt *gidxs; 4914ca8b9ea9SStefano Zampini 4915ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 49163bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4917f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4918f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4919f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4920f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 49214bc2dc4bSStefano 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); 4922f34684f1SStefano Zampini } 4923f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4924ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4925f34684f1SStefano Zampini } 4926f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4927302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4928f34684f1SStefano Zampini } 49298bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4930f34684f1SStefano Zampini /* get back data */ 4931f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4932f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4933674ae819SStefano Zampini PetscFunctionReturn(0); 4934674ae819SStefano Zampini } 4935674ae819SStefano Zampini 4936e456f2a8SStefano Zampini #undef __FUNCT__ 4937e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4938a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4939e456f2a8SStefano Zampini { 4940e456f2a8SStefano Zampini IS localis_t; 4941a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4942e456f2a8SStefano Zampini PetscScalar *vals; 4943e456f2a8SStefano Zampini PetscErrorCode ierr; 4944e456f2a8SStefano Zampini 4945e456f2a8SStefano Zampini PetscFunctionBegin; 4946a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4947e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4948854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4949e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4950e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4951a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4952a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 49531035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4954a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 49551035eff8SStefano Zampini } 4956a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4957e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4958e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4959a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4960a7dc3881SStefano Zampini /* now compute set in local ordering */ 4961a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4962a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4963a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4964a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4965a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4966ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4967e456f2a8SStefano Zampini lsize++; 4968e456f2a8SStefano Zampini } 4969e456f2a8SStefano Zampini } 4970854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4971a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4972ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4973e456f2a8SStefano Zampini idxs[lsize++] = i; 4974e456f2a8SStefano Zampini } 4975e456f2a8SStefano Zampini } 4976a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4977a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4978e456f2a8SStefano Zampini *localis = localis_t; 4979e456f2a8SStefano Zampini PetscFunctionReturn(0); 4980e456f2a8SStefano Zampini } 4981906d46d4SStefano Zampini 4982906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4983906d46d4SStefano Zampini #undef __FUNCT__ 4984906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4985906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4986906d46d4SStefano Zampini { 4987906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4988906d46d4SStefano Zampini PetscErrorCode ierr; 4989906d46d4SStefano Zampini 4990906d46d4SStefano Zampini PetscFunctionBegin; 4991906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4992906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4993906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4994906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4995906d46d4SStefano Zampini PetscFunctionReturn(0); 4996906d46d4SStefano Zampini } 4997906d46d4SStefano Zampini 4998906d46d4SStefano Zampini #undef __FUNCT__ 4999906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5000906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5001906d46d4SStefano Zampini { 5002906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5003906d46d4SStefano Zampini PetscErrorCode ierr; 5004906d46d4SStefano Zampini 5005906d46d4SStefano Zampini PetscFunctionBegin; 5006906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5007906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5008906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5009906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5010906d46d4SStefano Zampini PetscFunctionReturn(0); 5011906d46d4SStefano Zampini } 5012b96c3477SStefano Zampini 5013b96c3477SStefano Zampini #undef __FUNCT__ 5014b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 501508122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5016b96c3477SStefano Zampini { 5017a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5018b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5019b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5020a64f4aa4SStefano Zampini Mat S_j; 5021b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5022b96c3477SStefano Zampini PetscBool free_used_adj; 5023b96c3477SStefano Zampini PetscErrorCode ierr; 5024b96c3477SStefano Zampini 5025b96c3477SStefano Zampini PetscFunctionBegin; 5026b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5027b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 502808122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5029b96c3477SStefano Zampini used_xadj = NULL; 5030b96c3477SStefano Zampini used_adjncy = NULL; 5031b96c3477SStefano Zampini } else { 503208122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 503308122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 503408122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 503508122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5036b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5037b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5038b96c3477SStefano Zampini } else { 50392fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5040b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5041b96c3477SStefano Zampini PetscInt nvtxs; 5042b96c3477SStefano Zampini 50432fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 50442fffb893SStefano Zampini if (flg_row) { 5045b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5046b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5047b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5048b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 50492fffb893SStefano Zampini } else { 50502fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 50512fffb893SStefano Zampini used_xadj = NULL; 50522fffb893SStefano Zampini used_adjncy = NULL; 50532fffb893SStefano Zampini } 50542fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5055b96c3477SStefano Zampini } 5056b96c3477SStefano Zampini } 5057d5574798SStefano Zampini 5058d5574798SStefano Zampini /* setup sub_schurs data */ 5059a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5060a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5061a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5062a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 506306a4e24aSStefano 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); 5064a64f4aa4SStefano Zampini } else { 50656816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 506604708bb6SStefano Zampini PetscBool isseqaij; 50675feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 50685feab87aSStefano Zampini PetscInt n_vertices; 50695feab87aSStefano Zampini 50705feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 50712034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 50725feab87aSStefano Zampini } 507304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 507404708bb6SStefano Zampini if (!isseqaij) { 507504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 507604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 507704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 507804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 507904708bb6SStefano Zampini } else { 508004708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 508104708bb6SStefano Zampini } 508204708bb6SStefano Zampini } 508306a4e24aSStefano 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); 5084a64f4aa4SStefano Zampini } 5085a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5086b96c3477SStefano Zampini 5087b96c3477SStefano Zampini /* free adjacency */ 5088b96c3477SStefano Zampini if (free_used_adj) { 5089b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5090b96c3477SStefano Zampini } 5091b96c3477SStefano Zampini PetscFunctionReturn(0); 5092b96c3477SStefano Zampini } 5093b96c3477SStefano Zampini 5094b96c3477SStefano Zampini #undef __FUNCT__ 5095b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 509608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5097b96c3477SStefano Zampini { 5098b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5099b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5100b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5101b96c3477SStefano Zampini PCBDDCGraph graph; 5102b96c3477SStefano Zampini PetscErrorCode ierr; 5103b96c3477SStefano Zampini 5104b96c3477SStefano Zampini PetscFunctionBegin; 5105b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 510608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 51073301b35fSStefano Zampini IS verticesIS,verticescomm; 51083301b35fSStefano Zampini PetscInt vsize,*idxs; 5109b96c3477SStefano Zampini 5110b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 51113301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 51123301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51133301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 51143301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51153301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5116b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 51177fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 51183301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 51193301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5120b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5121b96c3477SStefano Zampini /* 5122b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5123b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5124b96c3477SStefano Zampini } 5125b96c3477SStefano Zampini */ 5126b96c3477SStefano Zampini } else { 5127b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5128b96c3477SStefano Zampini } 5129b96c3477SStefano Zampini 5130b96c3477SStefano Zampini /* sub_schurs init */ 5131a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5132a64f4aa4SStefano Zampini 5133b96c3477SStefano Zampini /* free graph struct */ 513408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5135b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5136b96c3477SStefano Zampini } 5137b96c3477SStefano Zampini PetscFunctionReturn(0); 5138b96c3477SStefano Zampini } 5139