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); 5368b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 5378b9f24d4SStefano 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; 622*d12edf2fSStefano Zampini Mat F,B0 = NULL; 623*d12edf2fSStefano Zampini Vec vec_R_benign,vec_R_benign_2; 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; 629*d12edf2fSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign,p0_lidx_I = 0; 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 IS dummy; 814019a44ceSStefano Zampini PetscInt ii[2]; 815019a44ceSStefano Zampini 816019a44ceSStefano Zampini ii[0] = 0; 817019a44ceSStefano Zampini ii[1] = pcbddc->B0_ncol; 818019a44ceSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 819019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 820019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 821019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 822019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 823019a44ceSStefano Zampini } 82425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 82588ebb749SStefano Zampini } 82688ebb749SStefano Zampini 82788ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 828f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 82906656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 83006656605SStefano Zampini if (pcbddc->coarse_phi_D) { 83106656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 83206656605SStefano Zampini } 833f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 83406656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 83506656605SStefano Zampini PetscScalar *marray; 83606656605SStefano Zampini 83706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 83806656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 839f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 840f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 841f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 842f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 843f4ddd8eeSStefano Zampini } 844f4ddd8eeSStefano Zampini } 84506656605SStefano Zampini 846f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 84706656605SStefano Zampini PetscScalar *marray; 84888ebb749SStefano Zampini 84906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8508eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 85106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 85288ebb749SStefano Zampini } 8533301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 85406656605SStefano Zampini n *= 2; 85588ebb749SStefano Zampini } 85606656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 85806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8598eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 86106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 86288ebb749SStefano Zampini } 8633301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 86406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 8658eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 86706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 86888ebb749SStefano Zampini } 86988ebb749SStefano Zampini } else { 870c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 871c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8721b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 873c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 874c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 875c0553b1fSStefano Zampini } 87688ebb749SStefano Zampini } 87706656605SStefano Zampini } 878019a44ceSStefano Zampini 87906656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 880*d12edf2fSStefano Zampini if (n_benign && (pcbddc->switch_static || pcbddc->dbg_flag)) { 881*d12edf2fSStefano Zampini const PetscInt *idxs; 882*d12edf2fSStefano Zampini 883*d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 884*d12edf2fSStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx,pcis->n-pcis->n_B,idxs,&p0_lidx_I);CHKERRQ(ierr); 885*d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 886*d12edf2fSStefano Zampini } 88706656605SStefano Zampini /* vertices */ 88806656605SStefano Zampini if (n_vertices) { 88916f15bc4SStefano Zampini 89004708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 89104708bb6SStefano Zampini 89216f15bc4SStefano Zampini if (n_R) { 89306656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 894019a44ceSStefano Zampini Mat S_V_benign; 89506656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 89616f15bc4SStefano Zampini PetscScalar *x,*y; 89704708bb6SStefano Zampini PetscBool isseqaij; 89806656605SStefano Zampini 89921eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 90006656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 90104708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 9026816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 90306656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 90406656605SStefano Zampini } else { 90506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 90606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 90706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 91206656605SStefano Zampini } 91306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 91406656605SStefano Zampini } 915019a44ceSStefano Zampini if (n_benign) { 916019a44ceSStefano Zampini const PetscScalar *vals; 917019a44ceSStefano Zampini const PetscInt *idxs; 918019a44ceSStefano Zampini PetscInt n; 919019a44ceSStefano Zampini 920019a44ceSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign);CHKERRQ(ierr); 921*d12edf2fSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign_2);CHKERRQ(ierr); 922*d12edf2fSStefano Zampini ierr = VecSet(vec_R_benign_2,0.);CHKERRQ(ierr); 923019a44ceSStefano Zampini ierr = MatGetRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 924*d12edf2fSStefano Zampini ierr = VecSetValues(vec_R_benign_2,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 925*d12edf2fSStefano Zampini ierr = VecAssemblyBegin(vec_R_benign_2);CHKERRQ(ierr); 926*d12edf2fSStefano Zampini ierr = VecAssemblyEnd(vec_R_benign_2);CHKERRQ(ierr); 927019a44ceSStefano Zampini ierr = MatRestoreRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 928*d12edf2fSStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,vec_R_benign_2,vec_R_benign);CHKERRQ(ierr); 929019a44ceSStefano Zampini ierr = VecScale(vec_R_benign,m_one);CHKERRQ(ierr); 930019a44ceSStefano Zampini } 93180677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 93206656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 93306656605SStefano Zampini if (n_constraints) { 93406656605SStefano Zampini Mat B; 93580677318SStefano Zampini 936b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 93780677318SStefano Zampini for (i=0;i<n_vertices;i++) { 93880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 93980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 94080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 94180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 94280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 94380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 94480677318SStefano Zampini } 94580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 94680677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 94780677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 94980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 95206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 95306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 95406656605SStefano Zampini } 95504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 95604708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 95704708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 95804708bb6SStefano Zampini } 95906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 960019a44ceSStefano Zampini if (n_benign) { 961019a44ceSStefano Zampini ierr = MatMatMult(B0_R,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_V_benign);CHKERRQ(ierr); 962019a44ceSStefano Zampini } 96380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 96406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 96506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 96606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 96706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 96806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 970019a44ceSStefano Zampini if (!n_benign) { 97106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 972019a44ceSStefano Zampini } else { 973019a44ceSStefano Zampini const PetscScalar *vals; 974019a44ceSStefano Zampini PetscScalar val; 975019a44ceSStefano Zampini const PetscInt *idxs; 976019a44ceSStefano Zampini PetscInt n,j; 977019a44ceSStefano Zampini 978019a44ceSStefano Zampini /* Use explicit loops to populate the local coarse matrix */ 979019a44ceSStefano Zampini ierr = MatDenseGetArray(S_VVt,&x);CHKERRQ(ierr); 980019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 981019a44ceSStefano Zampini for (i=0;i<n_vertices;i++) { 982019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+i] = x[j*n_vertices+i]; 983019a44ceSStefano Zampini } 984019a44ceSStefano Zampini } 985019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&x);CHKERRQ(ierr); 986019a44ceSStefano Zampini ierr = MatDenseGetArray(S_V_benign,&x);CHKERRQ(ierr); 987019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 988019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 989019a44ceSStefano Zampini coarse_submat_vals[n_vertices*pcbddc->local_primal_size+j] = x[j]+vals[j]; 990019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices] = x[j]+vals[j]; 991019a44ceSStefano Zampini } 992019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 993019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_V_benign,&x);CHKERRQ(ierr); 994019a44ceSStefano Zampini ierr = MatDestroy(&S_V_benign);CHKERRQ(ierr); 995*d12edf2fSStefano Zampini ierr = VecDot(vec_R_benign_2,vec_R_benign,&val);CHKERRQ(ierr); 996*d12edf2fSStefano Zampini coarse_submat_vals[pcbddc->local_primal_size*pcbddc->local_primal_size-1] = val; 997019a44ceSStefano Zampini } 99806656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 99916f15bc4SStefano Zampini } else { 1000019a44ceSStefano Zampini /* TODO benign case also here */ 1001019a44ceSStefano Zampini if (n_benign) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not yet implemented"); 100216f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 100316f15bc4SStefano Zampini } 100421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 100506656605SStefano Zampini /* coarse basis functions */ 100606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 100716f15bc4SStefano Zampini PetscScalar *y; 100816f15bc4SStefano Zampini 100906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 101106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 101506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 101606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 101706656605SStefano Zampini 101806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 101906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 102006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 102106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 102206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 102306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1024*d12edf2fSStefano Zampini y[n_D*i+p0_lidx_I] = 0.0; 102506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 102606656605SStefano Zampini } 102706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 102806656605SStefano Zampini } 1029019a44ceSStefano Zampini if (n_benign) { 1030019a44ceSStefano Zampini PetscScalar *y; 1031019a44ceSStefano Zampini 1032019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1033019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*n_vertices);CHKERRQ(ierr); 1034019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1035019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1036019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1037019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 1038019a44ceSStefano Zampini 1039019a44ceSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1040019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1041*d12edf2fSStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*n_vertices);CHKERRQ(ierr); 1042019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1043019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1044019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1045*d12edf2fSStefano Zampini y[n_D*n_vertices+p0_lidx_I] = 0.0; 1046019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1047019a44ceSStefano Zampini } 1048019a44ceSStefano Zampini } 104904708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 105004708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 105106656605SStefano Zampini } 105206656605SStefano Zampini 105306656605SStefano Zampini if (n_constraints) { 105406656605SStefano Zampini Mat B; 105506656605SStefano Zampini 105606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 105706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 105880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 105906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 106006656605SStefano Zampini if (n_vertices) { 106180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 106280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 106380677318SStefano Zampini } else { 106480677318SStefano Zampini Mat S_VCt; 106580677318SStefano Zampini 106680677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 106780677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 106880677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 106980677318SStefano Zampini } 107006656605SStefano Zampini } 107106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 107206656605SStefano Zampini /* coarse basis functions */ 107306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 107406656605SStefano Zampini PetscScalar *y; 107506656605SStefano Zampini 107606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 107706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 107806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 107906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 108206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 108306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 108406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 108506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 108606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 108906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 109006656605SStefano Zampini } 109106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 109206656605SStefano Zampini } 109306656605SStefano Zampini } 109480677318SStefano Zampini if (n_constraints) { 109580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 109680677318SStefano Zampini } 109706656605SStefano Zampini 1098019a44ceSStefano Zampini if (n_benign) { 1099019a44ceSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1100019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1101019a44ceSStefano Zampini ierr = VecDestroy(&vec_R_benign);CHKERRQ(ierr); 1102019a44ceSStefano Zampini } 1103019a44ceSStefano Zampini 110406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 11053301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 110606656605SStefano Zampini 110706656605SStefano Zampini if (n_constraints) { 110816f15bc4SStefano Zampini Mat S_CCT,B_C; 110906656605SStefano Zampini 111080677318SStefano Zampini /* this is a lazy thing */ 111180677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 111206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 111306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 111406656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 111516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 111606656605SStefano Zampini if (n_vertices) { 111716f15bc4SStefano Zampini Mat B_V,S_VCT; 111806656605SStefano Zampini 111906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 112006656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 112106656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 112206656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 112316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 112406656605SStefano Zampini } 112506656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 112680677318SStefano Zampini } else { /* if there are no constraints, reset work */ 112780677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 112806656605SStefano Zampini } 112916f15bc4SStefano Zampini if (n_vertices && n_R) { 113006656605SStefano Zampini Mat A_VRT; 113180677318SStefano Zampini PetscScalar *marray; 113206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 113306656605SStefano Zampini 113480677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 113580677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 113680677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 113706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 113880677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 113980677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 114016f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 114106656605SStefano Zampini } 114206656605SStefano Zampini 114306656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 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 = MatSolveTranspose(F,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 } else { 115206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 115306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 115406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 115506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 115606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 115806656605SStefano Zampini } 115906656605SStefano Zampini } 116006656605SStefano Zampini /* coarse basis functions */ 116106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 116206656605SStefano Zampini PetscScalar *y; 116306656605SStefano Zampini 116406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 116506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 116606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 116706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 116806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 116906656605SStefano Zampini if (i<n_vertices) { 117006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 117106656605SStefano Zampini } 117206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 117306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 117406656605SStefano Zampini 117506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 117606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 117706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 117806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 118006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 118106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 118206656605SStefano Zampini } 118306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 118406656605SStefano Zampini } 118506656605SStefano Zampini } 1186d62866d3SStefano Zampini /* free memory */ 118788ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 118806656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 118906656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 119006656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 119106656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1192d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1193d62866d3SStefano Zampini if (n_vertices) { 1194d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1195d62866d3SStefano Zampini } 1196d62866d3SStefano Zampini if (n_constraints) { 1197d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1198d62866d3SStefano Zampini } 119988ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 120088ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 120188ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1202*d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 120388ebb749SStefano Zampini Mat coarse_sub_mat; 120425084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 120588ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 120688ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 120788ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 12088bec7fa6SStefano Zampini Mat C_B,CPHI; 12098bec7fa6SStefano Zampini IS is_dummy; 12108bec7fa6SStefano Zampini Vec mones; 121188ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 121288ebb749SStefano Zampini PetscReal real_value; 121388ebb749SStefano Zampini 121488ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 121588ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 121688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 121788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 121888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 121988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1220c0553b1fSStefano Zampini if (unsymmetric_check) { 122188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 122288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 122388ebb749SStefano Zampini } 122488ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 122588ebb749SStefano Zampini 122625084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 12273301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 122825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1229c0553b1fSStefano Zampini if (unsymmetric_check) { 123088ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 123288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123388ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 123588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 123888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 124188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124288ebb749SStefano Zampini } else { 124388ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 124488ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 124588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 124788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 125088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 125188ebb749SStefano Zampini } 125288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 125388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 125488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 125588ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 1256*d12edf2fSStefano Zampini if (n_benign) { 1257*d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1258*d12edf2fSStefano Zampini PetscScalar *data,*data2; 1259*d12edf2fSStefano Zampini 1260*d12edf2fSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&is_dummy);CHKERRQ(ierr); 1261*d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 1262*d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1263*d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1264*d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1265*d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1266*d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 1267*d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1268*d12edf2fSStefano Zampini data[n_vertices*pcbddc->local_primal_size+i] += data2[i]; 1269*d12edf2fSStefano Zampini data[i*pcbddc->local_primal_size+n_vertices] += data2[i]; 1270*d12edf2fSStefano Zampini } 1271*d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1272*d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1273*d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1274*d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1275*d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1276*d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1277*d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1278*d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1279*d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1280*d12edf2fSStefano Zampini } 1281*d12edf2fSStefano Zampini #if 0 1282*d12edf2fSStefano Zampini { 1283*d12edf2fSStefano Zampini PetscViewer viewer; 1284*d12edf2fSStefano Zampini char filename[256]; 1285*d12edf2fSStefano Zampini sprintf(filename,"proj_local_coarse_mat%d.m",PetscGlobalRank); 1286*d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1287*d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1288*d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1289*d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1290*d12edf2fSStefano Zampini } 1291*d12edf2fSStefano Zampini #endif 129281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 12938bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 12940fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 129506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 12968bec7fa6SStefano Zampini 12978bec7fa6SStefano Zampini /* check constraints */ 1298*d12edf2fSStefano Zampini if (!n_benign) { 12998bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 13008bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 13018bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 13028bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 13038bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 13048bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 13058bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1306bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1307c0553b1fSStefano Zampini if (unsymmetric_check) { 1308bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1309bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1310bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1311bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1312bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 131388ebb749SStefano Zampini } 13148bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 13158bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 13168bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 13178bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1318*d12edf2fSStefano Zampini } 131925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 132088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 132188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 132288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 132388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 132488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 132588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 132688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 132788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 132888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 132988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1330c0553b1fSStefano Zampini if (unsymmetric_check) { 133188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 133288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 133388ebb749SStefano Zampini } 133488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 133588ebb749SStefano Zampini } 1336*d12edf2fSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 13378629588bSStefano Zampini /* get back data */ 13388629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 133988ebb749SStefano Zampini PetscFunctionReturn(0); 134088ebb749SStefano Zampini } 134188ebb749SStefano Zampini 134288ebb749SStefano Zampini #undef __FUNCT__ 1343d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1344d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1345aa0d41d4SStefano Zampini { 1346d65f70fdSStefano Zampini Mat *work_mat; 1347d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1348d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1349d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1350aa0d41d4SStefano Zampini PetscErrorCode ierr; 1351aa0d41d4SStefano Zampini 1352aa0d41d4SStefano Zampini PetscFunctionBegin; 1353d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1354d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1355d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1356d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1357aa0d41d4SStefano Zampini 1358d65f70fdSStefano Zampini if (!rsorted) { 1359906d46d4SStefano Zampini const PetscInt *idxs; 1360906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1361aa0d41d4SStefano Zampini 1362d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1363d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1364d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1365d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1366aa0d41d4SStefano Zampini } 1367d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1368d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1369d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1370d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1371aa0d41d4SStefano Zampini } 1372d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1373d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1374d65f70fdSStefano Zampini } else { 1375d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1376d65f70fdSStefano Zampini isrow_s = isrow; 1377aa0d41d4SStefano Zampini } 1378906d46d4SStefano Zampini 1379d65f70fdSStefano Zampini if (!csorted) { 1380d65f70fdSStefano Zampini if (isrow == iscol) { 1381d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1382d65f70fdSStefano Zampini iscol_s = isrow_s; 1383d65f70fdSStefano Zampini } else { 1384d65f70fdSStefano Zampini const PetscInt *idxs; 1385d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1386906d46d4SStefano Zampini 1387d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1388d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1389d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1390d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1391d65f70fdSStefano Zampini } 1392d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1393d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1394d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1395d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1396d65f70fdSStefano Zampini } 1397d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1398d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1399d65f70fdSStefano Zampini } 1400d65f70fdSStefano Zampini } else { 1401d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1402d65f70fdSStefano Zampini iscol_s = iscol; 1403d65f70fdSStefano Zampini } 1404d65f70fdSStefano Zampini 1405d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1406d65f70fdSStefano Zampini 1407d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1408906d46d4SStefano Zampini Mat new_mat; 1409d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1410906d46d4SStefano Zampini 1411d65f70fdSStefano Zampini if (!rsorted) { 1412d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1413d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1414d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1415d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1416906d46d4SStefano Zampini } 1417d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1418d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1419d65f70fdSStefano Zampini } else { 1420d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1421906d46d4SStefano Zampini } 1422d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1423d65f70fdSStefano Zampini 1424d65f70fdSStefano Zampini if (!csorted) { 1425d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1426d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1427d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1428d65f70fdSStefano Zampini } else { 1429d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1430d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1431d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1432d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1433d65f70fdSStefano Zampini } 1434d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1435d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1436d65f70fdSStefano Zampini } 1437d65f70fdSStefano Zampini } else { 1438d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1439d65f70fdSStefano Zampini } 1440d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1441d65f70fdSStefano Zampini 1442d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1443d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1444d65f70fdSStefano Zampini work_mat[0] = new_mat; 1445d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1446d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1447d65f70fdSStefano Zampini } 1448d65f70fdSStefano Zampini 1449d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1450d65f70fdSStefano Zampini *B = work_mat[0]; 1451d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1452d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1453d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1454d65f70fdSStefano Zampini PetscFunctionReturn(0); 1455d65f70fdSStefano Zampini } 1456d65f70fdSStefano Zampini 1457d65f70fdSStefano Zampini #undef __FUNCT__ 14585e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 14595e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1460aa0d41d4SStefano Zampini { 1461aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 14623bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 14635e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1464d65f70fdSStefano Zampini Mat new_mat; 14655e8657edSStefano Zampini IS is_local,is_global; 1466d65f70fdSStefano Zampini PetscInt local_size; 1467d65f70fdSStefano Zampini PetscBool isseqaij; 1468aa0d41d4SStefano Zampini PetscErrorCode ierr; 1469aa0d41d4SStefano Zampini 1470aa0d41d4SStefano Zampini PetscFunctionBegin; 1471aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 14725e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 14735e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 14743bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1475aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1476d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1477aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1478906d46d4SStefano Zampini 1479906d46d4SStefano Zampini /* check */ 1480906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1481906d46d4SStefano Zampini Vec x,x_change; 1482906d46d4SStefano Zampini PetscReal error; 1483906d46d4SStefano Zampini 14845e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1485906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 14865e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1487e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1488e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1489d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1490e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1491e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1492906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1493906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1494906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1495906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1496906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1497906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1498906d46d4SStefano Zampini } 1499906d46d4SStefano Zampini 150022d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 15019b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 150222d5777bSStefano Zampini if (isseqaij) { 1503d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1504aa0d41d4SStefano Zampini } else { 1505aa0d41d4SStefano Zampini Mat work_mat; 1506aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1507d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1508aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1509aa0d41d4SStefano Zampini } 15103301b35fSStefano Zampini if (matis->A->symmetric_set) { 15113301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1512e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 15133301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1514e496cd5dSStefano Zampini #endif 15153301b35fSStefano Zampini } 151645a1bb75SStefano Zampini /* 151745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1518d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 151945a1bb75SStefano Zampini */ 1520d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1521aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1522aa0d41d4SStefano Zampini } 1523aa0d41d4SStefano Zampini 1524aa0d41d4SStefano Zampini #undef __FUNCT__ 1525a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 15268ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1527a64d13efSStefano Zampini { 1528a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1529a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1530d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 153153892102SStefano Zampini PetscInt *idx_R_local=NULL; 15323a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 15333a50541eSStefano Zampini PetscInt vbs,bs; 15346816873aSStefano Zampini PetscBT bitmask=NULL; 1535a64d13efSStefano Zampini PetscErrorCode ierr; 1536a64d13efSStefano Zampini 1537a64d13efSStefano Zampini PetscFunctionBegin; 1538b23d619eSStefano Zampini /* 1539b23d619eSStefano Zampini No need to setup local scatters if 1540b23d619eSStefano Zampini - primal space is unchanged 1541b23d619eSStefano Zampini AND 1542b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1543b23d619eSStefano Zampini AND 1544b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1545b23d619eSStefano Zampini */ 1546b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1547f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1548f4ddd8eeSStefano Zampini } 1549f4ddd8eeSStefano Zampini /* destroy old objects */ 1550f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1551f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1552f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1553a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1554b371cd4fSStefano Zampini n_B = pcis->n_B; 1555b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1556b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15573a50541eSStefano Zampini 1558a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 15596816873aSStefano Zampini 156053892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 15616816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1562854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1563a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1564a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 15650e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1566a64d13efSStefano Zampini } 1567a64d13efSStefano Zampini 1568a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 15694641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 15706816873aSStefano Zampini idx_R_local[n_R++] = i; 1571a64d13efSStefano Zampini } 1572a64d13efSStefano Zampini } 157353892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 15746816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15756816873aSStefano Zampini 157653892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 157753892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 15786816873aSStefano Zampini } 15793a50541eSStefano Zampini 15803a50541eSStefano Zampini /* Block code */ 15813a50541eSStefano Zampini vbs = 1; 15823a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 15833a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 15843a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 15853a50541eSStefano Zampini PetscInt *vary; 1586d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1587785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 15883a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1589d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1590d3df7717SStefano 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 */ 15910e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1592d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 15933a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 15943a50541eSStefano Zampini is_blocked = PETSC_FALSE; 15953a50541eSStefano Zampini break; 15963a50541eSStefano Zampini } 15973a50541eSStefano Zampini } 1598d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1599d3df7717SStefano Zampini } else { 1600d3df7717SStefano Zampini /* Verify directly the R set */ 1601d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1602d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1603d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1604d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1605d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1606d3df7717SStefano Zampini break; 1607d3df7717SStefano Zampini } 1608d3df7717SStefano Zampini } 1609d3df7717SStefano Zampini } 1610d3df7717SStefano Zampini } 16113a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 16123a50541eSStefano Zampini vbs = bs; 16133a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 16143a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 16153a50541eSStefano Zampini } 16163a50541eSStefano Zampini } 16173a50541eSStefano Zampini } 16183a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 161953892102SStefano Zampini if (sub_schurs->reuse_mumps) { 162053892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 162153892102SStefano Zampini 162253892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 162353892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 162453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 162553892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 162653892102SStefano Zampini } else { 16273a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 162853892102SStefano Zampini } 1629a64d13efSStefano Zampini 1630a64d13efSStefano Zampini /* print some info if requested */ 1631a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1632019a44ceSStefano Zampini PetscInt benign; 1633019a44ceSStefano Zampini 1634019a44ceSStefano Zampini benign = (PetscInt)(!!pcbddc->zerodiag); 1635a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1636a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16370fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1638a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1639a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1640019a44ceSStefano 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); 1641a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1642a64d13efSStefano Zampini } 1643a64d13efSStefano Zampini 1644a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 16456816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16466816873aSStefano Zampini IS is_aux1,is_aux2; 16476816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 16486816873aSStefano Zampini 16493a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1650854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1651854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1652a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 16534641a718SStefano Zampini for (i=0; i<n_D; i++) { 16544641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 16554641a718SStefano Zampini } 1656a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1657a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16584641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 16594641a718SStefano Zampini aux_array1[j++] = i; 1660a64d13efSStefano Zampini } 1661a64d13efSStefano Zampini } 1662a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1663a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1664a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 16654641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 16664641a718SStefano Zampini aux_array2[j++] = i; 1667a64d13efSStefano Zampini } 1668a64d13efSStefano Zampini } 1669a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1670a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1671a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1672a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1673a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1674a64d13efSStefano Zampini 16758eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1676785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1677a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16784641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 16794641a718SStefano Zampini aux_array1[j++] = i; 1680a64d13efSStefano Zampini } 1681a64d13efSStefano Zampini } 1682a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1683a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1684a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1685a64d13efSStefano Zampini } 16864641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 16873a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1688d62866d3SStefano Zampini } else { 168953892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16906816873aSStefano Zampini IS tis; 16916816873aSStefano Zampini PetscInt schur_size; 16926816873aSStefano Zampini 169353892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 16946816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 169553892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 16966816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16976816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 16986816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 16996816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 17006816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1701d62866d3SStefano Zampini } 1702d62866d3SStefano Zampini } 1703a64d13efSStefano Zampini PetscFunctionReturn(0); 1704a64d13efSStefano Zampini } 1705a64d13efSStefano Zampini 1706304d26faSStefano Zampini 1707304d26faSStefano Zampini #undef __FUNCT__ 1708304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1709684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1710304d26faSStefano Zampini { 1711304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1712304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1713304d26faSStefano Zampini PC pc_temp; 1714304d26faSStefano Zampini Mat A_RR; 1715f4ddd8eeSStefano Zampini MatReuse reuse; 1716304d26faSStefano Zampini PetscScalar m_one = -1.0; 1717304d26faSStefano Zampini PetscReal value; 171804708bb6SStefano Zampini PetscInt n_D,n_R; 17199577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1720304d26faSStefano Zampini PetscErrorCode ierr; 1721e604994aSStefano Zampini /* prefixes stuff */ 1722312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1723e604994aSStefano Zampini size_t len; 1724304d26faSStefano Zampini 1725304d26faSStefano Zampini PetscFunctionBegin; 1726304d26faSStefano Zampini 1727e604994aSStefano Zampini /* compute prefixes */ 1728e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1729e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1730e604994aSStefano Zampini if (!pcbddc->current_level) { 1731e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1732e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1733e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1734e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1735e604994aSStefano Zampini } else { 1736e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1737312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1738e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1739e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1740312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1741312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 174234d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 174334d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1744e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1745e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1746e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1747e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1748e604994aSStefano Zampini } 1749e604994aSStefano Zampini 1750304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1751684f6988SStefano Zampini if (dirichlet) { 1752d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 17533301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 17543301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1755964fefecSStefano Zampini } 1756ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1757964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1758304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1759304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1760304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1761304d26faSStefano Zampini /* default */ 1762304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1763e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 17649577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1765304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 17669577ea80SStefano Zampini if (issbaij) { 17679577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 17689577ea80SStefano Zampini } else { 1769304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 17709577ea80SStefano Zampini } 1771304d26faSStefano Zampini /* Allow user's customization */ 1772304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1773304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1774304d26faSStefano Zampini } 1775d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1776d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1777d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1778d62866d3SStefano Zampini 1779d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1780d5574798SStefano Zampini } 1781304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1782304d26faSStefano Zampini if (!n_D) { 1783304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1784304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1785304d26faSStefano Zampini } 1786304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1787304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1788304d26faSStefano Zampini /* set ksp_D into pcis data */ 1789304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1790304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1791304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1792684f6988SStefano Zampini } 1793304d26faSStefano Zampini 1794304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1795684f6988SStefano Zampini A_RR = 0; 1796684f6988SStefano Zampini if (neumann) { 1797d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 179804708bb6SStefano Zampini PetscInt ibs,mbs; 179904708bb6SStefano Zampini PetscBool issbaij; 180004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1801f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 18028ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1803f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1804f4ddd8eeSStefano Zampini PetscInt nn_R; 180581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1806f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1807f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1808f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1809f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1810f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1811f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1812f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1813727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1814f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1815f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1816f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1817f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1818f4ddd8eeSStefano Zampini } 1819f4ddd8eeSStefano Zampini } 1820f4ddd8eeSStefano Zampini /* last check */ 1821d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1822f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1823f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1824f4ddd8eeSStefano Zampini } 1825f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1826f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1827f4ddd8eeSStefano Zampini } 1828f4ddd8eeSStefano Zampini /* extract A_RR */ 1829af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1830af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 183104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 183204708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 183304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 183404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 183504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1836af732b37SStefano Zampini } else { 183704708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 18386816873aSStefano Zampini } 183904708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 184004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 184104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 184204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 184304708bb6SStefano Zampini } else { 184404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 184504708bb6SStefano Zampini } 184604708bb6SStefano Zampini } 184704708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1848f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 18493301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 18503301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 18516816873aSStefano Zampini } 18526816873aSStefano Zampini } else { 18536816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 18546816873aSStefano Zampini 18556816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 18566816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 18576816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1858af732b37SStefano Zampini } 1859f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1860304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1861304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1862304d26faSStefano Zampini /* default */ 1863304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1864e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1865304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 18669577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 18679577ea80SStefano Zampini if (issbaij) { 18689577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 18699577ea80SStefano Zampini } else { 1870304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 18719577ea80SStefano Zampini } 1872304d26faSStefano Zampini /* Allow user's customization */ 1873304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1874304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1875304d26faSStefano Zampini } 1876d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1877304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1878304d26faSStefano Zampini if (!n_R) { 1879304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1880304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1881304d26faSStefano Zampini } 1882d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1883d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1884d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1885d62866d3SStefano Zampini 1886d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1887d62866d3SStefano Zampini } 1888304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1889304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1890684f6988SStefano Zampini } 18916816873aSStefano Zampini /* free Neumann problem's matrix */ 18926816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1893304d26faSStefano Zampini 1894304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 18950fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1896684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1897684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1898684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1899684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1900684f6988SStefano Zampini } 1901684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 19020fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 19030fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 19040fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 19050fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 19060fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1907304d26faSStefano Zampini /* need to be adapted? */ 1908b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1909b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1910b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1911304d26faSStefano Zampini /* print info */ 1912304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1913e604994aSStefano 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); 1914304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1915304d26faSStefano Zampini } 1916b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1917298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1918304d26faSStefano Zampini } 1919684f6988SStefano Zampini } 1920684f6988SStefano Zampini if (neumann) { /* Neumann */ 19216816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 19220fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 19230fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 19240fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 19250fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 19260fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1927304d26faSStefano Zampini /* need to be adapted? */ 1928b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1929b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1930304d26faSStefano Zampini /* print info */ 1931304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1932e604994aSStefano 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); 1933304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1934304d26faSStefano Zampini } 1935b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1936298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1937304d26faSStefano Zampini } 19380fccc4e9SStefano Zampini } 1939684f6988SStefano Zampini } 1940304d26faSStefano Zampini PetscFunctionReturn(0); 1941304d26faSStefano Zampini } 1942304d26faSStefano Zampini 1943304d26faSStefano Zampini #undef __FUNCT__ 1944ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 194580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1946674ae819SStefano Zampini { 1947674ae819SStefano Zampini PetscErrorCode ierr; 1948674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1949be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1950674ae819SStefano Zampini 1951674ae819SStefano Zampini PetscFunctionBegin; 1952be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 195380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 195420c7b377SStefano Zampini } 195580677318SStefano Zampini if (!pcbddc->switch_static) { 195680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 195780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 195880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 195920c7b377SStefano Zampini } 1960be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 196180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 196280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 196320c7b377SStefano Zampini } else { 1964be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1965be83ff47SStefano Zampini 196653892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196753892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196820c7b377SStefano Zampini } 1969be83ff47SStefano Zampini } else { 197080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 197580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 197680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 197780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1979674ae819SStefano Zampini } 1980674ae819SStefano Zampini } 1981be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 198280677318SStefano Zampini if (applytranspose) { 198380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 198480677318SStefano Zampini } else { 198580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 198680677318SStefano Zampini } 198753892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1988be83ff47SStefano Zampini } else { 1989be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1990be83ff47SStefano Zampini 1991be83ff47SStefano Zampini if (applytranspose) { 199253892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1993be83ff47SStefano Zampini } else { 199453892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1995be83ff47SStefano Zampini } 199653892102SStefano Zampini #endif 1997be83ff47SStefano Zampini } 199880677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 199980677318SStefano Zampini if (!pcbddc->switch_static) { 2000be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 200180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2003be83ff47SStefano Zampini } else { 2004be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2005be83ff47SStefano Zampini 200653892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 200753892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2008be83ff47SStefano Zampini } 200980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 201080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 201180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 201280677318SStefano Zampini } 201380677318SStefano Zampini } else { 201480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 201980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 202080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 202180677318SStefano Zampini } 202280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2026674ae819SStefano Zampini } 2027674ae819SStefano Zampini PetscFunctionReturn(0); 2028674ae819SStefano Zampini } 2029674ae819SStefano Zampini 2030dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2031674ae819SStefano Zampini #undef __FUNCT__ 2032674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2033dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2034674ae819SStefano Zampini { 2035674ae819SStefano Zampini PetscErrorCode ierr; 2036674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2037674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2038674ae819SStefano Zampini const PetscScalar zero = 0.0; 2039674ae819SStefano Zampini 2040674ae819SStefano Zampini PetscFunctionBegin; 2041dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2042dc359a40SStefano Zampini if (applytranspose) { 2043674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 20448eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2045dc359a40SStefano Zampini } else { 2046674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2047674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 204815aaf578SStefano Zampini } 2049efc2fbd9SStefano Zampini 2050efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to 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 array[pcbddc->local_primal_size-1] += pcbddc->benign_p0; 2056efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2057efc2fbd9SStefano Zampini } 2058efc2fbd9SStefano Zampini 205912edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 206012edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 206112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206312edc857SStefano Zampini 20649f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 206512edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 206612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 2067964fefecSStefano Zampini Vec rhs,sol; 2068964fefecSStefano Zampini 2069964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2070964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 207112edc857SStefano Zampini if (applytranspose) { 2072964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 207312edc857SStefano Zampini } else { 2074964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 207512edc857SStefano Zampini } 207612edc857SStefano Zampini } 2077674ae819SStefano Zampini 2078674ae819SStefano Zampini /* Local solution on R nodes */ 207980677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 208080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 20819f00e9b4SStefano Zampini } 2082674ae819SStefano Zampini 20839f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 20849f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 208512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2086674ae819SStefano Zampini 2087674ae819SStefano Zampini /* Sum contributions from two levels */ 2088dc359a40SStefano Zampini if (applytranspose) { 2089dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2090dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2091dc359a40SStefano Zampini } else { 2092674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 20938eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2094dc359a40SStefano Zampini } 2095efc2fbd9SStefano Zampini /* store p0 */ 2096efc2fbd9SStefano Zampini if (pcbddc->benign_change) { 2097efc2fbd9SStefano Zampini PetscScalar *array; 2098efc2fbd9SStefano Zampini 2099efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2100efc2fbd9SStefano Zampini pcbddc->benign_p0 = array[pcbddc->local_primal_size-1]; 2101efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2102efc2fbd9SStefano Zampini } 2103674ae819SStefano Zampini PetscFunctionReturn(0); 2104674ae819SStefano Zampini } 2105674ae819SStefano Zampini 2106674ae819SStefano Zampini #undef __FUNCT__ 2107674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 210812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2109674ae819SStefano Zampini { 2110674ae819SStefano Zampini PetscErrorCode ierr; 2111674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 211258da7f69SStefano Zampini PetscScalar *array; 211312edc857SStefano Zampini Vec from,to; 2114674ae819SStefano Zampini 2115674ae819SStefano Zampini PetscFunctionBegin; 211612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 211712edc857SStefano Zampini from = pcbddc->coarse_vec; 211812edc857SStefano Zampini to = pcbddc->vec1_P; 211912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 212012edc857SStefano Zampini Vec tvec; 212158da7f69SStefano Zampini 212258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 212358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 212412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 212558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 212658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 212758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 212812edc857SStefano Zampini } 212912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 213012edc857SStefano Zampini from = pcbddc->vec1_P; 213112edc857SStefano Zampini to = pcbddc->coarse_vec; 213212edc857SStefano Zampini } 213312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2134674ae819SStefano Zampini PetscFunctionReturn(0); 2135674ae819SStefano Zampini } 2136674ae819SStefano Zampini 2137674ae819SStefano Zampini #undef __FUNCT__ 2138674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 213912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2140674ae819SStefano Zampini { 2141674ae819SStefano Zampini PetscErrorCode ierr; 2142674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 214358da7f69SStefano Zampini PetscScalar *array; 214412edc857SStefano Zampini Vec from,to; 2145674ae819SStefano Zampini 2146674ae819SStefano Zampini PetscFunctionBegin; 214712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 214812edc857SStefano Zampini from = pcbddc->coarse_vec; 214912edc857SStefano Zampini to = pcbddc->vec1_P; 215012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 215112edc857SStefano Zampini from = pcbddc->vec1_P; 215212edc857SStefano Zampini to = pcbddc->coarse_vec; 215312edc857SStefano Zampini } 215412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 215512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 215612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 215712edc857SStefano Zampini Vec tvec; 215858da7f69SStefano Zampini 215912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 216058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 216158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 216258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 216358da7f69SStefano Zampini } 216458da7f69SStefano Zampini } else { 216558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 216658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 216712edc857SStefano Zampini } 216812edc857SStefano Zampini } 2169674ae819SStefano Zampini PetscFunctionReturn(0); 2170674ae819SStefano Zampini } 2171674ae819SStefano Zampini 2172984c4197SStefano Zampini /* uncomment for testing purposes */ 2173984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2174674ae819SStefano Zampini #undef __FUNCT__ 2175674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2176674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2177674ae819SStefano Zampini { 2178674ae819SStefano Zampini PetscErrorCode ierr; 2179674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2180674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2181674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2182984c4197SStefano Zampini /* one and zero */ 2183984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2184984c4197SStefano Zampini /* space to store constraints and their local indices */ 21859162d606SStefano Zampini PetscScalar *constraints_data; 21869162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 21879162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 21889162d606SStefano Zampini PetscInt *constraints_n; 2189984c4197SStefano Zampini /* iterators */ 2190b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2191984c4197SStefano Zampini /* BLAS integers */ 2192e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2193e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2194c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2195727cdba6SStefano Zampini /* reuse */ 21960e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 21970e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2198984c4197SStefano Zampini /* change of basis */ 2199b3d85658SStefano Zampini PetscBool qr_needed; 22009162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2201984c4197SStefano Zampini /* auxiliary stuff */ 220264efe560SStefano Zampini PetscInt *nnz,*is_indices; 22038a0068c3SStefano Zampini PetscInt ncc; 2204984c4197SStefano Zampini /* some quantities */ 220545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2206a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2207984c4197SStefano Zampini 2208674ae819SStefano Zampini PetscFunctionBegin; 22098e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 22108e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 22118e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2212088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2213088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 22140e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 22150e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 22160e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 22170e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 22180e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2219088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2220cf5a6209SStefano Zampini 2221cf5a6209SStefano Zampini /* print some info */ 2222cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2223cf5a6209SStefano Zampini IS vertices; 2224cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2225cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2226cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2227cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2228cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2229cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2230cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2231fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2232fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2233cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2234cf5a6209SStefano Zampini } 2235cf5a6209SStefano Zampini 2236cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 22379162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2238cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2239cf5a6209SStefano Zampini const Vec *nearnullvecs; 2240cf5a6209SStefano Zampini Vec *localnearnullsp; 2241cf5a6209SStefano Zampini PetscScalar *array; 2242cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2243cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2244674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2245b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2246674ae819SStefano Zampini PetscScalar *work; 2247674ae819SStefano Zampini PetscReal *singular_vals; 2248674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2249674ae819SStefano Zampini PetscReal *rwork; 2250674ae819SStefano Zampini #endif 2251674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2252674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2253674ae819SStefano Zampini #else 2254964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2255964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2256674ae819SStefano Zampini #endif 2257674ae819SStefano Zampini 2258674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2259d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2260d06fc5fdSStefano Zampini /* free unneeded index sets */ 2261d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2262d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2263674ae819SStefano Zampini } 2264d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2265d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2266d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2267d06fc5fdSStefano Zampini } 2268d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2269d06fc5fdSStefano Zampini n_ISForEdges = 0; 2270d06fc5fdSStefano Zampini } 2271d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2272d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2273d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2274d06fc5fdSStefano Zampini } 2275d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2276d06fc5fdSStefano Zampini n_ISForFaces = 0; 2277d06fc5fdSStefano Zampini } 227870022509SStefano Zampini 227970022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 228070022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 228170022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 228270022509SStefano Zampini if (pcbddc->NullSpace) { 228370022509SStefano Zampini PetscBool tbool[2],gbool[2]; 228470022509SStefano Zampini 228570022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2286b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2287d06fc5fdSStefano Zampini if (!ISForEdges) { 2288d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2289d06fc5fdSStefano Zampini } 2290b8ffe317SStefano Zampini } 2291d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2292d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2293d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2294d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2295d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 229698a51de6SStefano Zampini } 229770022509SStefano Zampini #endif 229808122e43SStefano Zampini 2299674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2300674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2301674ae819SStefano Zampini if (nearnullsp) { 2302674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2303f4ddd8eeSStefano Zampini /* remove any stored info */ 2304f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2305f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2306f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2307f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2308f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2309473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2310f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2311f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2312f4ddd8eeSStefano Zampini } 2313984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2314984c4197SStefano Zampini nnsp_size = 0; 2315674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2316674ae819SStefano Zampini } 2317984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2318984c4197SStefano Zampini max_constraints = nnsp_size; 2319984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2320984c4197SStefano Zampini 2321674ae819SStefano Zampini /* 2322674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 23239162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 23249162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 23259162d606SStefano Zampini There can be multiple constraints per connected component 2326674ae819SStefano Zampini */ 2327674ae819SStefano Zampini n_vertices = 0; 2328674ae819SStefano Zampini if (ISForVertices) { 2329674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2330674ae819SStefano Zampini } 23319162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 23329162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 23339162d606SStefano Zampini 23349162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 23359162d606SStefano Zampini total_counts *= max_constraints; 2336674ae819SStefano Zampini total_counts += n_vertices; 23374641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 23389162d606SStefano Zampini 2339674ae819SStefano Zampini total_counts = 0; 2340674ae819SStefano Zampini max_size_of_constraint = 0; 2341674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 23429162d606SStefano Zampini IS used_is; 2343674ae819SStefano Zampini if (i<n_ISForEdges) { 23449162d606SStefano Zampini used_is = ISForEdges[i]; 2345674ae819SStefano Zampini } else { 23469162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2347674ae819SStefano Zampini } 23489162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2349674ae819SStefano Zampini total_counts += j; 2350674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2351674ae819SStefano Zampini } 23529162d606SStefano 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); 23539162d606SStefano Zampini 2354984c4197SStefano Zampini /* get local part of global near null space vectors */ 2355785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2356984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2357984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2358e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2359e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2360984c4197SStefano Zampini } 2361674ae819SStefano Zampini 2362242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2363242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2364a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2365242a89d7SStefano Zampini 2366984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2367a773dcb8SStefano Zampini if (!skip_lapack) { 2368674ae819SStefano Zampini PetscScalar temp_work; 2369911cabfeSStefano Zampini 2370674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2371984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2372785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2373785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2374785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2375674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2376785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2377674ae819SStefano Zampini #endif 2378674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2379c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2380c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2381674ae819SStefano Zampini lwork = -1; 2382674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2383674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2384c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2385674ae819SStefano Zampini #else 2386c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2387674ae819SStefano Zampini #endif 2388674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2389984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2390674ae819SStefano Zampini #else /* on missing GESVD */ 2391674ae819SStefano Zampini /* SVD */ 2392674ae819SStefano Zampini PetscInt max_n,min_n; 2393674ae819SStefano Zampini max_n = max_size_of_constraint; 2394984c4197SStefano Zampini min_n = max_constraints; 2395984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2396674ae819SStefano Zampini min_n = max_size_of_constraint; 2397984c4197SStefano Zampini max_n = max_constraints; 2398674ae819SStefano Zampini } 2399785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2400674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2401785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2402674ae819SStefano Zampini #endif 2403674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2404674ae819SStefano Zampini lwork = -1; 2405e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2406e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2407b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2408674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2409674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24109162d606SStefano 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)); 2411674ae819SStefano Zampini #else 24129162d606SStefano 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)); 2413674ae819SStefano Zampini #endif 2414674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2415984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2416984c4197SStefano Zampini #endif /* on missing GESVD */ 2417674ae819SStefano Zampini /* Allocate optimal workspace */ 2418674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2419854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2420674ae819SStefano Zampini } 2421674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2422674ae819SStefano Zampini total_counts = 0; 24239162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 24249162d606SStefano Zampini constraints_data_ptr[0] = 0; 2425674ae819SStefano Zampini /* vertices */ 24269162d606SStefano Zampini if (n_vertices) { 2427674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 24289162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 24299162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2430674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 24319162d606SStefano Zampini constraints_n[total_counts] = 1; 24329162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24339162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24349162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2435674ae819SStefano Zampini total_counts++; 2436674ae819SStefano Zampini } 2437674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2438984c4197SStefano Zampini PetscBool used_vertex; 2439674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2440674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2441674ae819SStefano Zampini k = 0; 2442674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2443984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2444984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 24459162d606SStefano Zampini constraints_n[total_counts] = 1; 24469162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 24479162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24489162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24499162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2450674ae819SStefano Zampini total_counts++; 2451674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2452674ae819SStefano Zampini } 2453984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2454674ae819SStefano Zampini k++; 2455674ae819SStefano Zampini } 2456674ae819SStefano Zampini } 2457674ae819SStefano Zampini } 2458674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2459674ae819SStefano Zampini n_vertices = total_counts; 2460674ae819SStefano Zampini } 2461984c4197SStefano Zampini 2462674ae819SStefano Zampini /* edges and faces */ 24639162d606SStefano Zampini total_counts_cc = total_counts; 2464911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 24659162d606SStefano Zampini IS used_is; 24669162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 24679162d606SStefano Zampini 2468911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 24699162d606SStefano Zampini used_is = ISForEdges[ncc]; 2470984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2471674ae819SStefano Zampini } else { 24729162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2473984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2474674ae819SStefano Zampini } 2475674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 24769162d606SStefano Zampini 24779162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 24789162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2479984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2480984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2481674ae819SStefano Zampini if (nnsp_has_cnst) { 24825b08dc53SStefano Zampini PetscScalar quad_value; 24839162d606SStefano Zampini 24849162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24859162d606SStefano Zampini idxs_copied = PETSC_TRUE; 24869162d606SStefano Zampini 2487a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2488674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2489a773dcb8SStefano Zampini } else { 2490a773dcb8SStefano Zampini quad_value = 1.0; 2491a773dcb8SStefano Zampini } 2492674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24939162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2494674ae819SStefano Zampini } 24959162d606SStefano Zampini temp_constraints++; 2496674ae819SStefano Zampini total_counts++; 2497674ae819SStefano Zampini } 2498674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2499984c4197SStefano Zampini PetscReal real_value; 25009162d606SStefano Zampini PetscScalar *ptr_to_data; 25019162d606SStefano Zampini 2502984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 25039162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2504674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 25059162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2506674ae819SStefano Zampini } 2507984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2508984c4197SStefano Zampini /* check if array is null on the connected component */ 2509e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 25109162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 25115b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2512674ae819SStefano Zampini temp_constraints++; 2513674ae819SStefano Zampini total_counts++; 25149162d606SStefano Zampini if (!idxs_copied) { 25159162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 25169162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2517674ae819SStefano Zampini } 2518674ae819SStefano Zampini } 25199162d606SStefano Zampini } 25209162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 252145a1bb75SStefano Zampini valid_constraints = temp_constraints; 2522eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2523a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 25249162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 25259162d606SStefano Zampini 25269162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2527a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 25289162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2529a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 25309162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2531a773dcb8SStefano Zampini } else { /* perform SVD */ 2532984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 25339162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2534674ae819SStefano Zampini 2535674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2536984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2537984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2538984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2539984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2540984c4197SStefano Zampini from that computed using LAPACKgesvd 2541984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2542984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2543984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2544674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2545e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2546984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2547674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2548674ae819SStefano Zampini for (k=0;k<j+1;k++) { 25499162d606SStefano 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)); 2550674ae819SStefano Zampini } 2551674ae819SStefano Zampini } 2552e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2553e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2554e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2555674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2556c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2557674ae819SStefano Zampini #else 2558c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2559674ae819SStefano Zampini #endif 2560674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2561984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2562984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2563674ae819SStefano Zampini j = 0; 2564984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2565674ae819SStefano Zampini total_counts = total_counts-j; 256645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2567e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2568c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2569c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2570c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2571c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2572c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2573c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2574674ae819SStefano Zampini if (j<temp_constraints) { 2575984c4197SStefano Zampini PetscInt ii; 2576984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2577674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25789162d606SStefano 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)); 2579674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2580984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2581674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 25829162d606SStefano 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]; 2583674ae819SStefano Zampini } 2584674ae819SStefano Zampini } 2585674ae819SStefano Zampini } 2586674ae819SStefano Zampini #else /* on missing GESVD */ 2587e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2588e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2589b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2590674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2591674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25929162d606SStefano 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)); 2593674ae819SStefano Zampini #else 25949162d606SStefano 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)); 2595674ae819SStefano Zampini #endif 2596984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2597674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2598984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2599e310c8b4SStefano Zampini k = temp_constraints; 2600e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2601674ae819SStefano Zampini j = 0; 2602e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 260345a1bb75SStefano Zampini valid_constraints = k-j; 2604911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2605984c4197SStefano Zampini #endif /* on missing GESVD */ 2606674ae819SStefano Zampini } 2607a773dcb8SStefano Zampini } 26089162d606SStefano Zampini /* update pointers information */ 26099162d606SStefano Zampini if (valid_constraints) { 26109162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 26119162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 26129162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 26139162d606SStefano Zampini /* set change_of_basis flag */ 261445a1bb75SStefano Zampini if (boolforchange) { 2615b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 26169162d606SStefano Zampini } 2617b3d85658SStefano Zampini total_counts_cc++; 261845a1bb75SStefano Zampini } 261945a1bb75SStefano Zampini } 2620984c4197SStefano Zampini /* free workspace */ 26218f1c130eSStefano Zampini if (!skip_lapack) { 2622984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2623984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2624984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2625984c4197SStefano Zampini #endif 2626984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2627984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2628984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2629984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2630984c4197SStefano Zampini #endif 2631984c4197SStefano Zampini } 2632984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2633984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2634984c4197SStefano Zampini } 2635984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2636cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2637cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2638cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2639cf5a6209SStefano Zampini } 2640cf5a6209SStefano Zampini if (n_ISForFaces) { 2641cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2642cf5a6209SStefano Zampini } 2643cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2644cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2645cf5a6209SStefano Zampini } 2646cf5a6209SStefano Zampini if (n_ISForEdges) { 2647cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2648cf5a6209SStefano Zampini } 2649cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 265008122e43SStefano Zampini } else { 265108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2652984c4197SStefano Zampini 265308122e43SStefano Zampini total_counts = 0; 265408122e43SStefano Zampini n_vertices = 0; 2655d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2656d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 265708122e43SStefano Zampini } 265808122e43SStefano Zampini max_constraints = 0; 26599162d606SStefano Zampini total_counts_cc = 0; 266008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 266108122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 26629162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 266308122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 266408122e43SStefano Zampini } 26659162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 26669162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 26679162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 26689162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 266974d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 26709162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 26719162d606SStefano Zampini total_counts_cc = 0; 26729162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 26739162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 26749162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 267508122e43SStefano Zampini } 267608122e43SStefano Zampini } 26779162d606SStefano Zampini #if 0 26789162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 26799162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26809162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 26819162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 26829162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 26839162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 26849162d606SStefano Zampini } 26859162d606SStefano Zampini printf("\n"); 26869162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 26879162d606SStefano Zampini } 26881b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 26898bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 26901b968477SStefano Zampini } 26911b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 26928bec7fa6SStefano 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]); 26931b968477SStefano Zampini } 269408122e43SStefano Zampini #endif 269508122e43SStefano Zampini 26968bec7fa6SStefano Zampini max_size_of_constraint = 0; 26979162d606SStefano 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]); 26989162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 269908122e43SStefano Zampini /* Change of basis */ 2700b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 270108122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 270208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 270308122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2704b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 270508122e43SStefano Zampini } 270608122e43SStefano Zampini } 270708122e43SStefano Zampini } 270808122e43SStefano Zampini } 2709984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2710019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2711019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 271208122e43SStefano Zampini 27139162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 27149162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 27159162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 27169162d606SStefano 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); 271708122e43SStefano Zampini } 2718674ae819SStefano Zampini 2719674ae819SStefano Zampini /* Create constraint matrix */ 2720674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 272116f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2722984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2723984c4197SStefano Zampini 2724984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2725a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2726a717540cSStefano Zampini qr_needed = PETSC_FALSE; 272774d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2728984c4197SStefano Zampini total_primal_vertices=0; 2729b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 27309162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 27319162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2732984c4197SStefano Zampini if (size_of_constraint == 1) { 27339162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2734b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 273564efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 27369162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27379162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2738a717540cSStefano Zampini } 2739b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 274074d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2741a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2742a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2743a717540cSStefano Zampini } 2744fa434743SStefano Zampini } else { 2745b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2746fa434743SStefano Zampini } 2747a717540cSStefano Zampini } 2748b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2749b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2750674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 275170022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2752b3d85658SStefano Zampini 2753019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2754019a44ceSStefano 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); 27550e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 27560e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2757984c4197SStefano Zampini 2758984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 275974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2760785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2761984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 276274d5cdf7SStefano Zampini 2763984c4197SStefano Zampini j = total_primal_vertices; 276474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2765b3d85658SStefano Zampini cum = total_primal_vertices; 27669162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27674641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2768b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2769b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2770b3d85658SStefano Zampini cum++; 27719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 277274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 277374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 277474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 277574d5cdf7SStefano Zampini } 27769162d606SStefano Zampini j += constraints_n[i]; 2777674ae819SStefano Zampini } 2778674ae819SStefano Zampini } 2779674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2780674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2781088faed8SStefano Zampini 2782674ae819SStefano Zampini /* set values in constraint matrix */ 2783984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 27840e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2785674ae819SStefano Zampini } 2786984c4197SStefano Zampini total_counts = total_primal_vertices; 27879162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27884641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 27899162d606SStefano Zampini PetscInt *cols; 27909162d606SStefano Zampini 27919162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 27929162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 27939162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27949162d606SStefano Zampini PetscInt row = total_counts+k; 27959162d606SStefano Zampini PetscScalar *vals; 27969162d606SStefano Zampini 27979162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 27989162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 27999162d606SStefano Zampini } 28009162d606SStefano Zampini total_counts += constraints_n[i]; 2801674ae819SStefano Zampini } 2802674ae819SStefano Zampini } 2803674ae819SStefano Zampini /* assembling */ 2804674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2805674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2806088faed8SStefano Zampini 2807984c4197SStefano Zampini /* 280845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2809984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2810984c4197SStefano Zampini */ 2811674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2812674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2813026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2814984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2815984c4197SStefano Zampini /* working stuff for GEQRF */ 281681d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2817984c4197SStefano Zampini PetscBLASInt lqr_work; 2818984c4197SStefano Zampini /* working stuff for UNGQR */ 2819984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2820984c4197SStefano Zampini PetscBLASInt lgqr_work; 2821984c4197SStefano Zampini /* working stuff for TRTRS */ 2822984c4197SStefano Zampini PetscScalar *trs_rhs; 28233f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2824984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2825984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2826984c4197SStefano Zampini PetscScalar *start_vals; 2827984c4197SStefano Zampini /* working stuff for values insertion */ 28284641a718SStefano Zampini PetscBT is_primal; 282964efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2830906d46d4SStefano Zampini /* matrix sizes */ 2831906d46d4SStefano Zampini PetscInt global_size,local_size; 2832906d46d4SStefano Zampini /* temporary change of basis */ 2833906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2834cf5a6209SStefano Zampini /* extra space for debugging */ 2835cf5a6209SStefano Zampini PetscScalar *dbg_work; 2836984c4197SStefano Zampini 2837906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2838906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 283916f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2840bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2841906d46d4SStefano Zampini /* nonzeros for local mat */ 2842bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2843bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 28449162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2845a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 28469162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2847a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 28489162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2849a717540cSStefano Zampini } else { 28509162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 28519162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2852a717540cSStefano Zampini } 2853a717540cSStefano Zampini } 2854a717540cSStefano Zampini } 2855906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2856bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2857a717540cSStefano Zampini /* Set initial identity in the matrix */ 2858bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2859906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2860a717540cSStefano Zampini } 2861a717540cSStefano Zampini 2862a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2863a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2864a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2865a717540cSStefano Zampini } 2866a717540cSStefano Zampini 2867a717540cSStefano Zampini 2868a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2869a717540cSStefano Zampini /* 2870a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2871a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2872a717540cSStefano Zampini 2873a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2874a717540cSStefano Zampini 2875a6b551f4SStefano 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) 2876a6b551f4SStefano Zampini 2877a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2878a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2879a717540cSStefano Zampini | ... | 2880a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2881a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2882a717540cSStefano Zampini 2883a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2884a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2885a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2886a6b551f4SStefano Zampini 2887a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2888a717540cSStefano Zampini */ 2889a717540cSStefano Zampini if (qr_needed) { 2890984c4197SStefano Zampini /* space to store Q */ 2891854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2892984c4197SStefano Zampini /* first we issue queries for optimal work */ 28933f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28943f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 28953f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2896984c4197SStefano Zampini lqr_work = -1; 28973f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2898984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2899984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2900785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2901984c4197SStefano Zampini lgqr_work = -1; 29023f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 29033f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 29043f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 29053f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29063f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 29073f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2908984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2909984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2910785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2911984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2912785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2913984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2914785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2915a717540cSStefano Zampini /* allocating workspace for check */ 2916a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2917cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2918a717540cSStefano Zampini } 2919a717540cSStefano Zampini } 2920984c4197SStefano Zampini /* array to store whether a node is primal or not */ 29214641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2922473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 29230e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 292439e2fb2aSStefano Zampini if (i != total_primal_vertices) { 292539e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 29264641a718SStefano Zampini } 292739e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 292839e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 292939e2fb2aSStefano Zampini } 293039e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2931984c4197SStefano Zampini 2932a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 29339162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 29349162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 29354641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2936984c4197SStefano Zampini /* get constraint info */ 29379162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2938984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2939984c4197SStefano Zampini 2940984c4197SStefano Zampini if (pcbddc->dbg_flag) { 29419162d606SStefano 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); 2942674ae819SStefano Zampini } 2943984c4197SStefano Zampini 2944fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2945a717540cSStefano Zampini 2946a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2947a717540cSStefano Zampini if (pcbddc->dbg_flag) { 29489162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2949a717540cSStefano Zampini } 2950984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 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 /* compute QR decomposition of constraints */ 29543f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29553f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2957674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29583f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2959984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2960674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2961984c4197SStefano Zampini 2962984c4197SStefano Zampini /* explictly compute R^-T */ 2963984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2964984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 29653f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29663f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 29673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29683f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2969984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29703f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2971984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2972984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2973984c4197SStefano Zampini 2974a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 29753f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29763f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29773f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29783f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2979984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29803f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2981984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2982984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2983984c4197SStefano Zampini 2984984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2985984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2986984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 29873f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29883f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29893f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29903f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29913f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 29923f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2993984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29949162d606SStefano 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)); 2995984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 29969162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2997984c4197SStefano Zampini 2998984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 29999162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3000984c4197SStefano Zampini /* insert cols for primal dofs */ 3001984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3002984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 30039162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3004906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3005984c4197SStefano Zampini } 3006984c4197SStefano Zampini /* insert cols for dual dofs */ 3007984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 30089162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3009984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 30109162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3011906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3012984c4197SStefano Zampini j++; 3013674ae819SStefano Zampini } 3014674ae819SStefano Zampini } 3015984c4197SStefano Zampini 3016984c4197SStefano Zampini /* check change of basis */ 3017984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3018984c4197SStefano Zampini PetscInt ii,jj; 3019984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3020c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3021c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3022c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3023c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3024c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3025c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3026984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3027cf5a6209SStefano 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)); 3028984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3029984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3030984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3031cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3032cf5a6209SStefano 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; 3033674ae819SStefano Zampini } 3034674ae819SStefano Zampini } 3035984c4197SStefano Zampini if (!valid_qr) { 303622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3037984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3038984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3039cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3040cf5a6209SStefano 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])); 3041674ae819SStefano Zampini } 3042cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3043cf5a6209SStefano 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])); 3044984c4197SStefano Zampini } 3045984c4197SStefano Zampini } 3046984c4197SStefano Zampini } 3047674ae819SStefano Zampini } else { 304822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3049674ae819SStefano Zampini } 3050674ae819SStefano Zampini } 3051a717540cSStefano Zampini } else { /* simple transformation block */ 3052a717540cSStefano Zampini PetscInt row,col; 3053a6b551f4SStefano Zampini PetscScalar val,norm; 3054a6b551f4SStefano Zampini 3055a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30569162d606SStefano 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)); 3057a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 30589162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 30599162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3060bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 30619162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3062906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 30639162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3064a717540cSStefano Zampini } else { 3065a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 30669162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3067a717540cSStefano Zampini if (row != col) { 30689162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3069a717540cSStefano Zampini } else { 30709162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3071a717540cSStefano Zampini } 3072906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3073a717540cSStefano Zampini } 3074a717540cSStefano Zampini } 3075a717540cSStefano Zampini } 307698a51de6SStefano Zampini if (pcbddc->dbg_flag) { 307722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3078a717540cSStefano Zampini } 3079674ae819SStefano Zampini } 3080984c4197SStefano Zampini } else { 3081984c4197SStefano Zampini if (pcbddc->dbg_flag) { 30829162d606SStefano 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); 3083674ae819SStefano Zampini } 3084674ae819SStefano Zampini } 3085674ae819SStefano Zampini } 3086a717540cSStefano Zampini 3087a717540cSStefano Zampini /* free workspace */ 3088a717540cSStefano Zampini if (qr_needed) { 3089984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3090cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3091984c4197SStefano Zampini } 3092984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3093984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3094984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3095984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3096984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3097674ae819SStefano Zampini } 3098a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3099906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3100906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3101906d46d4SStefano Zampini 3102906d46d4SStefano Zampini /* assembling of global change of variable */ 3103bbb9e6c6SStefano Zampini { 3104bbb9e6c6SStefano Zampini Mat tmat; 310516f15bc4SStefano Zampini PetscInt bs; 310616f15bc4SStefano Zampini 3107906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3108906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3109bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3110bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3111bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3112bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 311316f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 311416f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3115906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3116bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3117bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3118bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3119bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3120bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3121e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3122e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3123bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3124bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3125906d46d4SStefano Zampini } 3126906d46d4SStefano Zampini /* check */ 3127906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3128906d46d4SStefano Zampini PetscReal error; 3129906d46d4SStefano Zampini Vec x,x_change; 3130906d46d4SStefano Zampini 3131906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3132906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3133906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3134906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3135e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3136e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3137bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3138e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3139e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3140906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3141906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3142906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3143906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3144bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3145906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3146906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3147906d46d4SStefano Zampini } 3148b96c3477SStefano Zampini 3149b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3150b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3151b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3152b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3153ac632422SStefano Zampini Mat S_new,tmat; 3154bbb9e6c6SStefano Zampini IS is_all_N; 3155bbb9e6c6SStefano Zampini 3156bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 31576816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3158bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3159ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3160b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3161ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3162ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3163ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3164ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3165ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3166b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3167ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3168ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3169ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3170ac632422SStefano Zampini } 3171b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3172b96c3477SStefano Zampini } 3173b96c3477SStefano Zampini } 3174906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3175906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3176b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3177b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3178b9b85e73SStefano Zampini } 3179906d46d4SStefano Zampini 3180906d46d4SStefano Zampini /* set up change of basis context */ 3181906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3182906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3183906d46d4SStefano Zampini 3184906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3185906d46d4SStefano Zampini PetscInt global_size,local_size; 3186906d46d4SStefano Zampini 3187906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3188906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3189906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3190906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3191906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3192906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3193906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3194906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3195906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3196906d46d4SStefano Zampini } else { 3197906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3198906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3199906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3200906d46d4SStefano Zampini } 3201906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3202906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3203906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3204906d46d4SStefano Zampini } else { 3205906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3206906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3207906d46d4SStefano Zampini } 3208906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3209906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3210906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3211906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3212b9b85e73SStefano Zampini } 3213a717540cSStefano Zampini 3214019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3215019a44ceSStefano Zampini if (pcbddc->zerodiag) { 3216019a44ceSStefano Zampini const PetscInt *idxs; 3217019a44ceSStefano Zampini PetscInt nz; 3218019a44ceSStefano Zampini 3219019a44ceSStefano Zampini ierr = ISGetLocalSize(pcbddc->zerodiag,&nz);CHKERRQ(ierr); 3220019a44ceSStefano Zampini ierr = ISGetIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3221019a44ceSStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3222019a44ceSStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3223019a44ceSStefano Zampini ierr = ISRestoreIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3224019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3225019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3226019a44ceSStefano Zampini pcbddc->local_primal_size++; 3227019a44ceSStefano Zampini } 3228019a44ceSStefano Zampini 3229019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3230727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3231727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3232aff50787SStefano 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); 3233c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 32340e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3235aff50787SStefano 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); 3236727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3237727cdba6SStefano Zampini } 32380e6343abSStefano Zampini } 32390e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3240727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3241727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3242727cdba6SStefano Zampini 3243a717540cSStefano Zampini /* flush dbg viewer */ 3244b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3245b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3246b8ffe317SStefano Zampini } 3247a717540cSStefano Zampini 3248e310c8b4SStefano Zampini /* free workspace */ 3249a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 32504641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 325108122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 32529162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 32539162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 325408122e43SStefano Zampini } else { 32559162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 32569162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 32579162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 325808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 325908122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 32609162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 32619162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 326208122e43SStefano Zampini } 3263674ae819SStefano Zampini PetscFunctionReturn(0); 3264674ae819SStefano Zampini } 3265674ae819SStefano Zampini 3266674ae819SStefano Zampini #undef __FUNCT__ 3267674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3268674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3269674ae819SStefano Zampini { 3270674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3271674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3272674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 32737fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3274674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3275674ae819SStefano Zampini 3276674ae819SStefano Zampini PetscFunctionBegin; 32778e61c736SStefano Zampini /* Reset previously computed graph */ 32788e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3279674ae819SStefano Zampini /* Init local Graph struct */ 32807fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 32813bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3282674ae819SStefano Zampini 3283575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3284575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3285575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3286575ad6abSStefano Zampini } 32879577ea80SStefano Zampini 3288674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 32894d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 32904d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 32914d379d7bSStefano Zampini PetscInt nvtxs; 3292e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3293674ae819SStefano Zampini 32944d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 32952fffb893SStefano Zampini 32962fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32972fffb893SStefano Zampini if (flg_row) { 32984d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3299b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 33002fffb893SStefano Zampini } 33012fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 33029b28b941SStefano 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 */ 33034d379d7bSStefano Zampini IS is_dummy; 33044d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 33054d379d7bSStefano Zampini PetscInt j,sum; 33064d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 33074d379d7bSStefano Zampini const PetscInt *idxs; 33084d379d7bSStefano Zampini PCBDDCGraph graph; 33094d379d7bSStefano Zampini PetscBT is_on_boundary; 33104d379d7bSStefano Zampini 33114d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 33124d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 33134d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 33144d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 33157fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 33164d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3317e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3318e496cd5dSStefano Zampini if (flg_row) { 33194d379d7bSStefano Zampini graph->xadj = xadj; 33204d379d7bSStefano Zampini graph->adjncy = adjncy; 3321e496cd5dSStefano Zampini } 33224d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 33234d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3324e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 33254d379d7bSStefano Zampini 33264d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 33279b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 33284d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33294d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 33304d379d7bSStefano Zampini } 33314d379d7bSStefano Zampini } 33324d379d7bSStefano Zampini 3333e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 33344d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 33354d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 33364d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 33374d379d7bSStefano Zampini } 33384d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 33394d379d7bSStefano Zampini 3340e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 33414d379d7bSStefano Zampini sum = 0; 33424d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33434d379d7bSStefano Zampini PetscInt sizecc = 0; 33444d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33454d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33464d379d7bSStefano Zampini sizecc++; 33474d379d7bSStefano Zampini } 33484d379d7bSStefano Zampini } 33494d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33504d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33514d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 33524d379d7bSStefano Zampini } 33534d379d7bSStefano Zampini } 33544d379d7bSStefano Zampini sum += sizecc*sizecc; 33554d379d7bSStefano Zampini } 33564d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 33574d379d7bSStefano Zampini sum = 0; 3358e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 33594d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 33604d379d7bSStefano Zampini cxadj[i] = sum; 33614d379d7bSStefano Zampini sum += temp; 33624d379d7bSStefano Zampini } 3363e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 33644d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33654d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33664d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33674d379d7bSStefano Zampini PetscInt k,sizecc = 0; 33684d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 33694d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 33704d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 33714d379d7bSStefano Zampini sizecc++; 33724d379d7bSStefano Zampini } 33734d379d7bSStefano Zampini } 33744d379d7bSStefano Zampini } 33754d379d7bSStefano Zampini } 33764d379d7bSStefano Zampini } 33779b28b941SStefano Zampini if (sum) { 3378e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 33794d379d7bSStefano Zampini } else { 33804d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 33814d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 33824d379d7bSStefano Zampini } 33834d379d7bSStefano Zampini graph->xadj = 0; 33844d379d7bSStefano Zampini graph->adjncy = 0; 33854d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 33864d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 33874d379d7bSStefano Zampini } 3388674ae819SStefano Zampini } 33899b28b941SStefano Zampini if (pcbddc->dbg_flag) { 33909b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3391674ae819SStefano Zampini } 3392674ae819SStefano Zampini 339363602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3394674ae819SStefano Zampini vertex_size = 1; 339563602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 339663602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 339795ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 339863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3399e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 340063602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3401674ae819SStefano Zampini } 340263602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 340363602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 340463602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3405674ae819SStefano Zampini } 340663602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3407674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 340863602bcaSStefano Zampini } else { 340963602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 341063602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3411854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 341263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 341363602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 341463602bcaSStefano Zampini } 341563602bcaSStefano Zampini } 3416674ae819SStefano Zampini } 3417674ae819SStefano Zampini 3418674ae819SStefano Zampini /* Setup of Graph */ 3419785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3420e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3421785d1243SStefano Zampini } 3422785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3423e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3424785d1243SStefano Zampini } 3425302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3426674ae819SStefano Zampini 3427674ae819SStefano Zampini /* Graph's connected components analysis */ 3428674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3429674ae819SStefano Zampini 3430674ae819SStefano Zampini /* print some info to stdout */ 3431674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3432302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3433674ae819SStefano Zampini } 3434fb180af4SStefano Zampini 3435fb180af4SStefano Zampini /* mark topography has done */ 3436fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3437674ae819SStefano Zampini PetscFunctionReturn(0); 3438674ae819SStefano Zampini } 3439674ae819SStefano Zampini 3440dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3441674ae819SStefano Zampini #undef __FUNCT__ 3442674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3443dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3444674ae819SStefano Zampini { 3445dc456d91SStefano Zampini PetscSF sf; 3446dc456d91SStefano Zampini PetscLayout map; 3447dc456d91SStefano Zampini const PetscInt *idxs; 3448dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3449dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3450dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3451dc456d91SStefano Zampini PetscMPIInt commsize; 3452674ae819SStefano Zampini PetscBool first_found; 3453674ae819SStefano Zampini PetscErrorCode ierr; 3454674ae819SStefano Zampini 3455674ae819SStefano Zampini PetscFunctionBegin; 3456dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3457dc456d91SStefano Zampini if (subset_mult) { 3458dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3459dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3460dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3461674ae819SStefano Zampini } 3462dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3463dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3464dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3465dc456d91SStefano Zampini for (i=0;i<n;i++) { 3466dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3467dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3468674ae819SStefano Zampini } 3469dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3470dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3471dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3472dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3473dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3474dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3475dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3476dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3477dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3478dc456d91SStefano Zampini 3479dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3480dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3481dc456d91SStefano Zampini if (subset_mult) { 3482dc456d91SStefano Zampini const PetscInt* idxs_mult; 3483dc456d91SStefano Zampini 3484dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3485dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3486dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3487674ae819SStefano Zampini } else { 3488dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3489674ae819SStefano Zampini } 3490dc456d91SStefano Zampini /* local size of new subset */ 3491dc456d91SStefano Zampini n_n = 0; 3492dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3493dc456d91SStefano Zampini 3494dc456d91SStefano Zampini /* global indexes in layout */ 3495dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3496dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3497dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3498dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3499dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3500dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3501dc456d91SStefano Zampini 3502dc456d91SStefano Zampini /* reduce from leaves to roots */ 3503dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 350464a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 350564a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3506dc456d91SStefano Zampini 3507dc456d91SStefano Zampini /* count indexes in local part of layout */ 3508674ae819SStefano Zampini nlocals = 0; 3509674ae819SStefano Zampini first_index = -1; 3510674ae819SStefano Zampini first_found = PETSC_FALSE; 3511dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3512dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3513674ae819SStefano Zampini first_found = PETSC_TRUE; 3514674ae819SStefano Zampini first_index = i; 3515674ae819SStefano Zampini } 3516dc456d91SStefano Zampini nlocals += root_data[i]; 3517674ae819SStefano Zampini } 3518dc456d91SStefano Zampini 3519dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 35205fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3521dc456d91SStefano Zampini start = 0; 352264a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 35235fa240b1SStefano Zampini #else 352464a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 35255fa240b1SStefano Zampini start = start-nlocals; 35265fa240b1SStefano Zampini #endif 35275fa240b1SStefano Zampini 3528dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3529dc456d91SStefano Zampini *N_n = start + nlocals; 3530dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3531dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3532674ae819SStefano Zampini } 35335fa240b1SStefano Zampini 35345fa240b1SStefano Zampini /* adapt root data with cumulative */ 3535674ae819SStefano Zampini if (first_found) { 3536dc456d91SStefano Zampini PetscInt old_index; 3537dc456d91SStefano Zampini 3538dc456d91SStefano Zampini root_data[first_index] += start; 3539674ae819SStefano Zampini old_index = first_index; 3540dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3541dc456d91SStefano Zampini if (root_data[i]) { 3542dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3543674ae819SStefano Zampini old_index = i; 3544674ae819SStefano Zampini } 3545674ae819SStefano Zampini } 3546674ae819SStefano Zampini } 3547dc456d91SStefano Zampini 3548dc456d91SStefano Zampini /* from roots to leaves */ 3549dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3550dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3551dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3552dc456d91SStefano Zampini 3553dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3554dc456d91SStefano Zampini if (subset_mult) { 3555dc456d91SStefano Zampini const PetscInt* idxs_mult; 3556dc456d91SStefano Zampini PetscInt cum; 3557dc456d91SStefano Zampini 3558dc456d91SStefano Zampini cum = 0; 3559dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3560dc456d91SStefano Zampini for (i=0;i<n;i++) { 3561dc456d91SStefano Zampini PetscInt j; 3562dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3563674ae819SStefano Zampini } 3564dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3565674ae819SStefano Zampini } else { 3566dc456d91SStefano Zampini for (i=0;i<n;i++) { 3567dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3568674ae819SStefano Zampini } 3569674ae819SStefano Zampini } 3570dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3571dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3572674ae819SStefano Zampini PetscFunctionReturn(0); 3573674ae819SStefano Zampini } 35749a7d3425SStefano Zampini 35759a7d3425SStefano Zampini #undef __FUNCT__ 35769a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 35779a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 35789a7d3425SStefano Zampini { 35799a7d3425SStefano Zampini PetscInt i,j; 35809a7d3425SStefano Zampini PetscScalar *alphas; 35819a7d3425SStefano Zampini PetscErrorCode ierr; 35829a7d3425SStefano Zampini 35839a7d3425SStefano Zampini PetscFunctionBegin; 35849a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3585785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 35869a7d3425SStefano Zampini for (i=0;i<n;i++) { 35879a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 35889a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 35899a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 35909a7d3425SStefano Zampini } 35919a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 35929a7d3425SStefano Zampini PetscFunctionReturn(0); 35939a7d3425SStefano Zampini } 35949a7d3425SStefano Zampini 3595e7931f94SStefano Zampini #undef __FUNCT__ 359670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3597b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3598e7931f94SStefano Zampini { 359952e5ac9dSStefano Zampini IS ranks_send_to; 3600e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3601e7931f94SStefano Zampini PetscMPIInt size,rank,color; 360252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 360352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 36043837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 36052b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3606e7931f94SStefano Zampini PetscSubcomm subcomm; 360752e5ac9dSStefano Zampini PetscErrorCode ierr; 3608a57a6d2fSStefano Zampini 3609e7931f94SStefano Zampini PetscFunctionBegin; 36102b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 36112b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 36122b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3613e7931f94SStefano Zampini 3614e7931f94SStefano Zampini /* Get info on mapping */ 36153bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 36163bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3617e7931f94SStefano Zampini 3618e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3619785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3620e7931f94SStefano Zampini xadj[0] = 0; 3621e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3622785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3623785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3624e7931f94SStefano Zampini 36252b510759SStefano Zampini if (threshold) { 3626d023bfaeSStefano Zampini PetscInt xadj_count = 0; 36272b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3628d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3629d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3630d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3631d023bfaeSStefano Zampini xadj_count++; 3632e7931f94SStefano Zampini } 3633e7931f94SStefano Zampini } 3634d023bfaeSStefano Zampini xadj[1] = xadj_count; 3635c8587f34SStefano Zampini } else { 3636e7931f94SStefano Zampini if (xadj[1]) { 3637e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3638e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3639c8587f34SStefano Zampini } 3640e7931f94SStefano Zampini } 36413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3642e7931f94SStefano Zampini if (use_square) { 3643e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3644e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3645e7931f94SStefano Zampini } 3646e7931f94SStefano Zampini } 3647e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3648e7931f94SStefano Zampini 36493837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3650e7931f94SStefano Zampini 3651e7931f94SStefano Zampini /* 3652e7931f94SStefano Zampini Restrict work on active processes only. 3653e7931f94SStefano Zampini */ 3654e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3655e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3656e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 36572b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3658d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3659e7931f94SStefano Zampini if (color) { 3660e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3661e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3662e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3663c8587f34SStefano Zampini } else { 366452e5ac9dSStefano Zampini Mat subdomain_adj; 366552e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 366652e5ac9dSStefano Zampini MatPartitioning partitioner; 366752e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 366852e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3669b0c7d250SStefano Zampini PetscBool aggregate; 3670b0c7d250SStefano Zampini 3671306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3672785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3673e7931f94SStefano Zampini prank = rank; 3674306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 36758002ef2cSStefano Zampini /* 3676e7931f94SStefano Zampini for (i=0;i<size;i++) { 3677e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3678c8587f34SStefano Zampini } 36798002ef2cSStefano Zampini */ 3680e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3681e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3682c8587f34SStefano Zampini } 3683e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3684b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3685b0c7d250SStefano Zampini if (aggregate) { 3686b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3687b0c7d250SStefano Zampini PetscMPIInt nrank; 3688b0c7d250SStefano Zampini PetscScalar *vals; 3689b0c7d250SStefano Zampini 3690b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3691b0c7d250SStefano Zampini lrows = 0; 3692b0c7d250SStefano Zampini if (nrank<redprocs) { 3693b0c7d250SStefano Zampini lrows = size/redprocs; 3694b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3695b0c7d250SStefano Zampini } 36965fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3697b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3698b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3699b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3700b0c7d250SStefano Zampini row = nrank; 3701b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3702b0c7d250SStefano Zampini cols = adjncy; 3703b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3704b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3705b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3706b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3707b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 370852e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 370952e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 371052e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3711b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3712b0c7d250SStefano Zampini } else { 3713306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3714b0c7d250SStefano Zampini } 371522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3716e7931f94SStefano Zampini 3717e7931f94SStefano Zampini /* Partition */ 3718306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3719e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3720e7931f94SStefano Zampini if (use_vwgt) { 37213837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3722e7931f94SStefano Zampini v_wgt[0] = local_size; 3723e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3724c8587f34SStefano Zampini } 372528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 372628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3727e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3728e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 372922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3730e7931f94SStefano Zampini 373152e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 373252e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 373352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 373452e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3735b0c7d250SStefano Zampini if (!redprocs) { 3736b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 373728143c3dSStefano Zampini } else { 3738b0c7d250SStefano Zampini PetscInt idxs[1]; 3739b0c7d250SStefano Zampini PetscMPIInt tag; 3740b0c7d250SStefano Zampini MPI_Request *reqs; 3741b0c7d250SStefano Zampini 3742b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3743b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3744b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3745b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 374628143c3dSStefano Zampini } 3747b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3748b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3749b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3750b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3751e7931f94SStefano Zampini } 375252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3753e7931f94SStefano Zampini /* clean up */ 3754e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 375552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3756e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3757e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3758e7931f94SStefano Zampini } 3759e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3760e7931f94SStefano Zampini 3761e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3762e7931f94SStefano Zampini i = 1; 3763e7931f94SStefano Zampini if (color) i=0; 3764e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3765e7931f94SStefano Zampini /* get back IS */ 3766e7931f94SStefano Zampini *is_sends = ranks_send_to; 3767e7931f94SStefano Zampini PetscFunctionReturn(0); 3768e7931f94SStefano Zampini } 3769e7931f94SStefano Zampini 3770e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3771e7931f94SStefano Zampini 3772e7931f94SStefano Zampini #undef __FUNCT__ 3773e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 377453a05cb3SStefano 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[]) 3775e7931f94SStefano Zampini { 377670cf5478SStefano Zampini Mat local_mat; 3777e7931f94SStefano Zampini IS is_sends_internal; 37789d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 377928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 37809d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3781e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3782e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3783e7931f94SStefano Zampini const PetscInt* is_indices; 3784e7931f94SStefano Zampini MatType new_local_type; 3785e7931f94SStefano Zampini /* buffers */ 3786e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 378728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 37889d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3789e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3790e7931f94SStefano Zampini /* MPI */ 379128143c3dSStefano Zampini MPI_Comm comm,comm_n; 379228143c3dSStefano Zampini PetscSubcomm subcomm; 3793e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 379428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 379528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 379628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 379728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 379828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3799e7931f94SStefano Zampini PetscErrorCode ierr; 3800e7931f94SStefano Zampini 3801e7931f94SStefano Zampini PetscFunctionBegin; 380228143c3dSStefano Zampini /* TODO: add missing checks */ 380328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 380428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 380528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 380628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3807e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 380828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3809e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3810e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3811e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3812e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3813e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 381428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 381570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 381670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 381728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 381870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 381970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 382070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 382170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 382270cf5478SStefano Zampini } 3823e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3824e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3825e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3826e7931f94SStefano Zampini if (!is_sends) { 382728143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3828b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3829c8587f34SStefano Zampini } else { 3830e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3831e7931f94SStefano Zampini is_sends_internal = is_sends; 3832c8587f34SStefano Zampini } 3833e7931f94SStefano Zampini 3834e7931f94SStefano Zampini /* get comm */ 3835a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3836e7931f94SStefano Zampini 3837e7931f94SStefano Zampini /* compute number of sends */ 3838e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3839e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3840e7931f94SStefano Zampini 3841e7931f94SStefano Zampini /* compute number of receives */ 3842e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3843785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3844e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3845e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3846e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3847e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3848e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3849e7931f94SStefano Zampini 385028143c3dSStefano Zampini /* restrict comm if requested */ 385128143c3dSStefano Zampini subcomm = 0; 385228143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 385328143c3dSStefano Zampini if (restrict_comm) { 3854779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3855779c1cceSStefano Zampini 385628143c3dSStefano Zampini color = 0; 385753a05cb3SStefano Zampini if (restrict_full) { 385853a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 385953a05cb3SStefano Zampini } else { 386053a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 386153a05cb3SStefano Zampini } 386228143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 386328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 386428143c3dSStefano Zampini /* check if reuse has been requested */ 386528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 386628143c3dSStefano Zampini if (*mat_n) { 386728143c3dSStefano Zampini PetscMPIInt subcommsize2; 386828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 386928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 387028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 387128143c3dSStefano Zampini } else { 387228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 387328143c3dSStefano Zampini } 387428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3875779c1cceSStefano Zampini PetscMPIInt rank; 3876779c1cceSStefano Zampini 3877779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 387828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 387928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 388028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3881306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 388228143c3dSStefano Zampini } 388328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 388428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 388528143c3dSStefano Zampini } else { 388628143c3dSStefano Zampini comm_n = comm; 388728143c3dSStefano Zampini } 388828143c3dSStefano Zampini 3889e7931f94SStefano Zampini /* prepare send/receive buffers */ 3890785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3891e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3892785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3893e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 389428143c3dSStefano Zampini if (nis) { 3895854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 389628143c3dSStefano Zampini } 3897e7931f94SStefano Zampini 389828143c3dSStefano Zampini /* Get data from local matrices */ 3899e7931f94SStefano Zampini if (!isdense) { 3900a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3901e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3902e7931f94SStefano Zampini /* 3903e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3904e7931f94SStefano Zampini send_buffer_idxs should contain: 3905e7931f94SStefano Zampini - MatType_PRIVATE type 3906e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3907e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3908e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3909e7931f94SStefano Zampini */ 3910e7931f94SStefano Zampini } else { 3911e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 39123bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3913854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3914e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3915e7931f94SStefano Zampini send_buffer_idxs[1] = i; 39163bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3917e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 39183bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3919e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3920e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3921e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3922e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3923c8587f34SStefano Zampini } 3924c8587f34SStefano Zampini } 3925e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 392628143c3dSStefano Zampini /* additional is (if any) */ 392728143c3dSStefano Zampini if (nis) { 392828143c3dSStefano Zampini PetscMPIInt psum; 392928143c3dSStefano Zampini PetscInt j; 393028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 393128143c3dSStefano Zampini PetscInt plen; 393228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 393328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 393428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 393528143c3dSStefano Zampini } 3936854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 393728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 393828143c3dSStefano Zampini PetscInt plen; 393928143c3dSStefano Zampini const PetscInt *is_array_idxs; 394028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 394128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 394228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 394328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 394428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 394528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 394628143c3dSStefano Zampini } 394728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 394828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 394928143c3dSStefano Zampini } 395028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 395128143c3dSStefano Zampini } 395228143c3dSStefano Zampini 3953e7931f94SStefano Zampini buf_size_idxs = 0; 3954e7931f94SStefano Zampini buf_size_vals = 0; 395528143c3dSStefano Zampini buf_size_idxs_is = 0; 3956e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3957e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3958e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 395928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3960e7931f94SStefano Zampini } 3961785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3962785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 396395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3964e7931f94SStefano Zampini 3965e7931f94SStefano Zampini /* get new tags for clean communications */ 3966e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3967e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 396828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3969e7931f94SStefano Zampini 3970e7931f94SStefano Zampini /* allocate for requests */ 3971785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3972785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 397395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3974785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3975785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 397695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3977e7931f94SStefano Zampini 3978e7931f94SStefano Zampini /* communications */ 3979e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3980e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 398128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3982e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3983e7931f94SStefano Zampini source_dest = onodes[i]; 3984e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3985e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3986e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3987e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 398828143c3dSStefano Zampini if (nis) { 398928143c3dSStefano 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); 399028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 399128143c3dSStefano Zampini } 3992e7931f94SStefano Zampini } 3993e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3994e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3995e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3996e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 399728143c3dSStefano Zampini if (nis) { 399828143c3dSStefano 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); 399928143c3dSStefano Zampini } 4000e7931f94SStefano Zampini } 4001e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4002e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4003e7931f94SStefano Zampini 4004e7931f94SStefano Zampini /* assemble new l2g map */ 4005e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4006e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 40079d30be91SStefano Zampini new_local_rows = 0; 4008e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 40099d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4010e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4011e7931f94SStefano Zampini } 40129d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4013e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 40149d30be91SStefano Zampini new_local_rows = 0; 4015e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 40169d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 40179d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4018e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4019e7931f94SStefano Zampini } 40209d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 40219d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4022e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4023e7931f94SStefano Zampini 4024e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4025e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4026e7931f94SStefano 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) */ 4027e7931f94SStefano Zampini if (n_recvs) { 402828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4029e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4030e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4031e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4032e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4033e7931f94SStefano Zampini break; 4034e7931f94SStefano Zampini } 4035e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4036e7931f94SStefano Zampini } 4037e7931f94SStefano Zampini switch (new_local_type_private) { 403828143c3dSStefano Zampini case MATDENSE_PRIVATE: 403928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4040e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4041e7931f94SStefano Zampini bs = 1; 404228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 404328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 404428143c3dSStefano Zampini bs = 1; 404528143c3dSStefano Zampini } 4046e7931f94SStefano Zampini break; 4047e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4048e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4049e7931f94SStefano Zampini bs = 1; 4050e7931f94SStefano Zampini break; 4051e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4052e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4053e7931f94SStefano Zampini break; 4054e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4055e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4056e7931f94SStefano Zampini break; 4057e7931f94SStefano Zampini default: 40589d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4059e7931f94SStefano Zampini break; 4060e7931f94SStefano Zampini } 406128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 406228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 406328143c3dSStefano Zampini bs = 1; 4064e7931f94SStefano Zampini } 4065e7931f94SStefano Zampini 406670cf5478SStefano Zampini /* create MATIS object if needed */ 406770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4068e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4069e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 407070cf5478SStefano Zampini } else { 407170cf5478SStefano Zampini /* it also destroys the local matrices */ 407270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 407370cf5478SStefano Zampini } 407470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4075e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 40769d30be91SStefano Zampini 40779d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 40789d30be91SStefano Zampini 40799d30be91SStefano Zampini /* Global to local map of received indices */ 40809d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 40819d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 40829d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 40839d30be91SStefano Zampini 40849d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 40859d30be91SStefano Zampini buf_size_idxs = 0; 40869d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40879d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 40889d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 40899d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 40909d30be91SStefano Zampini } 40919d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 40929d30be91SStefano Zampini 40939d30be91SStefano Zampini /* set preallocation */ 40949d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 40959d30be91SStefano Zampini if (!newisdense) { 40969d30be91SStefano Zampini PetscInt *new_local_nnz=0; 40979d30be91SStefano Zampini 40989d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 40999d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 41009d30be91SStefano Zampini if (n_recvs) { 41019d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 41029d30be91SStefano Zampini } 41039d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 41049d30be91SStefano Zampini PetscInt j; 41059d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 41069d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 41079d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 41089d30be91SStefano Zampini } 41099d30be91SStefano Zampini } else { 41109d30be91SStefano Zampini /* TODO */ 41119d30be91SStefano Zampini } 41129d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 41139d30be91SStefano Zampini } 41149d30be91SStefano Zampini if (new_local_nnz) { 41159d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 41169d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 41179d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 41189d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 41199d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 41209d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 41219d30be91SStefano Zampini } else { 41229d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 41239d30be91SStefano Zampini } 41249d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 41259d30be91SStefano Zampini } else { 41269d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 41279d30be91SStefano Zampini } 4128e7931f94SStefano Zampini 4129e7931f94SStefano Zampini /* set values */ 4130e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 41319d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4132e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4133e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4134e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 41359d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4136e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4137e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4138e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 413928143c3dSStefano Zampini } else { 414028143c3dSStefano Zampini /* TODO */ 4141e7931f94SStefano Zampini } 4142e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4143e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4144e7931f94SStefano Zampini } 4145e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4146e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 414770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 414870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41499d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 41509d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4151e7931f94SStefano Zampini 4152dfd14d43SStefano Zampini #if 0 415328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4154e7931f94SStefano Zampini Vec lvec,rvec; 4155e7931f94SStefano Zampini PetscReal infty_error; 4156e7931f94SStefano Zampini 41572a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4158e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4159e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4160e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 416170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4162e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4163e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4164e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4165e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4166e7931f94SStefano Zampini } 416728143c3dSStefano Zampini #endif 4168e7931f94SStefano Zampini 416928143c3dSStefano Zampini /* assemble new additional is (if any) */ 417028143c3dSStefano Zampini if (nis) { 417128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 417228143c3dSStefano Zampini 417328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4174854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 417528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 417628143c3dSStefano Zampini psum = 0; 417728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 417828143c3dSStefano Zampini for (j=0;j<nis;j++) { 417928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 418028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 418128143c3dSStefano Zampini psum += plen; 418228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 418328143c3dSStefano Zampini } 418428143c3dSStefano Zampini } 4185854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4186854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 418728143c3dSStefano Zampini for (i=1;i<nis;i++) { 418828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 418928143c3dSStefano Zampini } 419028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 419128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 419228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 419328143c3dSStefano Zampini for (j=0;j<nis;j++) { 419428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 419528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 419628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 419728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 419828143c3dSStefano Zampini } 419928143c3dSStefano Zampini } 420028143c3dSStefano Zampini for (i=0;i<nis;i++) { 420128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 420228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 420328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 420428143c3dSStefano Zampini } 420528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 420628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 420728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 420828143c3dSStefano Zampini } 4209e7931f94SStefano Zampini /* free workspace */ 421028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4211e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4212e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4213e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4214e7931f94SStefano Zampini if (isdense) { 4215e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4216e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4217e7931f94SStefano Zampini } else { 4218e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4219e7931f94SStefano Zampini } 422028143c3dSStefano Zampini if (nis) { 422128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 422228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 422328143c3dSStefano Zampini } 4224e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4225e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 422628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4227e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4228e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 422928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4230e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4231e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4232e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4233e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4234e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 423528143c3dSStefano Zampini if (nis) { 423628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 423728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 423828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 423928143c3dSStefano Zampini } 424028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 424128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 424228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 424328143c3dSStefano Zampini for (i=0;i<nis;i++) { 424428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 424528143c3dSStefano Zampini } 424653a05cb3SStefano Zampini *mat_n = NULL; 424728143c3dSStefano Zampini } 4248e7931f94SStefano Zampini PetscFunctionReturn(0); 4249e7931f94SStefano Zampini } 4250a57a6d2fSStefano Zampini 425112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4252af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 425312edc857SStefano Zampini 4254c8587f34SStefano Zampini #undef __FUNCT__ 4255c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4256c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4257c8587f34SStefano Zampini { 4258c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4259c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 426020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 42619881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 426220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 42636e683305SStefano Zampini IS coarse_is,*isarray; 42646e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 42656e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4266f9eb5b7dSStefano Zampini PC pc_temp; 4267c8587f34SStefano Zampini PCType coarse_pc_type; 4268c8587f34SStefano Zampini KSPType coarse_ksp_type; 4269f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 42704f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 42716e683305SStefano Zampini Mat t_coarse_mat_is; 42726e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 42736e683305SStefano Zampini PetscMPIInt all_procs; 427474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 427568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 427622bc73bbSStefano Zampini PetscScalar *array; 42779881197aSStefano Zampini PetscErrorCode ierr; 4278fdc09c96SStefano Zampini 4279c8587f34SStefano Zampini PetscFunctionBegin; 4280c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 428168457ee5SStefano 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 */ 4282fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 42835a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4284fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4285f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4286f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4287f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4288fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 428951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 429051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4291dc4bcba2SStefano Zampini PC pc; 4292dc4bcba2SStefano Zampini PetscBool isbddc; 4293dc4bcba2SStefano Zampini 4294dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4295dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4296dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4297dc4bcba2SStefano Zampini if (isbddc) { 4298dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4299dc4bcba2SStefano Zampini } 4300727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4301fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4302fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4303fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4304f4ddd8eeSStefano Zampini } 4305fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4306fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4307f4ddd8eeSStefano Zampini } 430870cf5478SStefano Zampini /* reset any subassembling information */ 430970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 43106e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 43116e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4312fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4313f4ddd8eeSStefano Zampini } 4314c8587f34SStefano Zampini 43156e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 43162b510759SStefano Zampini im_active = !!(pcis->n); 43172b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43186e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 43196e683305SStefano Zampini void_procs = all_procs-active_procs; 43206e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 432174e2c79eSStefano Zampini redist = PETSC_FALSE; 432222bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 43236e683305SStefano Zampini csin_ml = PETSC_TRUE; 43246e683305SStefano Zampini ncoarse_ml = void_procs; 4325779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4326779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 43276e683305SStefano Zampini csin_ds = PETSC_TRUE; 432818a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 432918a45a71SStefano Zampini redist = PETSC_TRUE; 433018a45a71SStefano Zampini } else { 43316e683305SStefano Zampini csin_ds = PETSC_TRUE; 4332779c1cceSStefano Zampini ncoarse_ds = active_procs; 4333779c1cceSStefano Zampini redist = PETSC_TRUE; 433418a45a71SStefano Zampini } 43356e683305SStefano Zampini } else { 43366e683305SStefano Zampini csin_ml = PETSC_FALSE; 43376e683305SStefano Zampini ncoarse_ml = all_procs; 43386e683305SStefano Zampini if (void_procs) { 43396e683305SStefano Zampini csin_ds = PETSC_TRUE; 43406e683305SStefano Zampini ncoarse_ds = void_procs; 43416e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 43426e683305SStefano Zampini } else { 4343779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 434474e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 434574e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 434674e2c79eSStefano Zampini redist = PETSC_TRUE; 434774e2c79eSStefano Zampini } else { 43486e683305SStefano Zampini csin_ds = PETSC_FALSE; 43496e683305SStefano Zampini ncoarse_ds = all_procs; 43506e683305SStefano Zampini } 43516e683305SStefano Zampini } 435274e2c79eSStefano Zampini } 43536e683305SStefano Zampini 43546e683305SStefano Zampini /* 43556e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 43566e683305SStefano Zampini - we have not exceeded the number of levels requested 43576e683305SStefano Zampini - we can actually subassemble the active processes 43586e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 43596e683305SStefano Zampini */ 43606e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 43616e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 43626e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 43636e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 43646e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4365f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 43662b510759SStefano Zampini } else { 4367f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4368c8587f34SStefano Zampini } 4369c8587f34SStefano Zampini } 43706e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 43716e683305SStefano Zampini if (multilevel_allowed) { 43726e683305SStefano Zampini ncoarse = ncoarse_ml; 43736e683305SStefano Zampini csin = csin_ml; 437458da7f69SStefano Zampini redist = PETSC_FALSE; 43756e683305SStefano Zampini } else { 43766e683305SStefano Zampini ncoarse = ncoarse_ds; 43776e683305SStefano Zampini csin = csin_ds; 43786e683305SStefano Zampini } 4379e7931f94SStefano Zampini 4380abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4381abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4382abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4383abbbba34SStefano Zampini 4384abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 438522bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 438622bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 438722bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 438822bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4389b9b85e73SStefano Zampini #if 0 4390b9b85e73SStefano Zampini { 4391b9b85e73SStefano Zampini PetscViewer viewer; 4392b9b85e73SStefano Zampini char filename[256]; 4393b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4394b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4395b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4396b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4397b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4398b9b85e73SStefano Zampini } 4399b9b85e73SStefano Zampini #endif 4400e176bc59SStefano 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); 44016e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 44026e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 44036e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4404abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4405abbbba34SStefano Zampini 44066e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 44076e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 44086e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 44096e683305SStefano Zampini const PetscInt *idxs; 44106e683305SStefano Zampini ISLocalToGlobalMapping tmap; 44116e683305SStefano Zampini 44126e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 44130be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 44146e683305SStefano Zampini /* allocate space for temporary storage */ 4415854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4416854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 44176e683305SStefano Zampini /* allocate for IS array */ 44186e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 44196e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 44206e683305SStefano Zampini nis = nisdofs + nisneu; 4421854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 44226e683305SStefano Zampini /* dofs splitting */ 44236e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 44246e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 44256e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 44266e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 44276e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 44286e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 44296e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 44306e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 44316e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 44326e683305SStefano Zampini } 44336e683305SStefano Zampini /* neumann boundaries */ 44346e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 44356e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 44366e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 44376e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 44386e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 44396e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 44406e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 44416e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 44426e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 44436e683305SStefano Zampini } 44446e683305SStefano Zampini /* free memory */ 44456e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 44466e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 44476e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 44486e683305SStefano Zampini } else { 44496e683305SStefano Zampini nis = 0; 44506e683305SStefano Zampini nisdofs = 0; 44516e683305SStefano Zampini nisneu = 0; 44526e683305SStefano Zampini isarray = NULL; 44536e683305SStefano Zampini } 44546e683305SStefano Zampini /* destroy no longer needed map */ 44556e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 44566e683305SStefano Zampini 44576e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 44586e683305SStefano Zampini coarse_mat_is = NULL; 44596e683305SStefano Zampini if (csin) { 44606e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 446174e2c79eSStefano Zampini if (redist) { 446274e2c79eSStefano Zampini PetscMPIInt rank; 4463779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 446474e2c79eSStefano Zampini 446574e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 446658da7f69SStefano Zampini spc = active_procs/ncoarse; 446758da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4468779c1cceSStefano Zampini if (im_active) { 4469779c1cceSStefano Zampini destsize = 1; 447074e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 447174e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 447274e2c79eSStefano Zampini } else { 447374e2c79eSStefano Zampini dest[0] = rank/(spc+1); 447474e2c79eSStefano Zampini } 447574e2c79eSStefano Zampini } else { 4476779c1cceSStefano Zampini destsize = 0; 44776e683305SStefano Zampini } 4478779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4479779c1cceSStefano Zampini } else if (csin_type_simple) { 44806e683305SStefano Zampini PetscMPIInt rank; 44816e683305SStefano Zampini PetscInt issize,isidx; 4482779c1cceSStefano Zampini 44836e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 44846e683305SStefano Zampini if (im_active) { 44856e683305SStefano Zampini issize = 1; 44866e683305SStefano Zampini isidx = (PetscInt)rank; 44876e683305SStefano Zampini } else { 44886e683305SStefano Zampini issize = 0; 44896e683305SStefano Zampini isidx = -1; 44906e683305SStefano Zampini } 44916e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4492779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4493b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 44946e683305SStefano Zampini } 4495779c1cceSStefano Zampini 4496779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4497779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4498779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4499779c1cceSStefano Zampini PetscInt *coarse_candidates; 4500779c1cceSStefano Zampini const PetscInt* tisindices; 4501779c1cceSStefano Zampini 4502779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4503779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4504779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4505779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4506779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4507779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4508779c1cceSStefano Zampini } 4509779c1cceSStefano Zampini } 4510779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4511779c1cceSStefano Zampini 4512779c1cceSStefano Zampini 45136e683305SStefano Zampini if (pcbddc->dbg_flag) { 45146e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45156e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 45166e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 45176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4518779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 45196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 45206e683305SStefano Zampini } 45216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 45226e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45236e683305SStefano Zampini } 45246e683305SStefano Zampini /* shift the pattern on coarse candidates */ 45256e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 45266e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4527854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 45286e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 45296e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 45306e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 45316e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 45326e683305SStefano Zampini } 45336e683305SStefano Zampini if (pcbddc->dbg_flag) { 45346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 45366e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 45376e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45386e683305SStefano Zampini } 4539779c1cceSStefano Zampini } 45406e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 454153a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 454253a05cb3SStefano 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); 454353a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 454453a05cb3SStefano 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); 454553a05cb3SStefano Zampini } 45466e683305SStefano Zampini } else { 45476e683305SStefano Zampini if (pcbddc->dbg_flag) { 45486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 45506e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45516e683305SStefano Zampini } 45526e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 45536e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 45546e683305SStefano Zampini } 45556e683305SStefano Zampini 45566e683305SStefano Zampini /* create local to global scatters for coarse problem */ 455768457ee5SStefano Zampini if (compute_vecs) { 45586e683305SStefano Zampini PetscInt lrows; 45596e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 45606e683305SStefano Zampini if (coarse_mat_is) { 45616e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 45626e683305SStefano Zampini } else { 45636e683305SStefano Zampini lrows = 0; 45646e683305SStefano Zampini } 45656e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 45666e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 45676e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 45686e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45696e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45706e683305SStefano Zampini } 45716e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 45726e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4573c8587f34SStefano Zampini 4574f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4575f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4576f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4577f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4578f9eb5b7dSStefano Zampini } else { 4579f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4580f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4581c8587f34SStefano Zampini } 4582c8587f34SStefano Zampini 45836e683305SStefano Zampini /* print some info if requested */ 45846e683305SStefano Zampini if (pcbddc->dbg_flag) { 45856e683305SStefano Zampini if (!multilevel_allowed) { 45866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45876e683305SStefano Zampini if (multilevel_requested) { 45886e683305SStefano 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); 45896e683305SStefano Zampini } else if (pcbddc->max_levels) { 45906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 45916e683305SStefano Zampini } 45926e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45936e683305SStefano Zampini } 45946e683305SStefano Zampini } 45956e683305SStefano Zampini 4596f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 45976e683305SStefano Zampini if (coarse_mat_is) { 45986e683305SStefano Zampini MatReuse coarse_mat_reuse; 45996a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 46006e683305SStefano Zampini if (pcbddc->dbg_flag) { 46016e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 46026e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 46036e683305SStefano Zampini } 4604f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4605312be037SStefano Zampini char prefix[256],str_level[16]; 4606e604994aSStefano Zampini size_t len; 46076e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4608422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4609c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4610f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 46115f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4612c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 46136e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4614c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4615c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4616e604994aSStefano Zampini /* prefix */ 4617e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4618e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4619e604994aSStefano Zampini if (!pcbddc->current_level) { 4620e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4621e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4622c8587f34SStefano Zampini } else { 4623e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4624312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4625312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 462634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4627312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4628e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4629e604994aSStefano Zampini } 4630e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 46313e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 46323e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 46333e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 46343e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4635f9eb5b7dSStefano Zampini /* allow user customization */ 4636f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 46373e3c6dadSStefano Zampini } 46383e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 463951bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 46403e3c6dadSStefano Zampini if (nisdofs) { 46413e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 46423e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 46433e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 46443e3c6dadSStefano Zampini } 46453e3c6dadSStefano Zampini } 46463e3c6dadSStefano Zampini if (nisneu) { 46473e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 46483e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4649312be037SStefano Zampini } 4650f9eb5b7dSStefano Zampini 4651f9eb5b7dSStefano Zampini /* get some info after set from options */ 4652f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4653f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 46544f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 46556e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4656f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4657f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4658f9eb5b7dSStefano Zampini } 465939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 46604f3a063dSStefano Zampini if (isredundant) { 46614f3a063dSStefano Zampini KSP inner_ksp; 46624f3a063dSStefano Zampini PC inner_pc; 46634f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 46644f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 46654f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 46664f3a063dSStefano Zampini } 4667f9eb5b7dSStefano Zampini 4668f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4669fa7f1dd8SStefano Zampini if (coarse_reuse) { 467081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4671fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 46726e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4673fa7f1dd8SStefano Zampini } else { 46746e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4675fa7f1dd8SStefano Zampini } 4676c8587f34SStefano Zampini if (isbddc || isnn) { 467722bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 467870cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4679b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 468022b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 46816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 46836e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 46846e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 468522b6e8a2SStefano Zampini } 468670cf5478SStefano Zampini } 468753a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 468870cf5478SStefano Zampini } else { 468922bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 469022bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 469122bc73bbSStefano Zampini } 469222bc73bbSStefano Zampini } else { 46932e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4694c8587f34SStefano Zampini } 4695c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4696c8587f34SStefano Zampini 46973301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 46985a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 46993301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 47003301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 47013301b35fSStefano Zampini } 47023301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 47033301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 47043301b35fSStefano Zampini } 47053301b35fSStefano Zampini if (pc->pmat->spd_set) { 47063301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 47073301b35fSStefano Zampini } 47086e683305SStefano Zampini /* set operators */ 47095f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 47106e683305SStefano Zampini if (pcbddc->dbg_flag) { 47116e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 47126e683305SStefano Zampini } 47136e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 47146e683305SStefano Zampini coarse_mat = 0; 47156e683305SStefano Zampini } 47166e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4717b9b85e73SStefano Zampini #if 0 4718b9b85e73SStefano Zampini { 4719b9b85e73SStefano Zampini PetscViewer viewer; 4720b9b85e73SStefano Zampini char filename[256]; 4721b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4722b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4723b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4724b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4725b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4726b9b85e73SStefano Zampini } 4727b9b85e73SStefano Zampini #endif 4728c8587f34SStefano Zampini 4729c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4730298c0119SStefano Zampini #if 0 4731c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4732c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 473398a51de6SStefano Zampini } 4734298c0119SStefano Zampini #endif 473598a51de6SStefano Zampini 473698a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 473798a51de6SStefano Zampini Vec crhs,csol; 473898a51de6SStefano Zampini PetscBool ispreonly; 473904708bb6SStefano Zampini 474098a51de6SStefano Zampini if (CoarseNullSpace) { 4741c8587f34SStefano Zampini if (isbddc) { 4742c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4743c8587f34SStefano Zampini } else { 47445fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4745c8587f34SStefano Zampini } 4746c8587f34SStefano Zampini } 4747f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4748f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4749f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4750f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 47516e683305SStefano Zampini /* hack */ 4752f347579bSStefano Zampini if (!csol) { 47532a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4754f9eb5b7dSStefano Zampini } 4755f347579bSStefano Zampini if (!crhs) { 47562a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4757f347579bSStefano Zampini } 4758cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4759cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 47606e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4761c8587f34SStefano Zampini KSP check_ksp; 47622b510759SStefano Zampini KSPType check_ksp_type; 4763c8587f34SStefano Zampini PC check_pc; 47646e683305SStefano Zampini Vec check_vec,coarse_vec; 47656a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 47662b510759SStefano Zampini PetscInt its; 47676e683305SStefano Zampini PetscBool compute_eigs; 47686e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 47696e683305SStefano Zampini PetscInt neigs; 47708e185a42SStefano Zampini const char *prefix; 4771c8587f34SStefano Zampini 47722b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 47736e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4774422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 477523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4776f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 47772b510759SStefano Zampini if (ispreonly) { 47782b510759SStefano Zampini check_ksp_type = KSPPREONLY; 47796e683305SStefano Zampini compute_eigs = PETSC_FALSE; 47802b510759SStefano Zampini } else { 4781cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 47826e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4783c8587f34SStefano Zampini } 4784c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 47856e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 47866e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 47876e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4788a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4789a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4790a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4791a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4792c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4793c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4794c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4795c8587f34SStefano Zampini /* create random vec */ 47966e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 47976e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4798c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4799c8587f34SStefano Zampini if (CoarseNullSpace) { 4800c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4801c8587f34SStefano Zampini } 48026e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4803c8587f34SStefano Zampini /* solve coarse problem */ 48046e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4805c8587f34SStefano Zampini if (CoarseNullSpace) { 48066e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4807c8587f34SStefano Zampini } 4808cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 48096e683305SStefano Zampini if (compute_eigs) { 4810854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4811854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 48126e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 48136e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 48146e683305SStefano Zampini lambda_min = eigs_r[0]; 48156e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 48166e683305SStefano Zampini if (lambda_max>lambda_min) { 4817cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4818cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4819cbcc2c2aSStefano Zampini } 4820c8587f34SStefano Zampini } 4821c8587f34SStefano Zampini } 4822cbcc2c2aSStefano Zampini 4823c8587f34SStefano Zampini /* check coarse problem residual error */ 48246e683305SStefano Zampini if (pcbddc->dbg_flag) { 48256e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 48266e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48276e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4828c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 48296e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 48306e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4831c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4832779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 48336e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 48346e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 48356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 48366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 48376e683305SStefano Zampini if (compute_eigs) { 48386e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4839deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4840c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 48416e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 48426e683305SStefano 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); 48436e683305SStefano Zampini for (i=0;i<neigs;i++) { 48446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4845c8587f34SStefano Zampini } 48466e683305SStefano Zampini } 48476e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 48486e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48496e683305SStefano Zampini } 4850c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 48516e683305SStefano Zampini if (compute_eigs) { 48526e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 48536e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4854c8587f34SStefano Zampini } 48556e683305SStefano Zampini } 48566e683305SStefano Zampini } 4857cbcc2c2aSStefano Zampini /* print additional info */ 4858cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 48596e683305SStefano Zampini /* waits until all processes reaches this point */ 48606e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4861cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4862cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4863cbcc2c2aSStefano Zampini } 4864cbcc2c2aSStefano Zampini 48652b510759SStefano Zampini /* free memory */ 4866c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4867fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4868c8587f34SStefano Zampini PetscFunctionReturn(0); 4869c8587f34SStefano Zampini } 4870674ae819SStefano Zampini 4871f34684f1SStefano Zampini #undef __FUNCT__ 4872f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4873f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4874f34684f1SStefano Zampini { 4875f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4876f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4877f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4878dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4879dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 488073be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4881dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4882f34684f1SStefano Zampini PetscErrorCode ierr; 4883f34684f1SStefano Zampini 4884f34684f1SStefano Zampini PetscFunctionBegin; 4885f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 48860e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 48870e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4888727cdba6SStefano Zampini } 4889dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 48903bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4891dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4892dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4893dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4894dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4895dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4896dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 48970e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 48980e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 48990e6343abSStefano Zampini } 4900dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4901dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4902dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4903dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4904dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4905f34684f1SStefano Zampini 4906f34684f1SStefano Zampini /* check numbering */ 4907f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4908019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 4909dc456d91SStefano Zampini PetscInt i; 4910b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4911f34684f1SStefano Zampini 4912f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4913f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4914f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 49150fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4916019a44ceSStefano Zampini /* counter */ 4917019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4918019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4919019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4920019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4921019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4922019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4923019a44ceSStefano Zampini 4924f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4925f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4926727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4927f34684f1SStefano Zampini } 4928f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4929f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4930f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4931e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4932e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4933e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4934e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4935f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4936019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4937f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4938019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 4939019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 4940019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 4941b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4942019a44ceSStefano 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); 4943f34684f1SStefano Zampini } 4944f34684f1SStefano Zampini } 4945019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4946b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4947f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4948f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4949f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4950f34684f1SStefano Zampini } 4951f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4952f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4953e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4954e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4955f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4956f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4957b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4958ca8b9ea9SStefano Zampini PetscInt *gidxs; 4959ca8b9ea9SStefano Zampini 4960ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 49613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4962f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4963f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4964f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4965f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 49664bc2dc4bSStefano 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); 4967f34684f1SStefano Zampini } 4968f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4969ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4970f34684f1SStefano Zampini } 4971f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4972302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4973f34684f1SStefano Zampini } 49748bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4975f34684f1SStefano Zampini /* get back data */ 4976f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4977f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4978674ae819SStefano Zampini PetscFunctionReturn(0); 4979674ae819SStefano Zampini } 4980674ae819SStefano Zampini 4981e456f2a8SStefano Zampini #undef __FUNCT__ 4982e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4983a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4984e456f2a8SStefano Zampini { 4985e456f2a8SStefano Zampini IS localis_t; 4986a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4987e456f2a8SStefano Zampini PetscScalar *vals; 4988e456f2a8SStefano Zampini PetscErrorCode ierr; 4989e456f2a8SStefano Zampini 4990e456f2a8SStefano Zampini PetscFunctionBegin; 4991a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4992e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4993854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4994e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4995e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4996a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4997a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 49981035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4999a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 50001035eff8SStefano Zampini } 5001a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5002e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5003e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5004a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5005a7dc3881SStefano Zampini /* now compute set in local ordering */ 5006a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5007a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5008a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5009a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5010a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5011ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5012e456f2a8SStefano Zampini lsize++; 5013e456f2a8SStefano Zampini } 5014e456f2a8SStefano Zampini } 5015854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5016a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5017ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5018e456f2a8SStefano Zampini idxs[lsize++] = i; 5019e456f2a8SStefano Zampini } 5020e456f2a8SStefano Zampini } 5021a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5022a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5023e456f2a8SStefano Zampini *localis = localis_t; 5024e456f2a8SStefano Zampini PetscFunctionReturn(0); 5025e456f2a8SStefano Zampini } 5026906d46d4SStefano Zampini 5027906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5028906d46d4SStefano Zampini #undef __FUNCT__ 5029906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5030906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5031906d46d4SStefano Zampini { 5032906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5033906d46d4SStefano Zampini PetscErrorCode ierr; 5034906d46d4SStefano Zampini 5035906d46d4SStefano Zampini PetscFunctionBegin; 5036906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5037906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5038906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5039906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5040906d46d4SStefano Zampini PetscFunctionReturn(0); 5041906d46d4SStefano Zampini } 5042906d46d4SStefano Zampini 5043906d46d4SStefano Zampini #undef __FUNCT__ 5044906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5045906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5046906d46d4SStefano Zampini { 5047906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5048906d46d4SStefano Zampini PetscErrorCode ierr; 5049906d46d4SStefano Zampini 5050906d46d4SStefano Zampini PetscFunctionBegin; 5051906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5052906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5053906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5054906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5055906d46d4SStefano Zampini PetscFunctionReturn(0); 5056906d46d4SStefano Zampini } 5057b96c3477SStefano Zampini 5058b96c3477SStefano Zampini #undef __FUNCT__ 5059b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 506008122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5061b96c3477SStefano Zampini { 5062a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5063b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5064b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5065a64f4aa4SStefano Zampini Mat S_j; 5066b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5067b96c3477SStefano Zampini PetscBool free_used_adj; 5068b96c3477SStefano Zampini PetscErrorCode ierr; 5069b96c3477SStefano Zampini 5070b96c3477SStefano Zampini PetscFunctionBegin; 5071b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5072b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 507308122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5074b96c3477SStefano Zampini used_xadj = NULL; 5075b96c3477SStefano Zampini used_adjncy = NULL; 5076b96c3477SStefano Zampini } else { 507708122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 507808122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 507908122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 508008122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5081b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5082b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5083b96c3477SStefano Zampini } else { 50842fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5085b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5086b96c3477SStefano Zampini PetscInt nvtxs; 5087b96c3477SStefano Zampini 50882fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 50892fffb893SStefano Zampini if (flg_row) { 5090b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5091b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5092b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5093b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 50942fffb893SStefano Zampini } else { 50952fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 50962fffb893SStefano Zampini used_xadj = NULL; 50972fffb893SStefano Zampini used_adjncy = NULL; 50982fffb893SStefano Zampini } 50992fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5100b96c3477SStefano Zampini } 5101b96c3477SStefano Zampini } 5102d5574798SStefano Zampini 5103d5574798SStefano Zampini /* setup sub_schurs data */ 5104a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5105a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5106a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5107a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 510806a4e24aSStefano 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); 5109a64f4aa4SStefano Zampini } else { 51106816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 511104708bb6SStefano Zampini PetscBool isseqaij; 51125feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 51135feab87aSStefano Zampini PetscInt n_vertices; 51145feab87aSStefano Zampini 51155feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 51162034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 51175feab87aSStefano Zampini } 511804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 511904708bb6SStefano Zampini if (!isseqaij) { 512004708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 512104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 512204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 512304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 512404708bb6SStefano Zampini } else { 512504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 512604708bb6SStefano Zampini } 512704708bb6SStefano Zampini } 512806a4e24aSStefano 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); 5129a64f4aa4SStefano Zampini } 5130a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5131b96c3477SStefano Zampini 5132b96c3477SStefano Zampini /* free adjacency */ 5133b96c3477SStefano Zampini if (free_used_adj) { 5134b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5135b96c3477SStefano Zampini } 5136b96c3477SStefano Zampini PetscFunctionReturn(0); 5137b96c3477SStefano Zampini } 5138b96c3477SStefano Zampini 5139b96c3477SStefano Zampini #undef __FUNCT__ 5140b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 514108122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5142b96c3477SStefano Zampini { 5143b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5144b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5145b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5146b96c3477SStefano Zampini PCBDDCGraph graph; 5147b96c3477SStefano Zampini PetscErrorCode ierr; 5148b96c3477SStefano Zampini 5149b96c3477SStefano Zampini PetscFunctionBegin; 5150b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 515108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 51523301b35fSStefano Zampini IS verticesIS,verticescomm; 51533301b35fSStefano Zampini PetscInt vsize,*idxs; 5154b96c3477SStefano Zampini 5155b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 51563301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 51573301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51583301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 51593301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51603301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5161b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 51627fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 51633301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 51643301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5165b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5166b96c3477SStefano Zampini /* 5167b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5168b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5169b96c3477SStefano Zampini } 5170b96c3477SStefano Zampini */ 5171b96c3477SStefano Zampini } else { 5172b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5173b96c3477SStefano Zampini } 5174b96c3477SStefano Zampini 5175b96c3477SStefano Zampini /* sub_schurs init */ 5176a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5177a64f4aa4SStefano Zampini 5178b96c3477SStefano Zampini /* free graph struct */ 517908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5180b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5181b96c3477SStefano Zampini } 5182b96c3477SStefano Zampini PetscFunctionReturn(0); 5183b96c3477SStefano Zampini } 5184