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__ 9*efc2fbd9SStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushP0" 10*efc2fbd9SStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushP0(PC pc, Vec v, PetscBool pop) 11*efc2fbd9SStefano Zampini { 12*efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13*efc2fbd9SStefano Zampini PetscErrorCode ierr; 14*efc2fbd9SStefano Zampini 15*efc2fbd9SStefano Zampini PetscFunctionBegin; 16*efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 17*efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 18*efc2fbd9SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,1,NULL,PETSC_OWN_POINTER,&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 19*efc2fbd9SStefano Zampini } 20*efc2fbd9SStefano Zampini if (pop) { /* use SF to get values */ 21*efc2fbd9SStefano Zampini PetscScalar *array; 22*efc2fbd9SStefano Zampini 23*efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 24*efc2fbd9SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 25*efc2fbd9SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 26*efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 27*efc2fbd9SStefano Zampini } else { /* use VecSetValue */ 28*efc2fbd9SStefano Zampini ierr = VecSetValue(v,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 29*efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 30*efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 31*efc2fbd9SStefano Zampini } 32*efc2fbd9SStefano Zampini PetscFunctionReturn(0); 33*efc2fbd9SStefano Zampini } 34*efc2fbd9SStefano Zampini 35*efc2fbd9SStefano 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 */ 47*efc2fbd9SStefano Zampini if (!pcbddc->zerodiag) { 48*efc2fbd9SStefano Zampini PetscFunctionReturn(0); 49*efc2fbd9SStefano 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 */ 56*efc2fbd9SStefano 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); 62*efc2fbd9SStefano 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); 65*efc2fbd9SStefano Zampini ierr = MatZeroRowsColumns(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,1.,NULL,NULL);CHKERRQ(ierr); 66c263805aSStefano Zampini } else { /* push */ 67*efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,pcbddc->B0_ncol,pcbddc->B0_cols,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 68*efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,pcbddc->B0_ncol,pcbddc->B0_cols,1,&pcbddc->benign_p0_lidx,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 69*efc2fbd9SStefano 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); 53681d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 537674ae819SStefano Zampini PetscFunctionReturn(0); 538674ae819SStefano Zampini } 539674ae819SStefano Zampini 540674ae819SStefano Zampini #undef __FUNCT__ 541f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 542f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 5436bfb1811SStefano Zampini { 5446bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5456bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 5466bfb1811SStefano Zampini VecType impVecType; 547019a44ceSStefano Zampini PetscInt n_constraints,n_R,old_size,n_benign; 5486bfb1811SStefano Zampini PetscErrorCode ierr; 5496bfb1811SStefano Zampini 5506bfb1811SStefano Zampini PetscFunctionBegin; 551f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 552019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 553f4ddd8eeSStefano Zampini } 554e7b262bdSStefano Zampini /* get sizes */ 555019a44ceSStefano Zampini n_benign = (PetscInt)(!!pcbddc->zerodiag); 556019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - pcbddc->n_vertices; 557019a44ceSStefano Zampini if (n_benign && n_constraints) { 558019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot handle benign susbspace trick without change of basis"); 559019a44ceSStefano Zampini } 560b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 5616bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 562e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 563e7b262bdSStefano Zampini /* R nodes */ 564e7b262bdSStefano Zampini old_size = -1; 565e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 566e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 567e7b262bdSStefano Zampini } 568e7b262bdSStefano Zampini if (n_R != old_size) { 569e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 570e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5716bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5726bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5736bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5746bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 575e7b262bdSStefano Zampini } 576e7b262bdSStefano Zampini /* local primal dofs */ 577e7b262bdSStefano Zampini old_size = -1; 578e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 579e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 580e7b262bdSStefano Zampini } 581e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 582e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 58383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 584e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5856bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 586e7b262bdSStefano Zampini } 587e7b262bdSStefano Zampini /* local explicit constraints */ 588e7b262bdSStefano Zampini old_size = -1; 589e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 590e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 591e7b262bdSStefano Zampini } 592e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 593e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 59483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 59583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 59683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 59783b7ccabSStefano Zampini } 5986bfb1811SStefano Zampini PetscFunctionReturn(0); 5996bfb1811SStefano Zampini } 6006bfb1811SStefano Zampini 6016bfb1811SStefano Zampini #undef __FUNCT__ 60247f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 60347f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 60488ebb749SStefano Zampini { 60525084f0cSStefano Zampini PetscErrorCode ierr; 60625084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 60788ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 60888ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 609d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 61025084f0cSStefano Zampini /* submatrices of local problem */ 61180677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 612019a44ceSStefano Zampini /* submatrices of benign trick */ 613019a44ceSStefano Zampini Mat B0_R,B0_V; 61406656605SStefano Zampini /* submatrices of local coarse problem */ 61506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 61625084f0cSStefano Zampini /* working matrices */ 61706656605SStefano Zampini Mat C_CR; 61825084f0cSStefano Zampini /* additional working stuff */ 61906656605SStefano Zampini PC pc_R; 62006656605SStefano Zampini Mat F; 621019a44ceSStefano Zampini Vec vec_R_benign; 62206656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 62306656605SStefano Zampini 62425084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 62506656605SStefano Zampini PetscScalar *work; 62606656605SStefano Zampini PetscInt *idx_V_B; 627019a44ceSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign; 62806656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 629b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 63045a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 63188ebb749SStefano Zampini MatType impMatType; 63225084f0cSStefano Zampini /* some shortcuts to scalars */ 63306656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 63488ebb749SStefano Zampini 63588ebb749SStefano Zampini PetscFunctionBegin; 636019a44ceSStefano Zampini n_benign = (PetscInt)(!!pcbddc->zerodiag); 637b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 638019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - n_vertices; 63988ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 640b371cd4fSStefano Zampini n_B = pcis->n_B; 641b371cd4fSStefano Zampini n_D = pcis->n - n_B; 64288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 64388ebb749SStefano Zampini 64488ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 64588ebb749SStefano Zampini impMatType = MATSEQDENSE; 64688ebb749SStefano Zampini 64788ebb749SStefano Zampini /* vertices in boundary numbering */ 648785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 6490e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 65088ebb749SStefano Zampini if (i != n_vertices) { 65122d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 65288ebb749SStefano Zampini } 65388ebb749SStefano Zampini 65406656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 655019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 656019a44ceSStefano Zampini if (!n_benign) { 65706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 65806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 65906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 66006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 66106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 66206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 66306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 66406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 665019a44ceSStefano Zampini } else { /* these wrappers are not needed with the benign trick */ 666019a44ceSStefano Zampini S_VV = S_CV = S_VC = S_CC = NULL; 667019a44ceSStefano Zampini } 66806656605SStefano Zampini 66906656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 67006656605SStefano Zampini /* allocate workspace */ 67106656605SStefano Zampini n = 0; 67206656605SStefano Zampini if (n_constraints) { 67306656605SStefano Zampini n += n_R*n_constraints; 67406656605SStefano Zampini } 67506656605SStefano Zampini if (n_vertices) { 67606656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 67780677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 67806656605SStefano Zampini } 6793301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 68006656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 68106656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 68206656605SStefano Zampini } 68306656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 68406656605SStefano Zampini 68506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 68606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 68706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 68806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 68906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 69006656605SStefano Zampini if (isLU || isILU || isCHOL) { 69106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 692d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 693d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 694d62866d3SStefano Zampini MatFactorType type; 695d62866d3SStefano Zampini 6966816873aSStefano Zampini F = reuse_mumps->F; 6976816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 698d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 69906656605SStefano Zampini } else { 70006656605SStefano Zampini F = NULL; 70106656605SStefano Zampini } 70206656605SStefano Zampini 70388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 70488ebb749SStefano Zampini if (n_constraints) { 70506656605SStefano Zampini Mat M1,M2,M3; 70680677318SStefano Zampini Mat auxmat; 70706656605SStefano Zampini IS is_aux; 70880677318SStefano Zampini PetscScalar *array,*array2; 70906656605SStefano Zampini 710f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 71180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 71288ebb749SStefano Zampini 71325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 71425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 7158ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 71680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 71788ebb749SStefano Zampini 71880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 71980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 72006656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 72188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 72206656605SStefano Zampini const PetscScalar *row_cmat_values; 72306656605SStefano Zampini const PetscInt *row_cmat_indices; 72406656605SStefano Zampini PetscInt size_of_constraint,j; 72588ebb749SStefano Zampini 72606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 72706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 72806656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 72906656605SStefano Zampini } 73006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 73106656605SStefano Zampini } 73280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 73306656605SStefano Zampini if (F) { 73406656605SStefano Zampini Mat B; 73506656605SStefano Zampini 73606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 73780677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 73806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 73906656605SStefano Zampini } else { 74080677318SStefano Zampini PetscScalar *marr; 74180677318SStefano Zampini 74280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 74306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 74406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 74580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 74606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 74706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 74806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 74906656605SStefano Zampini } 75080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 75106656605SStefano Zampini } 75280677318SStefano Zampini if (!pcbddc->switch_static) { 75380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 75480677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 75580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 75680677318SStefano Zampini for (i=0;i<n_constraints;i++) { 75780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 75880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 75980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 76080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 76180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 76280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 76380677318SStefano Zampini } 76480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 76580677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 76680677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 76780677318SStefano Zampini } else { 76880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 76980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 77025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 77180677318SStefano Zampini } 77280677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 77380677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 77480677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 77506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 77606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 77780677318SStefano Zampini if (isCHOL) { 77880677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 77980677318SStefano Zampini } else { 78025084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 78180677318SStefano Zampini } 78280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 78306656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 78425084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 78525084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 78625084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 78780677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 78880677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 78980677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 79006656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 79106656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 792f4ddd8eeSStefano Zampini } 79388ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 79488ebb749SStefano Zampini if (n_vertices) { 79506656605SStefano Zampini IS is_aux; 7963a50541eSStefano Zampini 7976816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7986816873aSStefano Zampini IS tis; 7996816873aSStefano Zampini 8006816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 8016816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 8026816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 8036816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 8046816873aSStefano Zampini } else { 8053a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 8066816873aSStefano Zampini } 8079577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 8089577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 80904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 810019a44ceSStefano Zampini if (n_benign) { 811019a44ceSStefano Zampini Mat B0; 812019a44ceSStefano Zampini IS dummy; 813019a44ceSStefano Zampini PetscInt ii[2]; 814019a44ceSStefano Zampini 815019a44ceSStefano Zampini ii[0] = 0; 816019a44ceSStefano Zampini ii[1] = pcbddc->B0_ncol; 817019a44ceSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 818019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 819019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 820019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 821019a44ceSStefano Zampini ierr = MatDestroy(&B0);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 */ 88006656605SStefano Zampini /* vertices */ 88106656605SStefano Zampini if (n_vertices) { 88216f15bc4SStefano Zampini 88304708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 88404708bb6SStefano Zampini 88516f15bc4SStefano Zampini if (n_R) { 88606656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 887019a44ceSStefano Zampini Mat S_V_benign; 88806656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 88916f15bc4SStefano Zampini PetscScalar *x,*y; 89004708bb6SStefano Zampini PetscBool isseqaij; 89106656605SStefano Zampini 89221eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 89404708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 8956816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 89606656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 89706656605SStefano Zampini } else { 89806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 89906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 90006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 90106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 90206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 90306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 90506656605SStefano Zampini } 90606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 90706656605SStefano Zampini } 908019a44ceSStefano Zampini if (n_benign) { 909019a44ceSStefano Zampini const PetscScalar *vals; 910019a44ceSStefano Zampini const PetscInt *idxs; 911019a44ceSStefano Zampini PetscInt n; 912019a44ceSStefano Zampini 913019a44ceSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign);CHKERRQ(ierr); 914019a44ceSStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 915019a44ceSStefano Zampini ierr = MatGetRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 916019a44ceSStefano Zampini ierr = VecSetValues(pcbddc->vec1_R,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 917019a44ceSStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_R);CHKERRQ(ierr); 918019a44ceSStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_R);CHKERRQ(ierr); 919019a44ceSStefano Zampini ierr = MatRestoreRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 920019a44ceSStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,vec_R_benign);CHKERRQ(ierr); 921019a44ceSStefano Zampini ierr = VecScale(vec_R_benign,m_one);CHKERRQ(ierr); 922019a44ceSStefano Zampini } 92380677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 92406656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 92506656605SStefano Zampini if (n_constraints) { 92606656605SStefano Zampini Mat B; 92780677318SStefano Zampini 928b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 92980677318SStefano Zampini for (i=0;i<n_vertices;i++) { 93080677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 93180677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 93280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 93380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 93480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 93580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 93680677318SStefano Zampini } 93780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 93880677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 93980677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 94180677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 94406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 94506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 94606656605SStefano Zampini } 94704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 94804708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 94904708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 95004708bb6SStefano Zampini } 95106656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 952019a44ceSStefano Zampini if (n_benign) { 953019a44ceSStefano Zampini ierr = MatMatMult(B0_R,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_V_benign);CHKERRQ(ierr); 954019a44ceSStefano Zampini } 95580677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 95706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 95906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 96006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 962019a44ceSStefano Zampini if (!n_benign) { 96306656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 964019a44ceSStefano Zampini } else { 965019a44ceSStefano Zampini const PetscScalar *vals; 966019a44ceSStefano Zampini PetscScalar val; 967019a44ceSStefano Zampini const PetscInt *idxs; 968019a44ceSStefano Zampini PetscInt n,j; 969019a44ceSStefano Zampini 970019a44ceSStefano Zampini /* Use explicit loops to populate the local coarse matrix */ 971019a44ceSStefano Zampini ierr = MatDenseGetArray(S_VVt,&x);CHKERRQ(ierr); 972019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 973019a44ceSStefano Zampini for (i=0;i<n_vertices;i++) { 974019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+i] = x[j*n_vertices+i]; 975019a44ceSStefano Zampini } 976019a44ceSStefano Zampini } 977019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&x);CHKERRQ(ierr); 978019a44ceSStefano Zampini ierr = MatDenseGetArray(S_V_benign,&x);CHKERRQ(ierr); 979019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 980019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 981019a44ceSStefano Zampini coarse_submat_vals[n_vertices*pcbddc->local_primal_size+j] = x[j]+vals[j]; 982019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices] = x[j]+vals[j]; 983019a44ceSStefano Zampini } 984019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 985019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_V_benign,&x);CHKERRQ(ierr); 986019a44ceSStefano Zampini ierr = MatDestroy(&S_V_benign);CHKERRQ(ierr); 987019a44ceSStefano Zampini ierr = VecDot(vec_R_benign,vec_R_benign,&val);CHKERRQ(ierr); 988019a44ceSStefano Zampini coarse_submat_vals[pcbddc->local_primal_size*pcbddc->local_primal_size-1] = -val; 989019a44ceSStefano Zampini } 99006656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 99116f15bc4SStefano Zampini } else { 992019a44ceSStefano Zampini /* TODO benign case also here */ 993019a44ceSStefano Zampini if (n_benign) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not yet implemented"); 99416f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 99516f15bc4SStefano Zampini } 99621eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 99706656605SStefano Zampini /* coarse basis functions */ 99806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 99916f15bc4SStefano Zampini PetscScalar *y; 100016f15bc4SStefano Zampini 100106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 100406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 100706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 100806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 100906656605SStefano Zampini 101006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 101106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 101606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 101706656605SStefano Zampini } 101806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 101906656605SStefano Zampini } 1020019a44ceSStefano Zampini if (n_benign) { 1021019a44ceSStefano Zampini PetscScalar *y; 1022019a44ceSStefano Zampini 1023019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1024019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*n_vertices);CHKERRQ(ierr); 1025019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1026019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1027019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1028019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 1029019a44ceSStefano Zampini 1030019a44ceSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1031019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1032019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 1033019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1034019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1035019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1036019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1037019a44ceSStefano Zampini } 1038019a44ceSStefano Zampini } 103904708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 104004708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 104106656605SStefano Zampini } 104206656605SStefano Zampini 104306656605SStefano Zampini if (n_constraints) { 104406656605SStefano Zampini Mat B; 104506656605SStefano Zampini 104606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 104706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 104880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 104906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 105006656605SStefano Zampini if (n_vertices) { 105180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 105280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 105380677318SStefano Zampini } else { 105480677318SStefano Zampini Mat S_VCt; 105580677318SStefano Zampini 105680677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 105780677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 105880677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 105980677318SStefano Zampini } 106006656605SStefano Zampini } 106106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 106206656605SStefano Zampini /* coarse basis functions */ 106306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 106406656605SStefano Zampini PetscScalar *y; 106506656605SStefano Zampini 106606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 106706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 106806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 106906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 107206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 107306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 107406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 107506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 107606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 107806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 107906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 108006656605SStefano Zampini } 108106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 108206656605SStefano Zampini } 108306656605SStefano Zampini } 108480677318SStefano Zampini if (n_constraints) { 108580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 108680677318SStefano Zampini } 108706656605SStefano Zampini 1088019a44ceSStefano Zampini if (n_benign) { 1089019a44ceSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1090019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1091019a44ceSStefano Zampini ierr = VecDestroy(&vec_R_benign);CHKERRQ(ierr); 1092019a44ceSStefano Zampini } 1093019a44ceSStefano Zampini 109406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 10953301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 109606656605SStefano Zampini 109706656605SStefano Zampini if (n_constraints) { 109816f15bc4SStefano Zampini Mat S_CCT,B_C; 109906656605SStefano Zampini 110080677318SStefano Zampini /* this is a lazy thing */ 110180677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 110206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 110306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 110406656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 110516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 110606656605SStefano Zampini if (n_vertices) { 110716f15bc4SStefano Zampini Mat B_V,S_VCT; 110806656605SStefano Zampini 110906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 111006656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 111106656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 111206656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 111316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 111406656605SStefano Zampini } 111506656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 111680677318SStefano Zampini } else { /* if there are no constraints, reset work */ 111780677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 111806656605SStefano Zampini } 111916f15bc4SStefano Zampini if (n_vertices && n_R) { 112006656605SStefano Zampini Mat A_VRT; 112180677318SStefano Zampini PetscScalar *marray; 112206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 112306656605SStefano Zampini 112480677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 112580677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 112680677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 112706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 112880677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 112980677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 113016f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 113106656605SStefano Zampini } 113206656605SStefano Zampini 113306656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 113406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 113506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 113606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 113706656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 113806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 113906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 114006656605SStefano Zampini } 114106656605SStefano Zampini } else { 114206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 114306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 114406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 114506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 114706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 114806656605SStefano Zampini } 114906656605SStefano Zampini } 115006656605SStefano Zampini /* coarse basis functions */ 115106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 115206656605SStefano Zampini PetscScalar *y; 115306656605SStefano Zampini 115406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 115506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 115606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 115806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 115906656605SStefano Zampini if (i<n_vertices) { 116006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 116106656605SStefano Zampini } 116206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 116306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 116406656605SStefano Zampini 116506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 116606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 116706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 116806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 116906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 117106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 117206656605SStefano Zampini } 117306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 117406656605SStefano Zampini } 117506656605SStefano Zampini } 1176d62866d3SStefano Zampini /* free memory */ 117788ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 117806656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 117906656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 118006656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 118106656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1182d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1183d62866d3SStefano Zampini if (n_vertices) { 1184d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1185d62866d3SStefano Zampini } 1186d62866d3SStefano Zampini if (n_constraints) { 1187d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1188d62866d3SStefano Zampini } 118988ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 119088ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 119188ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1192019a44ceSStefano Zampini //if (pcbddc->dbg_flag) { 1193019a44ceSStefano Zampini if (0) { 119488ebb749SStefano Zampini Mat coarse_sub_mat; 119525084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 119688ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 119788ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 119888ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 11998bec7fa6SStefano Zampini Mat C_B,CPHI; 12008bec7fa6SStefano Zampini IS is_dummy; 12018bec7fa6SStefano Zampini Vec mones; 120288ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 120388ebb749SStefano Zampini PetscReal real_value; 120488ebb749SStefano Zampini 120588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 120688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 120788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 120888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 120988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 121088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1211c0553b1fSStefano Zampini if (unsymmetric_check) { 121288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 121388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 121488ebb749SStefano Zampini } 121588ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 121688ebb749SStefano Zampini 121725084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 12183301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 121925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1220c0553b1fSStefano Zampini if (unsymmetric_check) { 122188ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 122288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 122388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 122488ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 122588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 122688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 122788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 122888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 122988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 123288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 123388ebb749SStefano Zampini } else { 123488ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 123588ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 123688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 123788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_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_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 124188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124288ebb749SStefano Zampini } 124388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 124488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 124588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 124688ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 124781d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 12488bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 12490fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 125006656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 12518bec7fa6SStefano Zampini 12528bec7fa6SStefano Zampini /* check constraints */ 12538bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 12548bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 12558bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 12568bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 12578bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 12588bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 12598bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1260bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1261c0553b1fSStefano Zampini if (unsymmetric_check) { 1262bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1263bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1264bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1265bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1266bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 126788ebb749SStefano Zampini } 12688bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 12698bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 12708bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 12718bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 12728bec7fa6SStefano Zampini 127325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 127488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 127588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 127688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 127788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 127888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 127988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 128088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 128188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 128288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 128388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1284c0553b1fSStefano Zampini if (unsymmetric_check) { 128588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 128688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 128788ebb749SStefano Zampini } 128888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 128988ebb749SStefano Zampini } 12908629588bSStefano Zampini /* get back data */ 12918629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 129288ebb749SStefano Zampini PetscFunctionReturn(0); 129388ebb749SStefano Zampini } 129488ebb749SStefano Zampini 129588ebb749SStefano Zampini #undef __FUNCT__ 1296d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1297d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1298aa0d41d4SStefano Zampini { 1299d65f70fdSStefano Zampini Mat *work_mat; 1300d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1301d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1302d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1303aa0d41d4SStefano Zampini PetscErrorCode ierr; 1304aa0d41d4SStefano Zampini 1305aa0d41d4SStefano Zampini PetscFunctionBegin; 1306d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1307d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1308d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1309d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1310aa0d41d4SStefano Zampini 1311d65f70fdSStefano Zampini if (!rsorted) { 1312906d46d4SStefano Zampini const PetscInt *idxs; 1313906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1314aa0d41d4SStefano Zampini 1315d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1316d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1317d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1318d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1319aa0d41d4SStefano Zampini } 1320d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1321d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1322d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1323d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1324aa0d41d4SStefano Zampini } 1325d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1326d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1327d65f70fdSStefano Zampini } else { 1328d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1329d65f70fdSStefano Zampini isrow_s = isrow; 1330aa0d41d4SStefano Zampini } 1331906d46d4SStefano Zampini 1332d65f70fdSStefano Zampini if (!csorted) { 1333d65f70fdSStefano Zampini if (isrow == iscol) { 1334d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1335d65f70fdSStefano Zampini iscol_s = isrow_s; 1336d65f70fdSStefano Zampini } else { 1337d65f70fdSStefano Zampini const PetscInt *idxs; 1338d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1339906d46d4SStefano Zampini 1340d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1341d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1342d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1343d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1344d65f70fdSStefano Zampini } 1345d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1346d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1347d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1348d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1349d65f70fdSStefano Zampini } 1350d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1351d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1352d65f70fdSStefano Zampini } 1353d65f70fdSStefano Zampini } else { 1354d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1355d65f70fdSStefano Zampini iscol_s = iscol; 1356d65f70fdSStefano Zampini } 1357d65f70fdSStefano Zampini 1358d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1359d65f70fdSStefano Zampini 1360d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1361906d46d4SStefano Zampini Mat new_mat; 1362d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1363906d46d4SStefano Zampini 1364d65f70fdSStefano Zampini if (!rsorted) { 1365d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1366d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1367d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1368d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1369906d46d4SStefano Zampini } 1370d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1371d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1372d65f70fdSStefano Zampini } else { 1373d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1374906d46d4SStefano Zampini } 1375d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1376d65f70fdSStefano Zampini 1377d65f70fdSStefano Zampini if (!csorted) { 1378d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1379d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1380d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1381d65f70fdSStefano Zampini } else { 1382d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1383d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1384d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1385d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1386d65f70fdSStefano Zampini } 1387d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1388d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1389d65f70fdSStefano Zampini } 1390d65f70fdSStefano Zampini } else { 1391d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1392d65f70fdSStefano Zampini } 1393d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1394d65f70fdSStefano Zampini 1395d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1396d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1397d65f70fdSStefano Zampini work_mat[0] = new_mat; 1398d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1399d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1400d65f70fdSStefano Zampini } 1401d65f70fdSStefano Zampini 1402d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1403d65f70fdSStefano Zampini *B = work_mat[0]; 1404d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1405d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1406d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1407d65f70fdSStefano Zampini PetscFunctionReturn(0); 1408d65f70fdSStefano Zampini } 1409d65f70fdSStefano Zampini 1410d65f70fdSStefano Zampini #undef __FUNCT__ 14115e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 14125e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1413aa0d41d4SStefano Zampini { 1414aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 14153bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 14165e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1417d65f70fdSStefano Zampini Mat new_mat; 14185e8657edSStefano Zampini IS is_local,is_global; 1419d65f70fdSStefano Zampini PetscInt local_size; 1420d65f70fdSStefano Zampini PetscBool isseqaij; 1421aa0d41d4SStefano Zampini PetscErrorCode ierr; 1422aa0d41d4SStefano Zampini 1423aa0d41d4SStefano Zampini PetscFunctionBegin; 1424aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 14255e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 14265e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 14273bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,is_local,&is_global);CHKERRQ(ierr); 1428aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1429d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1430aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1431906d46d4SStefano Zampini 1432906d46d4SStefano Zampini /* check */ 1433906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1434906d46d4SStefano Zampini Vec x,x_change; 1435906d46d4SStefano Zampini PetscReal error; 1436906d46d4SStefano Zampini 14375e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1438906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 14395e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1440e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1441e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1442d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1443e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1444e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1445906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1446906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1447906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1448906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1449906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1450906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1451906d46d4SStefano Zampini } 1452906d46d4SStefano Zampini 145322d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 14549b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 145522d5777bSStefano Zampini if (isseqaij) { 1456d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1457aa0d41d4SStefano Zampini } else { 1458aa0d41d4SStefano Zampini Mat work_mat; 1459aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1460d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1461aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1462aa0d41d4SStefano Zampini } 14633301b35fSStefano Zampini if (matis->A->symmetric_set) { 14643301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1465e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 14663301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1467e496cd5dSStefano Zampini #endif 14683301b35fSStefano Zampini } 146945a1bb75SStefano Zampini /* 147045a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1471d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 147245a1bb75SStefano Zampini */ 1473d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1474aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1475aa0d41d4SStefano Zampini } 1476aa0d41d4SStefano Zampini 1477aa0d41d4SStefano Zampini #undef __FUNCT__ 1478a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 14798ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1480a64d13efSStefano Zampini { 1481a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1482a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1483d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 148453892102SStefano Zampini PetscInt *idx_R_local=NULL; 14853a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 14863a50541eSStefano Zampini PetscInt vbs,bs; 14876816873aSStefano Zampini PetscBT bitmask=NULL; 1488a64d13efSStefano Zampini PetscErrorCode ierr; 1489a64d13efSStefano Zampini 1490a64d13efSStefano Zampini PetscFunctionBegin; 1491b23d619eSStefano Zampini /* 1492b23d619eSStefano Zampini No need to setup local scatters if 1493b23d619eSStefano Zampini - primal space is unchanged 1494b23d619eSStefano Zampini AND 1495b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1496b23d619eSStefano Zampini AND 1497b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1498b23d619eSStefano Zampini */ 1499b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1500f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1501f4ddd8eeSStefano Zampini } 1502f4ddd8eeSStefano Zampini /* destroy old objects */ 1503f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1504f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1505f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1506a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1507b371cd4fSStefano Zampini n_B = pcis->n_B; 1508b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1509b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15103a50541eSStefano Zampini 1511a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 15126816873aSStefano Zampini 151353892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 15146816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1515854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1516a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1517a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 15180e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1519a64d13efSStefano Zampini } 1520a64d13efSStefano Zampini 1521a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 15224641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 15236816873aSStefano Zampini idx_R_local[n_R++] = i; 1524a64d13efSStefano Zampini } 1525a64d13efSStefano Zampini } 152653892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 15276816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15286816873aSStefano Zampini 152953892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 153053892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 15316816873aSStefano Zampini } 15323a50541eSStefano Zampini 15333a50541eSStefano Zampini /* Block code */ 15343a50541eSStefano Zampini vbs = 1; 15353a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 15363a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 15373a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 15383a50541eSStefano Zampini PetscInt *vary; 1539d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1540785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 15413a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1542d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1543d3df7717SStefano 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 */ 15440e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1545d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 15463a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 15473a50541eSStefano Zampini is_blocked = PETSC_FALSE; 15483a50541eSStefano Zampini break; 15493a50541eSStefano Zampini } 15503a50541eSStefano Zampini } 1551d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1552d3df7717SStefano Zampini } else { 1553d3df7717SStefano Zampini /* Verify directly the R set */ 1554d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1555d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1556d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1557d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1558d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1559d3df7717SStefano Zampini break; 1560d3df7717SStefano Zampini } 1561d3df7717SStefano Zampini } 1562d3df7717SStefano Zampini } 1563d3df7717SStefano Zampini } 15643a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 15653a50541eSStefano Zampini vbs = bs; 15663a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 15673a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 15683a50541eSStefano Zampini } 15693a50541eSStefano Zampini } 15703a50541eSStefano Zampini } 15713a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 157253892102SStefano Zampini if (sub_schurs->reuse_mumps) { 157353892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 157453892102SStefano Zampini 157553892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 157653892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 157753892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 157853892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 157953892102SStefano Zampini } else { 15803a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 158153892102SStefano Zampini } 1582a64d13efSStefano Zampini 1583a64d13efSStefano Zampini /* print some info if requested */ 1584a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1585019a44ceSStefano Zampini PetscInt benign; 1586019a44ceSStefano Zampini 1587019a44ceSStefano Zampini benign = (PetscInt)(!!pcbddc->zerodiag); 1588a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1589a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 15900fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1591a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1592a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1593019a44ceSStefano 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); 1594a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1595a64d13efSStefano Zampini } 1596a64d13efSStefano Zampini 1597a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 15986816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 15996816873aSStefano Zampini IS is_aux1,is_aux2; 16006816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 16016816873aSStefano Zampini 16023a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1603854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1604854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1605a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 16064641a718SStefano Zampini for (i=0; i<n_D; i++) { 16074641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 16084641a718SStefano Zampini } 1609a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1610a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16114641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 16124641a718SStefano Zampini aux_array1[j++] = i; 1613a64d13efSStefano Zampini } 1614a64d13efSStefano Zampini } 1615a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1616a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1617a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 16184641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 16194641a718SStefano Zampini aux_array2[j++] = i; 1620a64d13efSStefano Zampini } 1621a64d13efSStefano Zampini } 1622a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1623a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1624a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1625a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1626a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1627a64d13efSStefano Zampini 16288eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1629785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1630a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16314641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 16324641a718SStefano Zampini aux_array1[j++] = i; 1633a64d13efSStefano Zampini } 1634a64d13efSStefano Zampini } 1635a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1636a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1637a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1638a64d13efSStefano Zampini } 16394641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 16403a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1641d62866d3SStefano Zampini } else { 164253892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16436816873aSStefano Zampini IS tis; 16446816873aSStefano Zampini PetscInt schur_size; 16456816873aSStefano Zampini 164653892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 16476816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 164853892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 16496816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16506816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 16516816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 16526816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 16536816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1654d62866d3SStefano Zampini } 1655d62866d3SStefano Zampini } 1656a64d13efSStefano Zampini PetscFunctionReturn(0); 1657a64d13efSStefano Zampini } 1658a64d13efSStefano Zampini 1659304d26faSStefano Zampini 1660304d26faSStefano Zampini #undef __FUNCT__ 1661304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1662684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1663304d26faSStefano Zampini { 1664304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1665304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1666304d26faSStefano Zampini PC pc_temp; 1667304d26faSStefano Zampini Mat A_RR; 1668f4ddd8eeSStefano Zampini MatReuse reuse; 1669304d26faSStefano Zampini PetscScalar m_one = -1.0; 1670304d26faSStefano Zampini PetscReal value; 167104708bb6SStefano Zampini PetscInt n_D,n_R; 16729577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1673304d26faSStefano Zampini PetscErrorCode ierr; 1674e604994aSStefano Zampini /* prefixes stuff */ 1675312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1676e604994aSStefano Zampini size_t len; 1677304d26faSStefano Zampini 1678304d26faSStefano Zampini PetscFunctionBegin; 1679304d26faSStefano Zampini 1680e604994aSStefano Zampini /* compute prefixes */ 1681e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1682e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1683e604994aSStefano Zampini if (!pcbddc->current_level) { 1684e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1685e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1686e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1687e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1688e604994aSStefano Zampini } else { 1689e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1690312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1691e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1692e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1693312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1694312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 169534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 169634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1697e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1698e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1699e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1700e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1701e604994aSStefano Zampini } 1702e604994aSStefano Zampini 1703304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1704684f6988SStefano Zampini if (dirichlet) { 1705d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 17063301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 17073301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1708964fefecSStefano Zampini } 1709ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1710964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1711304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1712304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1713304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1714304d26faSStefano Zampini /* default */ 1715304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1716e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 17179577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1718304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 17199577ea80SStefano Zampini if (issbaij) { 17209577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 17219577ea80SStefano Zampini } else { 1722304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 17239577ea80SStefano Zampini } 1724304d26faSStefano Zampini /* Allow user's customization */ 1725304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1726304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1727304d26faSStefano Zampini } 1728d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1729d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1730d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1731d62866d3SStefano Zampini 1732d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1733d5574798SStefano Zampini } 1734304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1735304d26faSStefano Zampini if (!n_D) { 1736304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1737304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1738304d26faSStefano Zampini } 1739304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1740304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1741304d26faSStefano Zampini /* set ksp_D into pcis data */ 1742304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1743304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1744304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1745684f6988SStefano Zampini } 1746304d26faSStefano Zampini 1747304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1748684f6988SStefano Zampini A_RR = 0; 1749684f6988SStefano Zampini if (neumann) { 1750d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 175104708bb6SStefano Zampini PetscInt ibs,mbs; 175204708bb6SStefano Zampini PetscBool issbaij; 175304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1754f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 17558ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1756f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1757f4ddd8eeSStefano Zampini PetscInt nn_R; 175881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1759f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1760f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1761f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1762f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1763f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1764f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1765f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1766727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1767f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1768f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1769f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1770f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1771f4ddd8eeSStefano Zampini } 1772f4ddd8eeSStefano Zampini } 1773f4ddd8eeSStefano Zampini /* last check */ 1774d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1775f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1776f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1777f4ddd8eeSStefano Zampini } 1778f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1779f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1780f4ddd8eeSStefano Zampini } 1781f4ddd8eeSStefano Zampini /* extract A_RR */ 1782af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1783af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 178404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 178504708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 178604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 178704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 178804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1789af732b37SStefano Zampini } else { 179004708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 17916816873aSStefano Zampini } 179204708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 179304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 179404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 179504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 179604708bb6SStefano Zampini } else { 179704708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 179804708bb6SStefano Zampini } 179904708bb6SStefano Zampini } 180004708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1801f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 18023301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 18033301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 18046816873aSStefano Zampini } 18056816873aSStefano Zampini } else { 18066816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 18076816873aSStefano Zampini 18086816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 18096816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 18106816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1811af732b37SStefano Zampini } 1812f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1813304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1814304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1815304d26faSStefano Zampini /* default */ 1816304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1817e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1818304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 18199577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 18209577ea80SStefano Zampini if (issbaij) { 18219577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 18229577ea80SStefano Zampini } else { 1823304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 18249577ea80SStefano Zampini } 1825304d26faSStefano Zampini /* Allow user's customization */ 1826304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1827304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1828304d26faSStefano Zampini } 1829d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1830304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1831304d26faSStefano Zampini if (!n_R) { 1832304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1833304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1834304d26faSStefano Zampini } 1835d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1836d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1837d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1838d62866d3SStefano Zampini 1839d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1840d62866d3SStefano Zampini } 1841304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1842304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1843684f6988SStefano Zampini } 18446816873aSStefano Zampini /* free Neumann problem's matrix */ 18456816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1846304d26faSStefano Zampini 1847304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 18480fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1849684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1850684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1851684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1852684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1853684f6988SStefano Zampini } 1854684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 18550fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 18560fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 18570fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 18580fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 18590fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1860304d26faSStefano Zampini /* need to be adapted? */ 1861b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1862b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1863b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1864304d26faSStefano Zampini /* print info */ 1865304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1866e604994aSStefano 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); 1867304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1868304d26faSStefano Zampini } 1869b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1870298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1871304d26faSStefano Zampini } 1872684f6988SStefano Zampini } 1873684f6988SStefano Zampini if (neumann) { /* Neumann */ 18746816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 18750fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 18760fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 18770fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 18780fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 18790fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1880304d26faSStefano Zampini /* need to be adapted? */ 1881b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1882b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1883304d26faSStefano Zampini /* print info */ 1884304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1885e604994aSStefano 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); 1886304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1887304d26faSStefano Zampini } 1888b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1889298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1890304d26faSStefano Zampini } 18910fccc4e9SStefano Zampini } 1892684f6988SStefano Zampini } 1893304d26faSStefano Zampini PetscFunctionReturn(0); 1894304d26faSStefano Zampini } 1895304d26faSStefano Zampini 1896304d26faSStefano Zampini #undef __FUNCT__ 1897ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 189880677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1899674ae819SStefano Zampini { 1900674ae819SStefano Zampini PetscErrorCode ierr; 1901674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1902be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1903674ae819SStefano Zampini 1904674ae819SStefano Zampini PetscFunctionBegin; 1905be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 190680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 190720c7b377SStefano Zampini } 190880677318SStefano Zampini if (!pcbddc->switch_static) { 190980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 191080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 191180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 191220c7b377SStefano Zampini } 1913be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 191480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 191580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 191620c7b377SStefano Zampini } else { 1917be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1918be83ff47SStefano Zampini 191953892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192053892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192120c7b377SStefano Zampini } 1922be83ff47SStefano Zampini } else { 192380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 192780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 192880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 192980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 193080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 193180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1932674ae819SStefano Zampini } 1933674ae819SStefano Zampini } 1934be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 193580677318SStefano Zampini if (applytranspose) { 193680677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 193780677318SStefano Zampini } else { 193880677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 193980677318SStefano Zampini } 194053892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1941be83ff47SStefano Zampini } else { 1942be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1943be83ff47SStefano Zampini 1944be83ff47SStefano Zampini if (applytranspose) { 194553892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1946be83ff47SStefano Zampini } else { 194753892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1948be83ff47SStefano Zampini } 194953892102SStefano Zampini #endif 1950be83ff47SStefano Zampini } 195180677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 195280677318SStefano Zampini if (!pcbddc->switch_static) { 1953be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 195480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1956be83ff47SStefano Zampini } else { 1957be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1958be83ff47SStefano Zampini 195953892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 196053892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1961be83ff47SStefano Zampini } 196280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 196380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 196480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 196580677318SStefano Zampini } 196680677318SStefano Zampini } else { 196780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 197280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 197380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 197480677318SStefano Zampini } 197580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1979674ae819SStefano Zampini } 1980674ae819SStefano Zampini PetscFunctionReturn(0); 1981674ae819SStefano Zampini } 1982674ae819SStefano Zampini 1983dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1984674ae819SStefano Zampini #undef __FUNCT__ 1985674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1986dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1987674ae819SStefano Zampini { 1988674ae819SStefano Zampini PetscErrorCode ierr; 1989674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1990674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1991674ae819SStefano Zampini const PetscScalar zero = 0.0; 1992674ae819SStefano Zampini 1993674ae819SStefano Zampini PetscFunctionBegin; 1994dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1995dc359a40SStefano Zampini if (applytranspose) { 1996674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 19978eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1998dc359a40SStefano Zampini } else { 1999674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2000674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 200115aaf578SStefano Zampini } 2002*efc2fbd9SStefano Zampini 2003*efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 2004*efc2fbd9SStefano Zampini if (pcbddc->benign_change) { 2005*efc2fbd9SStefano Zampini PetscScalar *array; 2006*efc2fbd9SStefano Zampini 2007*efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2008*efc2fbd9SStefano Zampini array[pcbddc->local_primal_size-1] += pcbddc->benign_p0; 2009*efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2010*efc2fbd9SStefano Zampini } 2011*efc2fbd9SStefano Zampini 201212edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 201312edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 201412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201612edc857SStefano Zampini 20179f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 201812edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 201912edc857SStefano Zampini if (pcbddc->coarse_ksp) { 2020964fefecSStefano Zampini Vec rhs,sol; 2021964fefecSStefano Zampini 2022964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2023964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 202412edc857SStefano Zampini if (applytranspose) { 2025964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 202612edc857SStefano Zampini } else { 2027964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 202812edc857SStefano Zampini } 202912edc857SStefano Zampini } 2030674ae819SStefano Zampini 2031674ae819SStefano Zampini /* Local solution on R nodes */ 203280677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 203380677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 20349f00e9b4SStefano Zampini } 2035674ae819SStefano Zampini 20369f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 20379f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 203812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2039674ae819SStefano Zampini 2040674ae819SStefano Zampini /* Sum contributions from two levels */ 2041dc359a40SStefano Zampini if (applytranspose) { 2042dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2043dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2044dc359a40SStefano Zampini } else { 2045674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 20468eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2047dc359a40SStefano Zampini } 2048*efc2fbd9SStefano Zampini /* store p0 */ 2049*efc2fbd9SStefano Zampini if (pcbddc->benign_change) { 2050*efc2fbd9SStefano Zampini PetscScalar *array; 2051*efc2fbd9SStefano Zampini 2052*efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2053*efc2fbd9SStefano Zampini pcbddc->benign_p0 = array[pcbddc->local_primal_size-1]; 2054*efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2055*efc2fbd9SStefano Zampini } 2056674ae819SStefano Zampini PetscFunctionReturn(0); 2057674ae819SStefano Zampini } 2058674ae819SStefano Zampini 2059674ae819SStefano Zampini #undef __FUNCT__ 2060674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 206112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2062674ae819SStefano Zampini { 2063674ae819SStefano Zampini PetscErrorCode ierr; 2064674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 206558da7f69SStefano Zampini PetscScalar *array; 206612edc857SStefano Zampini Vec from,to; 2067674ae819SStefano Zampini 2068674ae819SStefano Zampini PetscFunctionBegin; 206912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 207012edc857SStefano Zampini from = pcbddc->coarse_vec; 207112edc857SStefano Zampini to = pcbddc->vec1_P; 207212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 207312edc857SStefano Zampini Vec tvec; 207458da7f69SStefano Zampini 207558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 207658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 207712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 207858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 207958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 208058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 208112edc857SStefano Zampini } 208212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 208312edc857SStefano Zampini from = pcbddc->vec1_P; 208412edc857SStefano Zampini to = pcbddc->coarse_vec; 208512edc857SStefano Zampini } 208612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2087674ae819SStefano Zampini PetscFunctionReturn(0); 2088674ae819SStefano Zampini } 2089674ae819SStefano Zampini 2090674ae819SStefano Zampini #undef __FUNCT__ 2091674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 209212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2093674ae819SStefano Zampini { 2094674ae819SStefano Zampini PetscErrorCode ierr; 2095674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 209658da7f69SStefano Zampini PetscScalar *array; 209712edc857SStefano Zampini Vec from,to; 2098674ae819SStefano Zampini 2099674ae819SStefano Zampini PetscFunctionBegin; 210012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 210112edc857SStefano Zampini from = pcbddc->coarse_vec; 210212edc857SStefano Zampini to = pcbddc->vec1_P; 210312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 210412edc857SStefano Zampini from = pcbddc->vec1_P; 210512edc857SStefano Zampini to = pcbddc->coarse_vec; 210612edc857SStefano Zampini } 210712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 210812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 210912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 211012edc857SStefano Zampini Vec tvec; 211158da7f69SStefano Zampini 211212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 211358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 211458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 211558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 211658da7f69SStefano Zampini } 211758da7f69SStefano Zampini } else { 211858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 211958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 212012edc857SStefano Zampini } 212112edc857SStefano Zampini } 2122674ae819SStefano Zampini PetscFunctionReturn(0); 2123674ae819SStefano Zampini } 2124674ae819SStefano Zampini 2125984c4197SStefano Zampini /* uncomment for testing purposes */ 2126984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2127674ae819SStefano Zampini #undef __FUNCT__ 2128674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2129674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2130674ae819SStefano Zampini { 2131674ae819SStefano Zampini PetscErrorCode ierr; 2132674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2133674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2134674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2135984c4197SStefano Zampini /* one and zero */ 2136984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2137984c4197SStefano Zampini /* space to store constraints and their local indices */ 21389162d606SStefano Zampini PetscScalar *constraints_data; 21399162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 21409162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 21419162d606SStefano Zampini PetscInt *constraints_n; 2142984c4197SStefano Zampini /* iterators */ 2143b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2144984c4197SStefano Zampini /* BLAS integers */ 2145e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2146e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2147c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2148727cdba6SStefano Zampini /* reuse */ 21490e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 21500e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2151984c4197SStefano Zampini /* change of basis */ 2152b3d85658SStefano Zampini PetscBool qr_needed; 21539162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2154984c4197SStefano Zampini /* auxiliary stuff */ 215564efe560SStefano Zampini PetscInt *nnz,*is_indices; 21568a0068c3SStefano Zampini PetscInt ncc; 2157984c4197SStefano Zampini /* some quantities */ 215845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2159a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2160984c4197SStefano Zampini 2161674ae819SStefano Zampini PetscFunctionBegin; 21628e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 21638e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 21648e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2165088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2166088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 21670e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 21680e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 21690e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21700e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21710e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2172088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2173cf5a6209SStefano Zampini 2174cf5a6209SStefano Zampini /* print some info */ 2175cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2176cf5a6209SStefano Zampini IS vertices; 2177cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2178cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2179cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2180cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2181cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2182cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2183cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2184fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2185fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2186cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2187cf5a6209SStefano Zampini } 2188cf5a6209SStefano Zampini 2189cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 21909162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2191cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2192cf5a6209SStefano Zampini const Vec *nearnullvecs; 2193cf5a6209SStefano Zampini Vec *localnearnullsp; 2194cf5a6209SStefano Zampini PetscScalar *array; 2195cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2196cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2197674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2198b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2199674ae819SStefano Zampini PetscScalar *work; 2200674ae819SStefano Zampini PetscReal *singular_vals; 2201674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2202674ae819SStefano Zampini PetscReal *rwork; 2203674ae819SStefano Zampini #endif 2204674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2205674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2206674ae819SStefano Zampini #else 2207964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2208964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2209674ae819SStefano Zampini #endif 2210674ae819SStefano Zampini 2211674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2212d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2213d06fc5fdSStefano Zampini /* free unneeded index sets */ 2214d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2215d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2216674ae819SStefano Zampini } 2217d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2218d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2219d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2220d06fc5fdSStefano Zampini } 2221d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2222d06fc5fdSStefano Zampini n_ISForEdges = 0; 2223d06fc5fdSStefano Zampini } 2224d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2225d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2226d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2227d06fc5fdSStefano Zampini } 2228d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2229d06fc5fdSStefano Zampini n_ISForFaces = 0; 2230d06fc5fdSStefano Zampini } 223170022509SStefano Zampini 223270022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 223370022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 223470022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 223570022509SStefano Zampini if (pcbddc->NullSpace) { 223670022509SStefano Zampini PetscBool tbool[2],gbool[2]; 223770022509SStefano Zampini 223870022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2239b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2240d06fc5fdSStefano Zampini if (!ISForEdges) { 2241d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2242d06fc5fdSStefano Zampini } 2243b8ffe317SStefano Zampini } 2244d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2245d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2246d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2247d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2248d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 224998a51de6SStefano Zampini } 225070022509SStefano Zampini #endif 225108122e43SStefano Zampini 2252674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2253674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2254674ae819SStefano Zampini if (nearnullsp) { 2255674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2256f4ddd8eeSStefano Zampini /* remove any stored info */ 2257f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2258f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2259f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2260f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2261f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2262473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2263f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2264f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2265f4ddd8eeSStefano Zampini } 2266984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2267984c4197SStefano Zampini nnsp_size = 0; 2268674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2269674ae819SStefano Zampini } 2270984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2271984c4197SStefano Zampini max_constraints = nnsp_size; 2272984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2273984c4197SStefano Zampini 2274674ae819SStefano Zampini /* 2275674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 22769162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 22779162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 22789162d606SStefano Zampini There can be multiple constraints per connected component 2279674ae819SStefano Zampini */ 2280674ae819SStefano Zampini n_vertices = 0; 2281674ae819SStefano Zampini if (ISForVertices) { 2282674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2283674ae819SStefano Zampini } 22849162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 22859162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 22869162d606SStefano Zampini 22879162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 22889162d606SStefano Zampini total_counts *= max_constraints; 2289674ae819SStefano Zampini total_counts += n_vertices; 22904641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 22919162d606SStefano Zampini 2292674ae819SStefano Zampini total_counts = 0; 2293674ae819SStefano Zampini max_size_of_constraint = 0; 2294674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 22959162d606SStefano Zampini IS used_is; 2296674ae819SStefano Zampini if (i<n_ISForEdges) { 22979162d606SStefano Zampini used_is = ISForEdges[i]; 2298674ae819SStefano Zampini } else { 22999162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2300674ae819SStefano Zampini } 23019162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2302674ae819SStefano Zampini total_counts += j; 2303674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2304674ae819SStefano Zampini } 23059162d606SStefano 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); 23069162d606SStefano Zampini 2307984c4197SStefano Zampini /* get local part of global near null space vectors */ 2308785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2309984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2310984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2311e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2312e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2313984c4197SStefano Zampini } 2314674ae819SStefano Zampini 2315242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2316242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2317a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2318242a89d7SStefano Zampini 2319984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2320a773dcb8SStefano Zampini if (!skip_lapack) { 2321674ae819SStefano Zampini PetscScalar temp_work; 2322911cabfeSStefano Zampini 2323674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2324984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2325785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2326785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2327785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2328674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2329785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2330674ae819SStefano Zampini #endif 2331674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2332c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2333c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2334674ae819SStefano Zampini lwork = -1; 2335674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2336674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2337c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2338674ae819SStefano Zampini #else 2339c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2340674ae819SStefano Zampini #endif 2341674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2342984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2343674ae819SStefano Zampini #else /* on missing GESVD */ 2344674ae819SStefano Zampini /* SVD */ 2345674ae819SStefano Zampini PetscInt max_n,min_n; 2346674ae819SStefano Zampini max_n = max_size_of_constraint; 2347984c4197SStefano Zampini min_n = max_constraints; 2348984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2349674ae819SStefano Zampini min_n = max_size_of_constraint; 2350984c4197SStefano Zampini max_n = max_constraints; 2351674ae819SStefano Zampini } 2352785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2353674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2354785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2355674ae819SStefano Zampini #endif 2356674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2357674ae819SStefano Zampini lwork = -1; 2358e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2359e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2360b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2361674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2362674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23639162d606SStefano 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)); 2364674ae819SStefano Zampini #else 23659162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr)); 2366674ae819SStefano Zampini #endif 2367674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2368984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2369984c4197SStefano Zampini #endif /* on missing GESVD */ 2370674ae819SStefano Zampini /* Allocate optimal workspace */ 2371674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2372854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2373674ae819SStefano Zampini } 2374674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2375674ae819SStefano Zampini total_counts = 0; 23769162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 23779162d606SStefano Zampini constraints_data_ptr[0] = 0; 2378674ae819SStefano Zampini /* vertices */ 23799162d606SStefano Zampini if (n_vertices) { 2380674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 23819162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 23829162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2383674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 23849162d606SStefano Zampini constraints_n[total_counts] = 1; 23859162d606SStefano Zampini constraints_data[total_counts] = 1.0; 23869162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 23879162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2388674ae819SStefano Zampini total_counts++; 2389674ae819SStefano Zampini } 2390674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2391984c4197SStefano Zampini PetscBool used_vertex; 2392674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2393674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2394674ae819SStefano Zampini k = 0; 2395674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2396984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2397984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 23989162d606SStefano Zampini constraints_n[total_counts] = 1; 23999162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 24009162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24019162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24029162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2403674ae819SStefano Zampini total_counts++; 2404674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2405674ae819SStefano Zampini } 2406984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2407674ae819SStefano Zampini k++; 2408674ae819SStefano Zampini } 2409674ae819SStefano Zampini } 2410674ae819SStefano Zampini } 2411674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2412674ae819SStefano Zampini n_vertices = total_counts; 2413674ae819SStefano Zampini } 2414984c4197SStefano Zampini 2415674ae819SStefano Zampini /* edges and faces */ 24169162d606SStefano Zampini total_counts_cc = total_counts; 2417911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 24189162d606SStefano Zampini IS used_is; 24199162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 24209162d606SStefano Zampini 2421911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 24229162d606SStefano Zampini used_is = ISForEdges[ncc]; 2423984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2424674ae819SStefano Zampini } else { 24259162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2426984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2427674ae819SStefano Zampini } 2428674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 24299162d606SStefano Zampini 24309162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 24319162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2432984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2433984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2434674ae819SStefano Zampini if (nnsp_has_cnst) { 24355b08dc53SStefano Zampini PetscScalar quad_value; 24369162d606SStefano Zampini 24379162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24389162d606SStefano Zampini idxs_copied = PETSC_TRUE; 24399162d606SStefano Zampini 2440a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2441674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2442a773dcb8SStefano Zampini } else { 2443a773dcb8SStefano Zampini quad_value = 1.0; 2444a773dcb8SStefano Zampini } 2445674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24469162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2447674ae819SStefano Zampini } 24489162d606SStefano Zampini temp_constraints++; 2449674ae819SStefano Zampini total_counts++; 2450674ae819SStefano Zampini } 2451674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2452984c4197SStefano Zampini PetscReal real_value; 24539162d606SStefano Zampini PetscScalar *ptr_to_data; 24549162d606SStefano Zampini 2455984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 24569162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2457674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24589162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2459674ae819SStefano Zampini } 2460984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2461984c4197SStefano Zampini /* check if array is null on the connected component */ 2462e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24639162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 24645b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2465674ae819SStefano Zampini temp_constraints++; 2466674ae819SStefano Zampini total_counts++; 24679162d606SStefano Zampini if (!idxs_copied) { 24689162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24699162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2470674ae819SStefano Zampini } 2471674ae819SStefano Zampini } 24729162d606SStefano Zampini } 24739162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 247445a1bb75SStefano Zampini valid_constraints = temp_constraints; 2475eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2476a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 24779162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 24789162d606SStefano Zampini 24799162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2480a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24819162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2482a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 24839162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2484a773dcb8SStefano Zampini } else { /* perform SVD */ 2485984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 24869162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2487674ae819SStefano Zampini 2488674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2489984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2490984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2491984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2492984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2493984c4197SStefano Zampini from that computed using LAPACKgesvd 2494984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2495984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2496984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2497674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2498e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2499984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2500674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2501674ae819SStefano Zampini for (k=0;k<j+1;k++) { 25029162d606SStefano 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)); 2503674ae819SStefano Zampini } 2504674ae819SStefano Zampini } 2505e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2506e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2507e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2508674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2509c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2510674ae819SStefano Zampini #else 2511c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2512674ae819SStefano Zampini #endif 2513674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2514984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2515984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2516674ae819SStefano Zampini j = 0; 2517984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2518674ae819SStefano Zampini total_counts = total_counts-j; 251945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2520e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2521c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2522c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2523c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2524c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2525c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2526c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2527674ae819SStefano Zampini if (j<temp_constraints) { 2528984c4197SStefano Zampini PetscInt ii; 2529984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2530674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25319162d606SStefano 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)); 2532674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2533984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2534674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 25359162d606SStefano 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]; 2536674ae819SStefano Zampini } 2537674ae819SStefano Zampini } 2538674ae819SStefano Zampini } 2539674ae819SStefano Zampini #else /* on missing GESVD */ 2540e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2541e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2542b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2543674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2544674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25459162d606SStefano 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)); 2546674ae819SStefano Zampini #else 25479162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr)); 2548674ae819SStefano Zampini #endif 2549984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2550674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2551984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2552e310c8b4SStefano Zampini k = temp_constraints; 2553e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2554674ae819SStefano Zampini j = 0; 2555e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 255645a1bb75SStefano Zampini valid_constraints = k-j; 2557911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2558984c4197SStefano Zampini #endif /* on missing GESVD */ 2559674ae819SStefano Zampini } 2560a773dcb8SStefano Zampini } 25619162d606SStefano Zampini /* update pointers information */ 25629162d606SStefano Zampini if (valid_constraints) { 25639162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 25649162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 25659162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 25669162d606SStefano Zampini /* set change_of_basis flag */ 256745a1bb75SStefano Zampini if (boolforchange) { 2568b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 25699162d606SStefano Zampini } 2570b3d85658SStefano Zampini total_counts_cc++; 257145a1bb75SStefano Zampini } 257245a1bb75SStefano Zampini } 2573984c4197SStefano Zampini /* free workspace */ 25748f1c130eSStefano Zampini if (!skip_lapack) { 2575984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2576984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2577984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2578984c4197SStefano Zampini #endif 2579984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2580984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2581984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2582984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2583984c4197SStefano Zampini #endif 2584984c4197SStefano Zampini } 2585984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2586984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2587984c4197SStefano Zampini } 2588984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2589cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2590cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2591cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2592cf5a6209SStefano Zampini } 2593cf5a6209SStefano Zampini if (n_ISForFaces) { 2594cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2595cf5a6209SStefano Zampini } 2596cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2597cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2598cf5a6209SStefano Zampini } 2599cf5a6209SStefano Zampini if (n_ISForEdges) { 2600cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2601cf5a6209SStefano Zampini } 2602cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 260308122e43SStefano Zampini } else { 260408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2605984c4197SStefano Zampini 260608122e43SStefano Zampini total_counts = 0; 260708122e43SStefano Zampini n_vertices = 0; 2608d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2609d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 261008122e43SStefano Zampini } 261108122e43SStefano Zampini max_constraints = 0; 26129162d606SStefano Zampini total_counts_cc = 0; 261308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 261408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 26159162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 261608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 261708122e43SStefano Zampini } 26189162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 26199162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 26209162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 26219162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 262274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 26239162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 26249162d606SStefano Zampini total_counts_cc = 0; 26259162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 26269162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 26279162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 262808122e43SStefano Zampini } 262908122e43SStefano Zampini } 26309162d606SStefano Zampini #if 0 26319162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 26329162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26339162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 26349162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 26359162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 26369162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 26379162d606SStefano Zampini } 26389162d606SStefano Zampini printf("\n"); 26399162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 26409162d606SStefano Zampini } 26411b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 26428bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 26431b968477SStefano Zampini } 26441b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 26458bec7fa6SStefano 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]); 26461b968477SStefano Zampini } 264708122e43SStefano Zampini #endif 264808122e43SStefano Zampini 26498bec7fa6SStefano Zampini max_size_of_constraint = 0; 26509162d606SStefano 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]); 26519162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 265208122e43SStefano Zampini /* Change of basis */ 2653b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 265408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 265508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 265608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2657b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 265808122e43SStefano Zampini } 265908122e43SStefano Zampini } 266008122e43SStefano Zampini } 266108122e43SStefano Zampini } 2662984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2663019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2664019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 266508122e43SStefano Zampini 26669162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 26679162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 26689162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 26699162d606SStefano 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); 267008122e43SStefano Zampini } 2671674ae819SStefano Zampini 2672674ae819SStefano Zampini /* Create constraint matrix */ 2673674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 267416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2675984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2676984c4197SStefano Zampini 2677984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2678a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2679a717540cSStefano Zampini qr_needed = PETSC_FALSE; 268074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2681984c4197SStefano Zampini total_primal_vertices=0; 2682b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 26839162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2685984c4197SStefano Zampini if (size_of_constraint == 1) { 26869162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2687b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 268864efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 26899162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 26909162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2691a717540cSStefano Zampini } 2692b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 269374d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2694a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2695a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2696a717540cSStefano Zampini } 2697fa434743SStefano Zampini } else { 2698b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2699fa434743SStefano Zampini } 2700a717540cSStefano Zampini } 2701b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2702b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2703674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 270470022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2705b3d85658SStefano Zampini 2706019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2707019a44ceSStefano 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); 27080e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 27090e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2710984c4197SStefano Zampini 2711984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 271274d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2713785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2714984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 271574d5cdf7SStefano Zampini 2716984c4197SStefano Zampini j = total_primal_vertices; 271774d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2718b3d85658SStefano Zampini cum = total_primal_vertices; 27199162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27204641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2721b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2722b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2723b3d85658SStefano Zampini cum++; 27249162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 272574d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 272674d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 272774d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 272874d5cdf7SStefano Zampini } 27299162d606SStefano Zampini j += constraints_n[i]; 2730674ae819SStefano Zampini } 2731674ae819SStefano Zampini } 2732674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2733674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2734088faed8SStefano Zampini 2735674ae819SStefano Zampini /* set values in constraint matrix */ 2736984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 27370e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2738674ae819SStefano Zampini } 2739984c4197SStefano Zampini total_counts = total_primal_vertices; 27409162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27414641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 27429162d606SStefano Zampini PetscInt *cols; 27439162d606SStefano Zampini 27449162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 27459162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 27469162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27479162d606SStefano Zampini PetscInt row = total_counts+k; 27489162d606SStefano Zampini PetscScalar *vals; 27499162d606SStefano Zampini 27509162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 27519162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 27529162d606SStefano Zampini } 27539162d606SStefano Zampini total_counts += constraints_n[i]; 2754674ae819SStefano Zampini } 2755674ae819SStefano Zampini } 2756674ae819SStefano Zampini /* assembling */ 2757674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2758674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2759088faed8SStefano Zampini 2760984c4197SStefano Zampini /* 276145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2762984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2763984c4197SStefano Zampini */ 2764674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2765674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2766026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2767984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2768984c4197SStefano Zampini /* working stuff for GEQRF */ 276981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2770984c4197SStefano Zampini PetscBLASInt lqr_work; 2771984c4197SStefano Zampini /* working stuff for UNGQR */ 2772984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2773984c4197SStefano Zampini PetscBLASInt lgqr_work; 2774984c4197SStefano Zampini /* working stuff for TRTRS */ 2775984c4197SStefano Zampini PetscScalar *trs_rhs; 27763f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2777984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2778984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2779984c4197SStefano Zampini PetscScalar *start_vals; 2780984c4197SStefano Zampini /* working stuff for values insertion */ 27814641a718SStefano Zampini PetscBT is_primal; 278264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2783906d46d4SStefano Zampini /* matrix sizes */ 2784906d46d4SStefano Zampini PetscInt global_size,local_size; 2785906d46d4SStefano Zampini /* temporary change of basis */ 2786906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2787cf5a6209SStefano Zampini /* extra space for debugging */ 2788cf5a6209SStefano Zampini PetscScalar *dbg_work; 2789984c4197SStefano Zampini 2790906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2791906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 279216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2793bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2794906d46d4SStefano Zampini /* nonzeros for local mat */ 2795bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2796bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 27979162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2798a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 27999162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2800a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 28019162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2802a717540cSStefano Zampini } else { 28039162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 28049162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2805a717540cSStefano Zampini } 2806a717540cSStefano Zampini } 2807a717540cSStefano Zampini } 2808906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2809bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2810a717540cSStefano Zampini /* Set initial identity in the matrix */ 2811bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2812906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2813a717540cSStefano Zampini } 2814a717540cSStefano Zampini 2815a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2816a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2817a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2818a717540cSStefano Zampini } 2819a717540cSStefano Zampini 2820a717540cSStefano Zampini 2821a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2822a717540cSStefano Zampini /* 2823a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2824a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2825a717540cSStefano Zampini 2826a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2827a717540cSStefano Zampini 2828a6b551f4SStefano 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) 2829a6b551f4SStefano Zampini 2830a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2831a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2832a717540cSStefano Zampini | ... | 2833a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2834a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2835a717540cSStefano Zampini 2836a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2837a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2838a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2839a6b551f4SStefano Zampini 2840a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2841a717540cSStefano Zampini */ 2842a717540cSStefano Zampini if (qr_needed) { 2843984c4197SStefano Zampini /* space to store Q */ 2844854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2845984c4197SStefano Zampini /* first we issue queries for optimal work */ 28463f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28473f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 28483f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2849984c4197SStefano Zampini lqr_work = -1; 28503f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2851984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2852984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2853785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2854984c4197SStefano Zampini lgqr_work = -1; 28553f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28563f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 28573f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 28583f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 28593f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 28603f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2861984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2862984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2863785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2864984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2865785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2866984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2867785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2868a717540cSStefano Zampini /* allocating workspace for check */ 2869a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2870cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2871a717540cSStefano Zampini } 2872a717540cSStefano Zampini } 2873984c4197SStefano Zampini /* array to store whether a node is primal or not */ 28744641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2875473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 28760e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 287739e2fb2aSStefano Zampini if (i != total_primal_vertices) { 287839e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 28794641a718SStefano Zampini } 288039e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 288139e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 288239e2fb2aSStefano Zampini } 288339e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2884984c4197SStefano Zampini 2885a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 28869162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 28879162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 28884641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2889984c4197SStefano Zampini /* get constraint info */ 28909162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2891984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2892984c4197SStefano Zampini 2893984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28949162d606SStefano 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); 2895674ae819SStefano Zampini } 2896984c4197SStefano Zampini 2897fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2898a717540cSStefano Zampini 2899a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2900a717540cSStefano Zampini if (pcbddc->dbg_flag) { 29019162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2902a717540cSStefano Zampini } 2903984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 29049162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2905984c4197SStefano Zampini 2906984c4197SStefano Zampini /* compute QR decomposition of constraints */ 29073f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29083f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29093f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2910674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29113f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2912984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2913674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2914984c4197SStefano Zampini 2915984c4197SStefano Zampini /* explictly compute R^-T */ 2916984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2917984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 29183f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29193f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 29203f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29213f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2922984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29233f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2924984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2925984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2926984c4197SStefano Zampini 2927a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 29283f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29303f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2932984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29333f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2934984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2935984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2936984c4197SStefano Zampini 2937984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2938984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2939984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 29403f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29413f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29423f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29433f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29443f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 29453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2946984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29479162d606SStefano 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)); 2948984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 29499162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2950984c4197SStefano Zampini 2951984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 29529162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2953984c4197SStefano Zampini /* insert cols for primal dofs */ 2954984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2955984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 29569162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2957906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2958984c4197SStefano Zampini } 2959984c4197SStefano Zampini /* insert cols for dual dofs */ 2960984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 29619162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2962984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 29639162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2964906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2965984c4197SStefano Zampini j++; 2966674ae819SStefano Zampini } 2967674ae819SStefano Zampini } 2968984c4197SStefano Zampini 2969984c4197SStefano Zampini /* check change of basis */ 2970984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2971984c4197SStefano Zampini PetscInt ii,jj; 2972984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2973c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2974c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2975c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2976c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2977c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2978c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2979984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2980cf5a6209SStefano 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)); 2981984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2982984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2983984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2984cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2985cf5a6209SStefano 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; 2986674ae819SStefano Zampini } 2987674ae819SStefano Zampini } 2988984c4197SStefano Zampini if (!valid_qr) { 298922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2990984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2991984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2992cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2993cf5a6209SStefano 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])); 2994674ae819SStefano Zampini } 2995cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2996cf5a6209SStefano 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])); 2997984c4197SStefano Zampini } 2998984c4197SStefano Zampini } 2999984c4197SStefano Zampini } 3000674ae819SStefano Zampini } else { 300122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3002674ae819SStefano Zampini } 3003674ae819SStefano Zampini } 3004a717540cSStefano Zampini } else { /* simple transformation block */ 3005a717540cSStefano Zampini PetscInt row,col; 3006a6b551f4SStefano Zampini PetscScalar val,norm; 3007a6b551f4SStefano Zampini 3008a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30099162d606SStefano 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)); 3010a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 30119162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 30129162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3013bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 30149162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3015906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 30169162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3017a717540cSStefano Zampini } else { 3018a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 30199162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3020a717540cSStefano Zampini if (row != col) { 30219162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3022a717540cSStefano Zampini } else { 30239162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3024a717540cSStefano Zampini } 3025906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3026a717540cSStefano Zampini } 3027a717540cSStefano Zampini } 3028a717540cSStefano Zampini } 302998a51de6SStefano Zampini if (pcbddc->dbg_flag) { 303022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3031a717540cSStefano Zampini } 3032674ae819SStefano Zampini } 3033984c4197SStefano Zampini } else { 3034984c4197SStefano Zampini if (pcbddc->dbg_flag) { 30359162d606SStefano 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); 3036674ae819SStefano Zampini } 3037674ae819SStefano Zampini } 3038674ae819SStefano Zampini } 3039a717540cSStefano Zampini 3040a717540cSStefano Zampini /* free workspace */ 3041a717540cSStefano Zampini if (qr_needed) { 3042984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3043cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3044984c4197SStefano Zampini } 3045984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3046984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3047984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3048984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3049984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3050674ae819SStefano Zampini } 3051a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3052906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3053906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3054906d46d4SStefano Zampini 3055906d46d4SStefano Zampini /* assembling of global change of variable */ 3056bbb9e6c6SStefano Zampini { 3057bbb9e6c6SStefano Zampini Mat tmat; 305816f15bc4SStefano Zampini PetscInt bs; 305916f15bc4SStefano Zampini 3060906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3061906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3062bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3063bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3064bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3065bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 306616f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 306716f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3068906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3069bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3070bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3071bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3072bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3073bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3074e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3075e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3076bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3077bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3078906d46d4SStefano Zampini } 3079906d46d4SStefano Zampini /* check */ 3080906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3081906d46d4SStefano Zampini PetscReal error; 3082906d46d4SStefano Zampini Vec x,x_change; 3083906d46d4SStefano Zampini 3084906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3085906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3086906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3087906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3088e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3089e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3090bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3091e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3092e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3093906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3094906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3095906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3096906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3097bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3098906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3099906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3100906d46d4SStefano Zampini } 3101b96c3477SStefano Zampini 3102b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3103b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3104b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3105b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3106ac632422SStefano Zampini Mat S_new,tmat; 3107bbb9e6c6SStefano Zampini IS is_all_N; 3108bbb9e6c6SStefano Zampini 3109bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 31106816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3111bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3112ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3113b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3114ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3115ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3116ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3117ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3118ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3119b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3120ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3121ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3122ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3123ac632422SStefano Zampini } 3124b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3125b96c3477SStefano Zampini } 3126b96c3477SStefano Zampini } 3127906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3128906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3129b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3130b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3131b9b85e73SStefano Zampini } 3132906d46d4SStefano Zampini 3133906d46d4SStefano Zampini /* set up change of basis context */ 3134906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3135906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3136906d46d4SStefano Zampini 3137906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3138906d46d4SStefano Zampini PetscInt global_size,local_size; 3139906d46d4SStefano Zampini 3140906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3141906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3142906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3143906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3144906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3145906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3146906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3147906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3148906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3149906d46d4SStefano Zampini } else { 3150906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3151906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3152906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3153906d46d4SStefano Zampini } 3154906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3155906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3156906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3157906d46d4SStefano Zampini } else { 3158906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3159906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3160906d46d4SStefano Zampini } 3161906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3162906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3163906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3164906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3165b9b85e73SStefano Zampini } 3166a717540cSStefano Zampini 3167019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3168019a44ceSStefano Zampini if (pcbddc->zerodiag) { 3169019a44ceSStefano Zampini const PetscInt *idxs; 3170019a44ceSStefano Zampini PetscInt nz; 3171019a44ceSStefano Zampini 3172019a44ceSStefano Zampini ierr = ISGetLocalSize(pcbddc->zerodiag,&nz);CHKERRQ(ierr); 3173019a44ceSStefano Zampini ierr = ISGetIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3174019a44ceSStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3175019a44ceSStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size_cc] = idxs[nz-1]; 3176019a44ceSStefano Zampini ierr = ISRestoreIndices(pcbddc->zerodiag,&idxs);CHKERRQ(ierr); 3177019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3178019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3179019a44ceSStefano Zampini pcbddc->local_primal_size++; 3180019a44ceSStefano Zampini } 3181019a44ceSStefano Zampini 3182019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3183727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3184727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3185aff50787SStefano 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); 3186c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 31870e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3188aff50787SStefano 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); 3189727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3190727cdba6SStefano Zampini } 31910e6343abSStefano Zampini } 31920e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3193727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3194727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3195727cdba6SStefano Zampini 3196a717540cSStefano Zampini /* flush dbg viewer */ 3197b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3198b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3199b8ffe317SStefano Zampini } 3200a717540cSStefano Zampini 3201e310c8b4SStefano Zampini /* free workspace */ 3202a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 32034641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 320408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 32059162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 32069162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 320708122e43SStefano Zampini } else { 32089162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 32099162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 32109162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 321108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 321208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 32139162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 32149162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 321508122e43SStefano Zampini } 3216674ae819SStefano Zampini PetscFunctionReturn(0); 3217674ae819SStefano Zampini } 3218674ae819SStefano Zampini 3219674ae819SStefano Zampini #undef __FUNCT__ 3220674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3221674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3222674ae819SStefano Zampini { 3223674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3224674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3225674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 32267fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3227674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3228674ae819SStefano Zampini 3229674ae819SStefano Zampini PetscFunctionBegin; 32308e61c736SStefano Zampini /* Reset previously computed graph */ 32318e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3232674ae819SStefano Zampini /* Init local Graph struct */ 32337fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 32343bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3235674ae819SStefano Zampini 3236575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3237575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3238575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3239575ad6abSStefano Zampini } 32409577ea80SStefano Zampini 3241674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 32424d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 32434d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 32444d379d7bSStefano Zampini PetscInt nvtxs; 3245e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3246674ae819SStefano Zampini 32474d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 32482fffb893SStefano Zampini 32492fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32502fffb893SStefano Zampini if (flg_row) { 32514d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3252b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 32532fffb893SStefano Zampini } 32542fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32559b28b941SStefano 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 */ 32564d379d7bSStefano Zampini IS is_dummy; 32574d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 32584d379d7bSStefano Zampini PetscInt j,sum; 32594d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 32604d379d7bSStefano Zampini const PetscInt *idxs; 32614d379d7bSStefano Zampini PCBDDCGraph graph; 32624d379d7bSStefano Zampini PetscBT is_on_boundary; 32634d379d7bSStefano Zampini 32644d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 32654d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 32664d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 32674d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 32687fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 32694d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3270e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3271e496cd5dSStefano Zampini if (flg_row) { 32724d379d7bSStefano Zampini graph->xadj = xadj; 32734d379d7bSStefano Zampini graph->adjncy = adjncy; 3274e496cd5dSStefano Zampini } 32754d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 32764d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3277e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32784d379d7bSStefano Zampini 32794d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 32809b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 32814d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32824d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 32834d379d7bSStefano Zampini } 32844d379d7bSStefano Zampini } 32854d379d7bSStefano Zampini 3286e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 32874d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 32884d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 32894d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 32904d379d7bSStefano Zampini } 32914d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 32924d379d7bSStefano Zampini 3293e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 32944d379d7bSStefano Zampini sum = 0; 32954d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 32964d379d7bSStefano Zampini PetscInt sizecc = 0; 32974d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 32984d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 32994d379d7bSStefano Zampini sizecc++; 33004d379d7bSStefano Zampini } 33014d379d7bSStefano Zampini } 33024d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33034d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33044d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 33054d379d7bSStefano Zampini } 33064d379d7bSStefano Zampini } 33074d379d7bSStefano Zampini sum += sizecc*sizecc; 33084d379d7bSStefano Zampini } 33094d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 33104d379d7bSStefano Zampini sum = 0; 3311e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 33124d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 33134d379d7bSStefano Zampini cxadj[i] = sum; 33144d379d7bSStefano Zampini sum += temp; 33154d379d7bSStefano Zampini } 3316e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 33174d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33184d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33194d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33204d379d7bSStefano Zampini PetscInt k,sizecc = 0; 33214d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 33224d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 33234d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 33244d379d7bSStefano Zampini sizecc++; 33254d379d7bSStefano Zampini } 33264d379d7bSStefano Zampini } 33274d379d7bSStefano Zampini } 33284d379d7bSStefano Zampini } 33294d379d7bSStefano Zampini } 33309b28b941SStefano Zampini if (sum) { 3331e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 33324d379d7bSStefano Zampini } else { 33334d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 33344d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 33354d379d7bSStefano Zampini } 33364d379d7bSStefano Zampini graph->xadj = 0; 33374d379d7bSStefano Zampini graph->adjncy = 0; 33384d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 33394d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 33404d379d7bSStefano Zampini } 3341674ae819SStefano Zampini } 33429b28b941SStefano Zampini if (pcbddc->dbg_flag) { 33439b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3344674ae819SStefano Zampini } 3345674ae819SStefano Zampini 334663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3347674ae819SStefano Zampini vertex_size = 1; 334863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 334963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 335095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 335163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3352e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 335363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3354674ae819SStefano Zampini } 335563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 335663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 335763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3358674ae819SStefano Zampini } 335963602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3360674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 336163602bcaSStefano Zampini } else { 336263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 336363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3364854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 336563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 336663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 336763602bcaSStefano Zampini } 336863602bcaSStefano Zampini } 3369674ae819SStefano Zampini } 3370674ae819SStefano Zampini 3371674ae819SStefano Zampini /* Setup of Graph */ 3372785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3373e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3374785d1243SStefano Zampini } 3375785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3376e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3377785d1243SStefano Zampini } 3378302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3379674ae819SStefano Zampini 3380674ae819SStefano Zampini /* Graph's connected components analysis */ 3381674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3382674ae819SStefano Zampini 3383674ae819SStefano Zampini /* print some info to stdout */ 3384674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3385302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3386674ae819SStefano Zampini } 3387fb180af4SStefano Zampini 3388fb180af4SStefano Zampini /* mark topography has done */ 3389fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3390674ae819SStefano Zampini PetscFunctionReturn(0); 3391674ae819SStefano Zampini } 3392674ae819SStefano Zampini 3393dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3394674ae819SStefano Zampini #undef __FUNCT__ 3395674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3396dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3397674ae819SStefano Zampini { 3398dc456d91SStefano Zampini PetscSF sf; 3399dc456d91SStefano Zampini PetscLayout map; 3400dc456d91SStefano Zampini const PetscInt *idxs; 3401dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3402dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3403dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3404dc456d91SStefano Zampini PetscMPIInt commsize; 3405674ae819SStefano Zampini PetscBool first_found; 3406674ae819SStefano Zampini PetscErrorCode ierr; 3407674ae819SStefano Zampini 3408674ae819SStefano Zampini PetscFunctionBegin; 3409dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3410dc456d91SStefano Zampini if (subset_mult) { 3411dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3412dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3413dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3414674ae819SStefano Zampini } 3415dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3416dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3417dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3418dc456d91SStefano Zampini for (i=0;i<n;i++) { 3419dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3420dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3421674ae819SStefano Zampini } 3422dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3423dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3424dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3425dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3426dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3427dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3428dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3429dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3430dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3431dc456d91SStefano Zampini 3432dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3433dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3434dc456d91SStefano Zampini if (subset_mult) { 3435dc456d91SStefano Zampini const PetscInt* idxs_mult; 3436dc456d91SStefano Zampini 3437dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3438dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3439dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3440674ae819SStefano Zampini } else { 3441dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3442674ae819SStefano Zampini } 3443dc456d91SStefano Zampini /* local size of new subset */ 3444dc456d91SStefano Zampini n_n = 0; 3445dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3446dc456d91SStefano Zampini 3447dc456d91SStefano Zampini /* global indexes in layout */ 3448dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3449dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3450dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3451dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3452dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3453dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3454dc456d91SStefano Zampini 3455dc456d91SStefano Zampini /* reduce from leaves to roots */ 3456dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 345764a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 345864a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3459dc456d91SStefano Zampini 3460dc456d91SStefano Zampini /* count indexes in local part of layout */ 3461674ae819SStefano Zampini nlocals = 0; 3462674ae819SStefano Zampini first_index = -1; 3463674ae819SStefano Zampini first_found = PETSC_FALSE; 3464dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3465dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3466674ae819SStefano Zampini first_found = PETSC_TRUE; 3467674ae819SStefano Zampini first_index = i; 3468674ae819SStefano Zampini } 3469dc456d91SStefano Zampini nlocals += root_data[i]; 3470674ae819SStefano Zampini } 3471dc456d91SStefano Zampini 3472dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 34735fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3474dc456d91SStefano Zampini start = 0; 347564a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 34765fa240b1SStefano Zampini #else 347764a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 34785fa240b1SStefano Zampini start = start-nlocals; 34795fa240b1SStefano Zampini #endif 34805fa240b1SStefano Zampini 3481dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3482dc456d91SStefano Zampini *N_n = start + nlocals; 3483dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3484dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3485674ae819SStefano Zampini } 34865fa240b1SStefano Zampini 34875fa240b1SStefano Zampini /* adapt root data with cumulative */ 3488674ae819SStefano Zampini if (first_found) { 3489dc456d91SStefano Zampini PetscInt old_index; 3490dc456d91SStefano Zampini 3491dc456d91SStefano Zampini root_data[first_index] += start; 3492674ae819SStefano Zampini old_index = first_index; 3493dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3494dc456d91SStefano Zampini if (root_data[i]) { 3495dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3496674ae819SStefano Zampini old_index = i; 3497674ae819SStefano Zampini } 3498674ae819SStefano Zampini } 3499674ae819SStefano Zampini } 3500dc456d91SStefano Zampini 3501dc456d91SStefano Zampini /* from roots to leaves */ 3502dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3503dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3504dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3505dc456d91SStefano Zampini 3506dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3507dc456d91SStefano Zampini if (subset_mult) { 3508dc456d91SStefano Zampini const PetscInt* idxs_mult; 3509dc456d91SStefano Zampini PetscInt cum; 3510dc456d91SStefano Zampini 3511dc456d91SStefano Zampini cum = 0; 3512dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3513dc456d91SStefano Zampini for (i=0;i<n;i++) { 3514dc456d91SStefano Zampini PetscInt j; 3515dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3516674ae819SStefano Zampini } 3517dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3518674ae819SStefano Zampini } else { 3519dc456d91SStefano Zampini for (i=0;i<n;i++) { 3520dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3521674ae819SStefano Zampini } 3522674ae819SStefano Zampini } 3523dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3524dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3525674ae819SStefano Zampini PetscFunctionReturn(0); 3526674ae819SStefano Zampini } 35279a7d3425SStefano Zampini 35289a7d3425SStefano Zampini #undef __FUNCT__ 35299a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 35309a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 35319a7d3425SStefano Zampini { 35329a7d3425SStefano Zampini PetscInt i,j; 35339a7d3425SStefano Zampini PetscScalar *alphas; 35349a7d3425SStefano Zampini PetscErrorCode ierr; 35359a7d3425SStefano Zampini 35369a7d3425SStefano Zampini PetscFunctionBegin; 35379a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3538785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 35399a7d3425SStefano Zampini for (i=0;i<n;i++) { 35409a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 35419a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 35429a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 35439a7d3425SStefano Zampini } 35449a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 35459a7d3425SStefano Zampini PetscFunctionReturn(0); 35469a7d3425SStefano Zampini } 35479a7d3425SStefano Zampini 3548e7931f94SStefano Zampini #undef __FUNCT__ 354970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3550b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3551e7931f94SStefano Zampini { 355252e5ac9dSStefano Zampini IS ranks_send_to; 3553e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3554e7931f94SStefano Zampini PetscMPIInt size,rank,color; 355552e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 355652e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 35573837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 35582b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3559e7931f94SStefano Zampini PetscSubcomm subcomm; 356052e5ac9dSStefano Zampini PetscErrorCode ierr; 3561a57a6d2fSStefano Zampini 3562e7931f94SStefano Zampini PetscFunctionBegin; 35632b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 35642b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 35652b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3566e7931f94SStefano Zampini 3567e7931f94SStefano Zampini /* Get info on mapping */ 35683bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 35693bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3570e7931f94SStefano Zampini 3571e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3572785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3573e7931f94SStefano Zampini xadj[0] = 0; 3574e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3575785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3576785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3577e7931f94SStefano Zampini 35782b510759SStefano Zampini if (threshold) { 3579d023bfaeSStefano Zampini PetscInt xadj_count = 0; 35802b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3581d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3582d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3583d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3584d023bfaeSStefano Zampini xadj_count++; 3585e7931f94SStefano Zampini } 3586e7931f94SStefano Zampini } 3587d023bfaeSStefano Zampini xadj[1] = xadj_count; 3588c8587f34SStefano Zampini } else { 3589e7931f94SStefano Zampini if (xadj[1]) { 3590e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3591e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3592c8587f34SStefano Zampini } 3593e7931f94SStefano Zampini } 35943bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3595e7931f94SStefano Zampini if (use_square) { 3596e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3597e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3598e7931f94SStefano Zampini } 3599e7931f94SStefano Zampini } 3600e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3601e7931f94SStefano Zampini 36023837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3603e7931f94SStefano Zampini 3604e7931f94SStefano Zampini /* 3605e7931f94SStefano Zampini Restrict work on active processes only. 3606e7931f94SStefano Zampini */ 3607e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3608e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3609e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 36102b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3611d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3612e7931f94SStefano Zampini if (color) { 3613e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3614e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3615e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3616c8587f34SStefano Zampini } else { 361752e5ac9dSStefano Zampini Mat subdomain_adj; 361852e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 361952e5ac9dSStefano Zampini MatPartitioning partitioner; 362052e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 362152e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3622b0c7d250SStefano Zampini PetscBool aggregate; 3623b0c7d250SStefano Zampini 3624306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3625785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3626e7931f94SStefano Zampini prank = rank; 3627306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 36288002ef2cSStefano Zampini /* 3629e7931f94SStefano Zampini for (i=0;i<size;i++) { 3630e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3631c8587f34SStefano Zampini } 36328002ef2cSStefano Zampini */ 3633e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3634e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3635c8587f34SStefano Zampini } 3636e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3637b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3638b0c7d250SStefano Zampini if (aggregate) { 3639b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3640b0c7d250SStefano Zampini PetscMPIInt nrank; 3641b0c7d250SStefano Zampini PetscScalar *vals; 3642b0c7d250SStefano Zampini 3643b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3644b0c7d250SStefano Zampini lrows = 0; 3645b0c7d250SStefano Zampini if (nrank<redprocs) { 3646b0c7d250SStefano Zampini lrows = size/redprocs; 3647b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3648b0c7d250SStefano Zampini } 36495fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3650b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3651b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3652b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3653b0c7d250SStefano Zampini row = nrank; 3654b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3655b0c7d250SStefano Zampini cols = adjncy; 3656b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3657b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3658b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3659b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3660b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 366152e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 366252e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 366352e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3664b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3665b0c7d250SStefano Zampini } else { 3666306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3667b0c7d250SStefano Zampini } 366822b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3669e7931f94SStefano Zampini 3670e7931f94SStefano Zampini /* Partition */ 3671306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3672e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3673e7931f94SStefano Zampini if (use_vwgt) { 36743837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3675e7931f94SStefano Zampini v_wgt[0] = local_size; 3676e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3677c8587f34SStefano Zampini } 367828143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 367928143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3680e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3681e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 368222b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3683e7931f94SStefano Zampini 368452e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 368552e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 368652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 368752e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3688b0c7d250SStefano Zampini if (!redprocs) { 3689b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 369028143c3dSStefano Zampini } else { 3691b0c7d250SStefano Zampini PetscInt idxs[1]; 3692b0c7d250SStefano Zampini PetscMPIInt tag; 3693b0c7d250SStefano Zampini MPI_Request *reqs; 3694b0c7d250SStefano Zampini 3695b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3696b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3697b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3698b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 369928143c3dSStefano Zampini } 3700b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3701b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3702b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3703b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3704e7931f94SStefano Zampini } 370552e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3706e7931f94SStefano Zampini /* clean up */ 3707e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 370852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3709e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3710e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3711e7931f94SStefano Zampini } 3712e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3713e7931f94SStefano Zampini 3714e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3715e7931f94SStefano Zampini i = 1; 3716e7931f94SStefano Zampini if (color) i=0; 3717e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3718e7931f94SStefano Zampini /* get back IS */ 3719e7931f94SStefano Zampini *is_sends = ranks_send_to; 3720e7931f94SStefano Zampini PetscFunctionReturn(0); 3721e7931f94SStefano Zampini } 3722e7931f94SStefano Zampini 3723e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3724e7931f94SStefano Zampini 3725e7931f94SStefano Zampini #undef __FUNCT__ 3726e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 372753a05cb3SStefano 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[]) 3728e7931f94SStefano Zampini { 372970cf5478SStefano Zampini Mat local_mat; 3730e7931f94SStefano Zampini IS is_sends_internal; 37319d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 373228143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 37339d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3734e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3735e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3736e7931f94SStefano Zampini const PetscInt* is_indices; 3737e7931f94SStefano Zampini MatType new_local_type; 3738e7931f94SStefano Zampini /* buffers */ 3739e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 374028143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 37419d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3742e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3743e7931f94SStefano Zampini /* MPI */ 374428143c3dSStefano Zampini MPI_Comm comm,comm_n; 374528143c3dSStefano Zampini PetscSubcomm subcomm; 3746e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 374728143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 374828143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 374928143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 375028143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 375128143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3752e7931f94SStefano Zampini PetscErrorCode ierr; 3753e7931f94SStefano Zampini 3754e7931f94SStefano Zampini PetscFunctionBegin; 375528143c3dSStefano Zampini /* TODO: add missing checks */ 375628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 375728143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 375828143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 375928143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3760e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 376128143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3762e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3763e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3764e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3765e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3766e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 376728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 376870cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 376970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 377028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 377170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 377270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 377370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 377470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 377570cf5478SStefano Zampini } 3776e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3777e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3778e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3779e7931f94SStefano Zampini if (!is_sends) { 378028143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3781b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3782c8587f34SStefano Zampini } else { 3783e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3784e7931f94SStefano Zampini is_sends_internal = is_sends; 3785c8587f34SStefano Zampini } 3786e7931f94SStefano Zampini 3787e7931f94SStefano Zampini /* get comm */ 3788a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3789e7931f94SStefano Zampini 3790e7931f94SStefano Zampini /* compute number of sends */ 3791e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3792e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3793e7931f94SStefano Zampini 3794e7931f94SStefano Zampini /* compute number of receives */ 3795e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3796785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3797e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3798e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3799e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3800e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3801e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3802e7931f94SStefano Zampini 380328143c3dSStefano Zampini /* restrict comm if requested */ 380428143c3dSStefano Zampini subcomm = 0; 380528143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 380628143c3dSStefano Zampini if (restrict_comm) { 3807779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3808779c1cceSStefano Zampini 380928143c3dSStefano Zampini color = 0; 381053a05cb3SStefano Zampini if (restrict_full) { 381153a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 381253a05cb3SStefano Zampini } else { 381353a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 381453a05cb3SStefano Zampini } 381528143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 381628143c3dSStefano Zampini subcommsize = commsize - subcommsize; 381728143c3dSStefano Zampini /* check if reuse has been requested */ 381828143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 381928143c3dSStefano Zampini if (*mat_n) { 382028143c3dSStefano Zampini PetscMPIInt subcommsize2; 382128143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 382228143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 382328143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 382428143c3dSStefano Zampini } else { 382528143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 382628143c3dSStefano Zampini } 382728143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3828779c1cceSStefano Zampini PetscMPIInt rank; 3829779c1cceSStefano Zampini 3830779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 383128143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 383228143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 383328143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3834306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 383528143c3dSStefano Zampini } 383628143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 383728143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 383828143c3dSStefano Zampini } else { 383928143c3dSStefano Zampini comm_n = comm; 384028143c3dSStefano Zampini } 384128143c3dSStefano Zampini 3842e7931f94SStefano Zampini /* prepare send/receive buffers */ 3843785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3844e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3845785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3846e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 384728143c3dSStefano Zampini if (nis) { 3848854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 384928143c3dSStefano Zampini } 3850e7931f94SStefano Zampini 385128143c3dSStefano Zampini /* Get data from local matrices */ 3852e7931f94SStefano Zampini if (!isdense) { 3853a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3854e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3855e7931f94SStefano Zampini /* 3856e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3857e7931f94SStefano Zampini send_buffer_idxs should contain: 3858e7931f94SStefano Zampini - MatType_PRIVATE type 3859e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3860e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3861e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3862e7931f94SStefano Zampini */ 3863e7931f94SStefano Zampini } else { 3864e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 38653bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3866854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3867e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3868e7931f94SStefano Zampini send_buffer_idxs[1] = i; 38693bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3870e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 38713bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3872e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3873e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3874e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3875e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3876c8587f34SStefano Zampini } 3877c8587f34SStefano Zampini } 3878e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 387928143c3dSStefano Zampini /* additional is (if any) */ 388028143c3dSStefano Zampini if (nis) { 388128143c3dSStefano Zampini PetscMPIInt psum; 388228143c3dSStefano Zampini PetscInt j; 388328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 388428143c3dSStefano Zampini PetscInt plen; 388528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 388628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 388728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 388828143c3dSStefano Zampini } 3889854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 389028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 389128143c3dSStefano Zampini PetscInt plen; 389228143c3dSStefano Zampini const PetscInt *is_array_idxs; 389328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 389428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 389528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 389628143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 389728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 389828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 389928143c3dSStefano Zampini } 390028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 390128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 390228143c3dSStefano Zampini } 390328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 390428143c3dSStefano Zampini } 390528143c3dSStefano Zampini 3906e7931f94SStefano Zampini buf_size_idxs = 0; 3907e7931f94SStefano Zampini buf_size_vals = 0; 390828143c3dSStefano Zampini buf_size_idxs_is = 0; 3909e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3910e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3911e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 391228143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3913e7931f94SStefano Zampini } 3914785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3915785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 391695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3917e7931f94SStefano Zampini 3918e7931f94SStefano Zampini /* get new tags for clean communications */ 3919e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3920e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 392128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3922e7931f94SStefano Zampini 3923e7931f94SStefano Zampini /* allocate for requests */ 3924785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3925785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 392695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3927785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3928785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 392995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3930e7931f94SStefano Zampini 3931e7931f94SStefano Zampini /* communications */ 3932e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3933e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 393428143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3935e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3936e7931f94SStefano Zampini source_dest = onodes[i]; 3937e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3938e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3939e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3940e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 394128143c3dSStefano Zampini if (nis) { 394228143c3dSStefano 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); 394328143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 394428143c3dSStefano Zampini } 3945e7931f94SStefano Zampini } 3946e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3947e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3948e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3949e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 395028143c3dSStefano Zampini if (nis) { 395128143c3dSStefano 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); 395228143c3dSStefano Zampini } 3953e7931f94SStefano Zampini } 3954e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3955e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3956e7931f94SStefano Zampini 3957e7931f94SStefano Zampini /* assemble new l2g map */ 3958e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3959e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39609d30be91SStefano Zampini new_local_rows = 0; 3961e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 39629d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3963e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3964e7931f94SStefano Zampini } 39659d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3966e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39679d30be91SStefano Zampini new_local_rows = 0; 3968e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 39699d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 39709d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3971e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3972e7931f94SStefano Zampini } 39739d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 39749d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3975e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3976e7931f94SStefano Zampini 3977e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3978e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3979e7931f94SStefano 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) */ 3980e7931f94SStefano Zampini if (n_recvs) { 398128143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3982e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3983e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3984e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3985e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3986e7931f94SStefano Zampini break; 3987e7931f94SStefano Zampini } 3988e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3989e7931f94SStefano Zampini } 3990e7931f94SStefano Zampini switch (new_local_type_private) { 399128143c3dSStefano Zampini case MATDENSE_PRIVATE: 399228143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3993e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3994e7931f94SStefano Zampini bs = 1; 399528143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 399628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 399728143c3dSStefano Zampini bs = 1; 399828143c3dSStefano Zampini } 3999e7931f94SStefano Zampini break; 4000e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4001e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4002e7931f94SStefano Zampini bs = 1; 4003e7931f94SStefano Zampini break; 4004e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4005e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4006e7931f94SStefano Zampini break; 4007e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4008e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4009e7931f94SStefano Zampini break; 4010e7931f94SStefano Zampini default: 40119d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4012e7931f94SStefano Zampini break; 4013e7931f94SStefano Zampini } 401428143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 401528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 401628143c3dSStefano Zampini bs = 1; 4017e7931f94SStefano Zampini } 4018e7931f94SStefano Zampini 401970cf5478SStefano Zampini /* create MATIS object if needed */ 402070cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4021e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4022e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 402370cf5478SStefano Zampini } else { 402470cf5478SStefano Zampini /* it also destroys the local matrices */ 402570cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 402670cf5478SStefano Zampini } 402770cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4028e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 40299d30be91SStefano Zampini 40309d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 40319d30be91SStefano Zampini 40329d30be91SStefano Zampini /* Global to local map of received indices */ 40339d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 40349d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 40359d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 40369d30be91SStefano Zampini 40379d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 40389d30be91SStefano Zampini buf_size_idxs = 0; 40399d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40409d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 40419d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 40429d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 40439d30be91SStefano Zampini } 40449d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 40459d30be91SStefano Zampini 40469d30be91SStefano Zampini /* set preallocation */ 40479d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 40489d30be91SStefano Zampini if (!newisdense) { 40499d30be91SStefano Zampini PetscInt *new_local_nnz=0; 40509d30be91SStefano Zampini 40519d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 40529d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 40539d30be91SStefano Zampini if (n_recvs) { 40549d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 40559d30be91SStefano Zampini } 40569d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40579d30be91SStefano Zampini PetscInt j; 40589d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 40599d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 40609d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 40619d30be91SStefano Zampini } 40629d30be91SStefano Zampini } else { 40639d30be91SStefano Zampini /* TODO */ 40649d30be91SStefano Zampini } 40659d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 40669d30be91SStefano Zampini } 40679d30be91SStefano Zampini if (new_local_nnz) { 40689d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 40699d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 40709d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 40719d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 40729d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 40739d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 40749d30be91SStefano Zampini } else { 40759d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 40769d30be91SStefano Zampini } 40779d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 40789d30be91SStefano Zampini } else { 40799d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 40809d30be91SStefano Zampini } 4081e7931f94SStefano Zampini 4082e7931f94SStefano Zampini /* set values */ 4083e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 40849d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4085e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4086e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4087e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 40889d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4089e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4090e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4091e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 409228143c3dSStefano Zampini } else { 409328143c3dSStefano Zampini /* TODO */ 4094e7931f94SStefano Zampini } 4095e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4096e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4097e7931f94SStefano Zampini } 4098e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4099e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 410070cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 410170cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41029d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 41039d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4104e7931f94SStefano Zampini 4105dfd14d43SStefano Zampini #if 0 410628143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4107e7931f94SStefano Zampini Vec lvec,rvec; 4108e7931f94SStefano Zampini PetscReal infty_error; 4109e7931f94SStefano Zampini 41102a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4111e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4112e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4113e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 411470cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4115e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4116e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4117e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4118e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4119e7931f94SStefano Zampini } 412028143c3dSStefano Zampini #endif 4121e7931f94SStefano Zampini 412228143c3dSStefano Zampini /* assemble new additional is (if any) */ 412328143c3dSStefano Zampini if (nis) { 412428143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 412528143c3dSStefano Zampini 412628143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4127854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 412828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 412928143c3dSStefano Zampini psum = 0; 413028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 413128143c3dSStefano Zampini for (j=0;j<nis;j++) { 413228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 413328143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 413428143c3dSStefano Zampini psum += plen; 413528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 413628143c3dSStefano Zampini } 413728143c3dSStefano Zampini } 4138854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4139854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 414028143c3dSStefano Zampini for (i=1;i<nis;i++) { 414128143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 414228143c3dSStefano Zampini } 414328143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 414428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 414528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 414628143c3dSStefano Zampini for (j=0;j<nis;j++) { 414728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 414828143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 414928143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 415028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 415128143c3dSStefano Zampini } 415228143c3dSStefano Zampini } 415328143c3dSStefano Zampini for (i=0;i<nis;i++) { 415428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 415528143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 415628143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 415728143c3dSStefano Zampini } 415828143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 415928143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 416028143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 416128143c3dSStefano Zampini } 4162e7931f94SStefano Zampini /* free workspace */ 416328143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4164e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4165e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4166e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4167e7931f94SStefano Zampini if (isdense) { 4168e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4169e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4170e7931f94SStefano Zampini } else { 4171e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4172e7931f94SStefano Zampini } 417328143c3dSStefano Zampini if (nis) { 417428143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 417528143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 417628143c3dSStefano Zampini } 4177e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4178e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 417928143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4180e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4181e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 418228143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4183e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4184e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4185e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4186e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4187e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 418828143c3dSStefano Zampini if (nis) { 418928143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 419028143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 419128143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 419228143c3dSStefano Zampini } 419328143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 419428143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 419528143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 419628143c3dSStefano Zampini for (i=0;i<nis;i++) { 419728143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 419828143c3dSStefano Zampini } 419953a05cb3SStefano Zampini *mat_n = NULL; 420028143c3dSStefano Zampini } 4201e7931f94SStefano Zampini PetscFunctionReturn(0); 4202e7931f94SStefano Zampini } 4203a57a6d2fSStefano Zampini 420412edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4205af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 420612edc857SStefano Zampini 4207c8587f34SStefano Zampini #undef __FUNCT__ 4208c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4209c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4210c8587f34SStefano Zampini { 4211c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4212c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 421320a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 42149881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 421520a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 42166e683305SStefano Zampini IS coarse_is,*isarray; 42176e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 42186e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4219f9eb5b7dSStefano Zampini PC pc_temp; 4220c8587f34SStefano Zampini PCType coarse_pc_type; 4221c8587f34SStefano Zampini KSPType coarse_ksp_type; 4222f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 42234f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 42246e683305SStefano Zampini Mat t_coarse_mat_is; 42256e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 42266e683305SStefano Zampini PetscMPIInt all_procs; 422774e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 422868457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 422922bc73bbSStefano Zampini PetscScalar *array; 42309881197aSStefano Zampini PetscErrorCode ierr; 4231fdc09c96SStefano Zampini 4232c8587f34SStefano Zampini PetscFunctionBegin; 4233c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 423468457ee5SStefano 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 */ 4235fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 42365a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4237fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4238f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4239f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4240f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4241fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 424251bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 424351bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4244dc4bcba2SStefano Zampini PC pc; 4245dc4bcba2SStefano Zampini PetscBool isbddc; 4246dc4bcba2SStefano Zampini 4247dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4248dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4249dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4250dc4bcba2SStefano Zampini if (isbddc) { 4251dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4252dc4bcba2SStefano Zampini } 4253727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4254fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4255fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4256fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4257f4ddd8eeSStefano Zampini } 4258fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4259fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4260f4ddd8eeSStefano Zampini } 426170cf5478SStefano Zampini /* reset any subassembling information */ 426270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 42636e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42646e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4265fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4266f4ddd8eeSStefano Zampini } 4267c8587f34SStefano Zampini 42686e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 42692b510759SStefano Zampini im_active = !!(pcis->n); 42702b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 42716e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 42726e683305SStefano Zampini void_procs = all_procs-active_procs; 42736e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 427474e2c79eSStefano Zampini redist = PETSC_FALSE; 427522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 42766e683305SStefano Zampini csin_ml = PETSC_TRUE; 42776e683305SStefano Zampini ncoarse_ml = void_procs; 4278779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4279779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 42806e683305SStefano Zampini csin_ds = PETSC_TRUE; 428118a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 428218a45a71SStefano Zampini redist = PETSC_TRUE; 428318a45a71SStefano Zampini } else { 42846e683305SStefano Zampini csin_ds = PETSC_TRUE; 4285779c1cceSStefano Zampini ncoarse_ds = active_procs; 4286779c1cceSStefano Zampini redist = PETSC_TRUE; 428718a45a71SStefano Zampini } 42886e683305SStefano Zampini } else { 42896e683305SStefano Zampini csin_ml = PETSC_FALSE; 42906e683305SStefano Zampini ncoarse_ml = all_procs; 42916e683305SStefano Zampini if (void_procs) { 42926e683305SStefano Zampini csin_ds = PETSC_TRUE; 42936e683305SStefano Zampini ncoarse_ds = void_procs; 42946e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 42956e683305SStefano Zampini } else { 4296779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 429774e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 429874e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 429974e2c79eSStefano Zampini redist = PETSC_TRUE; 430074e2c79eSStefano Zampini } else { 43016e683305SStefano Zampini csin_ds = PETSC_FALSE; 43026e683305SStefano Zampini ncoarse_ds = all_procs; 43036e683305SStefano Zampini } 43046e683305SStefano Zampini } 430574e2c79eSStefano Zampini } 43066e683305SStefano Zampini 43076e683305SStefano Zampini /* 43086e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 43096e683305SStefano Zampini - we have not exceeded the number of levels requested 43106e683305SStefano Zampini - we can actually subassemble the active processes 43116e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 43126e683305SStefano Zampini */ 43136e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 43146e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 43156e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 43166e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 43176e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4318f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 43192b510759SStefano Zampini } else { 4320f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4321c8587f34SStefano Zampini } 4322c8587f34SStefano Zampini } 43236e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 43246e683305SStefano Zampini if (multilevel_allowed) { 43256e683305SStefano Zampini ncoarse = ncoarse_ml; 43266e683305SStefano Zampini csin = csin_ml; 432758da7f69SStefano Zampini redist = PETSC_FALSE; 43286e683305SStefano Zampini } else { 43296e683305SStefano Zampini ncoarse = ncoarse_ds; 43306e683305SStefano Zampini csin = csin_ds; 43316e683305SStefano Zampini } 4332e7931f94SStefano Zampini 4333abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4334abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4335abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4336abbbba34SStefano Zampini 4337abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 433822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 433922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 434022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 434122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4342b9b85e73SStefano Zampini #if 0 4343b9b85e73SStefano Zampini { 4344b9b85e73SStefano Zampini PetscViewer viewer; 4345b9b85e73SStefano Zampini char filename[256]; 4346b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4347b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4348b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4349b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4350b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4351b9b85e73SStefano Zampini } 4352b9b85e73SStefano Zampini #endif 4353e176bc59SStefano 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); 43546e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 43556e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43566e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4357abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4358abbbba34SStefano Zampini 43596e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 43606e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 43616e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 43626e683305SStefano Zampini const PetscInt *idxs; 43636e683305SStefano Zampini ISLocalToGlobalMapping tmap; 43646e683305SStefano Zampini 43656e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 43660be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 43676e683305SStefano Zampini /* allocate space for temporary storage */ 4368854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4369854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 43706e683305SStefano Zampini /* allocate for IS array */ 43716e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 43726e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 43736e683305SStefano Zampini nis = nisdofs + nisneu; 4374854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 43756e683305SStefano Zampini /* dofs splitting */ 43766e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 43776e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 43786e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 43796e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 43806e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 43816e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 43826e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 43836e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 43846e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 43856e683305SStefano Zampini } 43866e683305SStefano Zampini /* neumann boundaries */ 43876e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 43886e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 43896e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 43906e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 43916e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 43926e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 43936e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 43946e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 43956e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 43966e683305SStefano Zampini } 43976e683305SStefano Zampini /* free memory */ 43986e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 43996e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 44006e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 44016e683305SStefano Zampini } else { 44026e683305SStefano Zampini nis = 0; 44036e683305SStefano Zampini nisdofs = 0; 44046e683305SStefano Zampini nisneu = 0; 44056e683305SStefano Zampini isarray = NULL; 44066e683305SStefano Zampini } 44076e683305SStefano Zampini /* destroy no longer needed map */ 44086e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 44096e683305SStefano Zampini 44106e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 44116e683305SStefano Zampini coarse_mat_is = NULL; 44126e683305SStefano Zampini if (csin) { 44136e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 441474e2c79eSStefano Zampini if (redist) { 441574e2c79eSStefano Zampini PetscMPIInt rank; 4416779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 441774e2c79eSStefano Zampini 441874e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 441958da7f69SStefano Zampini spc = active_procs/ncoarse; 442058da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4421779c1cceSStefano Zampini if (im_active) { 4422779c1cceSStefano Zampini destsize = 1; 442374e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 442474e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 442574e2c79eSStefano Zampini } else { 442674e2c79eSStefano Zampini dest[0] = rank/(spc+1); 442774e2c79eSStefano Zampini } 442874e2c79eSStefano Zampini } else { 4429779c1cceSStefano Zampini destsize = 0; 44306e683305SStefano Zampini } 4431779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4432779c1cceSStefano Zampini } else if (csin_type_simple) { 44336e683305SStefano Zampini PetscMPIInt rank; 44346e683305SStefano Zampini PetscInt issize,isidx; 4435779c1cceSStefano Zampini 44366e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 44376e683305SStefano Zampini if (im_active) { 44386e683305SStefano Zampini issize = 1; 44396e683305SStefano Zampini isidx = (PetscInt)rank; 44406e683305SStefano Zampini } else { 44416e683305SStefano Zampini issize = 0; 44426e683305SStefano Zampini isidx = -1; 44436e683305SStefano Zampini } 44446e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4445779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4446b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 44476e683305SStefano Zampini } 4448779c1cceSStefano Zampini 4449779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4450779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4451779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4452779c1cceSStefano Zampini PetscInt *coarse_candidates; 4453779c1cceSStefano Zampini const PetscInt* tisindices; 4454779c1cceSStefano Zampini 4455779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4456779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4457779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4458779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4459779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4460779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4461779c1cceSStefano Zampini } 4462779c1cceSStefano Zampini } 4463779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4464779c1cceSStefano Zampini 4465779c1cceSStefano Zampini 44666e683305SStefano Zampini if (pcbddc->dbg_flag) { 44676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 44696e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 44706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4471779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 44726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 44736e683305SStefano Zampini } 44746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 44756e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44766e683305SStefano Zampini } 44776e683305SStefano Zampini /* shift the pattern on coarse candidates */ 44786e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 44796e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4480854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 44816e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 44826e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 44836e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 44846e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 44856e683305SStefano Zampini } 44866e683305SStefano Zampini if (pcbddc->dbg_flag) { 44876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 44896e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 44906e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 44916e683305SStefano Zampini } 4492779c1cceSStefano Zampini } 44936e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 449453a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 449553a05cb3SStefano 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); 449653a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 449753a05cb3SStefano 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); 449853a05cb3SStefano Zampini } 44996e683305SStefano Zampini } else { 45006e683305SStefano Zampini if (pcbddc->dbg_flag) { 45016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 45036e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45046e683305SStefano Zampini } 45056e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 45066e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 45076e683305SStefano Zampini } 45086e683305SStefano Zampini 45096e683305SStefano Zampini /* create local to global scatters for coarse problem */ 451068457ee5SStefano Zampini if (compute_vecs) { 45116e683305SStefano Zampini PetscInt lrows; 45126e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 45136e683305SStefano Zampini if (coarse_mat_is) { 45146e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 45156e683305SStefano Zampini } else { 45166e683305SStefano Zampini lrows = 0; 45176e683305SStefano Zampini } 45186e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 45196e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 45206e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 45216e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45226e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45236e683305SStefano Zampini } 45246e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 45256e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4526c8587f34SStefano Zampini 4527f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4528f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4529f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4530f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4531f9eb5b7dSStefano Zampini } else { 4532f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4533f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4534c8587f34SStefano Zampini } 4535c8587f34SStefano Zampini 45366e683305SStefano Zampini /* print some info if requested */ 45376e683305SStefano Zampini if (pcbddc->dbg_flag) { 45386e683305SStefano Zampini if (!multilevel_allowed) { 45396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45406e683305SStefano Zampini if (multilevel_requested) { 45416e683305SStefano 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); 45426e683305SStefano Zampini } else if (pcbddc->max_levels) { 45436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 45446e683305SStefano Zampini } 45456e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45466e683305SStefano Zampini } 45476e683305SStefano Zampini } 45486e683305SStefano Zampini 4549f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 45506e683305SStefano Zampini if (coarse_mat_is) { 45516e683305SStefano Zampini MatReuse coarse_mat_reuse; 45526a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 45536e683305SStefano Zampini if (pcbddc->dbg_flag) { 45546e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 45556e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 45566e683305SStefano Zampini } 4557f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4558312be037SStefano Zampini char prefix[256],str_level[16]; 4559e604994aSStefano Zampini size_t len; 45606e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4561422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4562c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4563f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 45645f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4565c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 45666e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4567c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4568c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4569e604994aSStefano Zampini /* prefix */ 4570e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4571e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4572e604994aSStefano Zampini if (!pcbddc->current_level) { 4573e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4574e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4575c8587f34SStefano Zampini } else { 4576e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4577312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4578312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 457934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4580312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4581e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4582e604994aSStefano Zampini } 4583e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 45843e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 45853e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 45863e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 45873e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4588f9eb5b7dSStefano Zampini /* allow user customization */ 4589f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 45903e3c6dadSStefano Zampini } 45913e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 459251bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 45933e3c6dadSStefano Zampini if (nisdofs) { 45943e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 45953e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 45963e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 45973e3c6dadSStefano Zampini } 45983e3c6dadSStefano Zampini } 45993e3c6dadSStefano Zampini if (nisneu) { 46003e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 46013e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4602312be037SStefano Zampini } 4603f9eb5b7dSStefano Zampini 4604f9eb5b7dSStefano Zampini /* get some info after set from options */ 4605f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4606f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 46074f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 46086e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4609f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4610f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4611f9eb5b7dSStefano Zampini } 461239f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 46134f3a063dSStefano Zampini if (isredundant) { 46144f3a063dSStefano Zampini KSP inner_ksp; 46154f3a063dSStefano Zampini PC inner_pc; 46164f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 46174f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 46184f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 46194f3a063dSStefano Zampini } 4620f9eb5b7dSStefano Zampini 4621f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4622fa7f1dd8SStefano Zampini if (coarse_reuse) { 462381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4624fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 46256e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4626fa7f1dd8SStefano Zampini } else { 46276e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4628fa7f1dd8SStefano Zampini } 4629c8587f34SStefano Zampini if (isbddc || isnn) { 463022bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 463170cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4632b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 463322b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 46346e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 46366e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 46376e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 463822b6e8a2SStefano Zampini } 463970cf5478SStefano Zampini } 464053a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 464170cf5478SStefano Zampini } else { 464222bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 464322bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 464422bc73bbSStefano Zampini } 464522bc73bbSStefano Zampini } else { 46462e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4647c8587f34SStefano Zampini } 4648c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4649c8587f34SStefano Zampini 46503301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 46515a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 46523301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 46533301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 46543301b35fSStefano Zampini } 46553301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 46563301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 46573301b35fSStefano Zampini } 46583301b35fSStefano Zampini if (pc->pmat->spd_set) { 46593301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 46603301b35fSStefano Zampini } 46616e683305SStefano Zampini /* set operators */ 46625f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 46636e683305SStefano Zampini if (pcbddc->dbg_flag) { 46646e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 46656e683305SStefano Zampini } 46666e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 46676e683305SStefano Zampini coarse_mat = 0; 46686e683305SStefano Zampini } 46696e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4670b9b85e73SStefano Zampini #if 0 4671b9b85e73SStefano Zampini { 4672b9b85e73SStefano Zampini PetscViewer viewer; 4673b9b85e73SStefano Zampini char filename[256]; 4674b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4675b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4676b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4677b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4678b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4679b9b85e73SStefano Zampini } 4680b9b85e73SStefano Zampini #endif 4681c8587f34SStefano Zampini 4682c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4683298c0119SStefano Zampini #if 0 4684c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4685c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 468698a51de6SStefano Zampini } 4687298c0119SStefano Zampini #endif 468898a51de6SStefano Zampini 468998a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 469098a51de6SStefano Zampini Vec crhs,csol; 469198a51de6SStefano Zampini PetscBool ispreonly; 469204708bb6SStefano Zampini 469398a51de6SStefano Zampini if (CoarseNullSpace) { 4694c8587f34SStefano Zampini if (isbddc) { 4695c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4696c8587f34SStefano Zampini } else { 46975fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4698c8587f34SStefano Zampini } 4699c8587f34SStefano Zampini } 4700f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4701f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4702f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4703f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 47046e683305SStefano Zampini /* hack */ 4705f347579bSStefano Zampini if (!csol) { 47062a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4707f9eb5b7dSStefano Zampini } 4708f347579bSStefano Zampini if (!crhs) { 47092a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4710f347579bSStefano Zampini } 4711cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4712cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 47136e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4714c8587f34SStefano Zampini KSP check_ksp; 47152b510759SStefano Zampini KSPType check_ksp_type; 4716c8587f34SStefano Zampini PC check_pc; 47176e683305SStefano Zampini Vec check_vec,coarse_vec; 47186a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 47192b510759SStefano Zampini PetscInt its; 47206e683305SStefano Zampini PetscBool compute_eigs; 47216e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 47226e683305SStefano Zampini PetscInt neigs; 47238e185a42SStefano Zampini const char *prefix; 4724c8587f34SStefano Zampini 47252b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 47266e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4727422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 472823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4729f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 47302b510759SStefano Zampini if (ispreonly) { 47312b510759SStefano Zampini check_ksp_type = KSPPREONLY; 47326e683305SStefano Zampini compute_eigs = PETSC_FALSE; 47332b510759SStefano Zampini } else { 4734cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 47356e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4736c8587f34SStefano Zampini } 4737c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 47386e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 47396e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 47406e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4741a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4742a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4743a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4744a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4745c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4746c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4747c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4748c8587f34SStefano Zampini /* create random vec */ 47496e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 47506e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4751c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4752c8587f34SStefano Zampini if (CoarseNullSpace) { 4753c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4754c8587f34SStefano Zampini } 47556e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4756c8587f34SStefano Zampini /* solve coarse problem */ 47576e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4758c8587f34SStefano Zampini if (CoarseNullSpace) { 47596e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4760c8587f34SStefano Zampini } 4761cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 47626e683305SStefano Zampini if (compute_eigs) { 4763854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4764854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 47656e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 47666e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 47676e683305SStefano Zampini lambda_min = eigs_r[0]; 47686e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 47696e683305SStefano Zampini if (lambda_max>lambda_min) { 4770cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4771cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4772cbcc2c2aSStefano Zampini } 4773c8587f34SStefano Zampini } 4774c8587f34SStefano Zampini } 4775cbcc2c2aSStefano Zampini 4776c8587f34SStefano Zampini /* check coarse problem residual error */ 47776e683305SStefano Zampini if (pcbddc->dbg_flag) { 47786e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 47796e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 47806e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4781c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 47826e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 47836e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4784c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4785779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 47866e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 47876e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 47886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 47896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 47906e683305SStefano Zampini if (compute_eigs) { 47916e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4792deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4793c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 47946e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 47956e683305SStefano 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); 47966e683305SStefano Zampini for (i=0;i<neigs;i++) { 47976e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4798c8587f34SStefano Zampini } 47996e683305SStefano Zampini } 48006e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 48016e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48026e683305SStefano Zampini } 4803c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 48046e683305SStefano Zampini if (compute_eigs) { 48056e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 48066e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4807c8587f34SStefano Zampini } 48086e683305SStefano Zampini } 48096e683305SStefano Zampini } 4810cbcc2c2aSStefano Zampini /* print additional info */ 4811cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 48126e683305SStefano Zampini /* waits until all processes reaches this point */ 48136e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4814cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4815cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4816cbcc2c2aSStefano Zampini } 4817cbcc2c2aSStefano Zampini 48182b510759SStefano Zampini /* free memory */ 4819c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4820fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4821c8587f34SStefano Zampini PetscFunctionReturn(0); 4822c8587f34SStefano Zampini } 4823674ae819SStefano Zampini 4824f34684f1SStefano Zampini #undef __FUNCT__ 4825f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4826f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4827f34684f1SStefano Zampini { 4828f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4829f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4830f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4831dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4832dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 483373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4834dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4835f34684f1SStefano Zampini PetscErrorCode ierr; 4836f34684f1SStefano Zampini 4837f34684f1SStefano Zampini PetscFunctionBegin; 4838f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 48390e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 48400e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4841727cdba6SStefano Zampini } 4842dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 48433bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4844dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4845dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4846dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4847dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4848dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4849dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 48500e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 48510e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 48520e6343abSStefano Zampini } 4853dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4854dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4855dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4856dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4857dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4858f34684f1SStefano Zampini 4859f34684f1SStefano Zampini /* check numbering */ 4860f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4861019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 4862dc456d91SStefano Zampini PetscInt i; 4863b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4864f34684f1SStefano Zampini 4865f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4866f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4867f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 48680fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4869019a44ceSStefano Zampini /* counter */ 4870019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4871019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4872019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4873019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4874019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4875019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4876019a44ceSStefano Zampini 4877f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4878f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4879727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4880f34684f1SStefano Zampini } 4881f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4882f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4883f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4884e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4885e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4886e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4887e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4888f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4889019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4890f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4891019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 4892019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 4893019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 4894b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4895019a44ceSStefano 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); 4896f34684f1SStefano Zampini } 4897f34684f1SStefano Zampini } 4898019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4899b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4900f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4901f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4902f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4903f34684f1SStefano Zampini } 4904f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4905f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4906e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4907e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4908f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4909f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4910b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4911ca8b9ea9SStefano Zampini PetscInt *gidxs; 4912ca8b9ea9SStefano Zampini 4913ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 49143bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4915f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4916f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4917f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4918f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 49194bc2dc4bSStefano 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); 4920f34684f1SStefano Zampini } 4921f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4922ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4923f34684f1SStefano Zampini } 4924f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4925302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4926f34684f1SStefano Zampini } 49278bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4928f34684f1SStefano Zampini /* get back data */ 4929f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4930f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4931674ae819SStefano Zampini PetscFunctionReturn(0); 4932674ae819SStefano Zampini } 4933674ae819SStefano Zampini 4934e456f2a8SStefano Zampini #undef __FUNCT__ 4935e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4936a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4937e456f2a8SStefano Zampini { 4938e456f2a8SStefano Zampini IS localis_t; 4939a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4940e456f2a8SStefano Zampini PetscScalar *vals; 4941e456f2a8SStefano Zampini PetscErrorCode ierr; 4942e456f2a8SStefano Zampini 4943e456f2a8SStefano Zampini PetscFunctionBegin; 4944a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4945e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4946854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4947e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4948e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4949a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4950a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 49511035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4952a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 49531035eff8SStefano Zampini } 4954a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4955e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4956e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4957a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4958a7dc3881SStefano Zampini /* now compute set in local ordering */ 4959a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4960a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4961a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4962a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4963a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4964ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4965e456f2a8SStefano Zampini lsize++; 4966e456f2a8SStefano Zampini } 4967e456f2a8SStefano Zampini } 4968854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4969a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4970ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4971e456f2a8SStefano Zampini idxs[lsize++] = i; 4972e456f2a8SStefano Zampini } 4973e456f2a8SStefano Zampini } 4974a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4975a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4976e456f2a8SStefano Zampini *localis = localis_t; 4977e456f2a8SStefano Zampini PetscFunctionReturn(0); 4978e456f2a8SStefano Zampini } 4979906d46d4SStefano Zampini 4980906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4981906d46d4SStefano Zampini #undef __FUNCT__ 4982906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4983906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4984906d46d4SStefano Zampini { 4985906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4986906d46d4SStefano Zampini PetscErrorCode ierr; 4987906d46d4SStefano Zampini 4988906d46d4SStefano Zampini PetscFunctionBegin; 4989906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4990906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4991906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4992906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4993906d46d4SStefano Zampini PetscFunctionReturn(0); 4994906d46d4SStefano Zampini } 4995906d46d4SStefano Zampini 4996906d46d4SStefano Zampini #undef __FUNCT__ 4997906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4998906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4999906d46d4SStefano Zampini { 5000906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5001906d46d4SStefano Zampini PetscErrorCode ierr; 5002906d46d4SStefano Zampini 5003906d46d4SStefano Zampini PetscFunctionBegin; 5004906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5005906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5006906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5007906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5008906d46d4SStefano Zampini PetscFunctionReturn(0); 5009906d46d4SStefano Zampini } 5010b96c3477SStefano Zampini 5011b96c3477SStefano Zampini #undef __FUNCT__ 5012b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 501308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5014b96c3477SStefano Zampini { 5015a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5016b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5017b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5018a64f4aa4SStefano Zampini Mat S_j; 5019b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5020b96c3477SStefano Zampini PetscBool free_used_adj; 5021b96c3477SStefano Zampini PetscErrorCode ierr; 5022b96c3477SStefano Zampini 5023b96c3477SStefano Zampini PetscFunctionBegin; 5024b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5025b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 502608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5027b96c3477SStefano Zampini used_xadj = NULL; 5028b96c3477SStefano Zampini used_adjncy = NULL; 5029b96c3477SStefano Zampini } else { 503008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 503108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 503208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 503308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5034b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5035b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5036b96c3477SStefano Zampini } else { 50372fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5038b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5039b96c3477SStefano Zampini PetscInt nvtxs; 5040b96c3477SStefano Zampini 50412fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 50422fffb893SStefano Zampini if (flg_row) { 5043b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5044b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5045b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5046b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 50472fffb893SStefano Zampini } else { 50482fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 50492fffb893SStefano Zampini used_xadj = NULL; 50502fffb893SStefano Zampini used_adjncy = NULL; 50512fffb893SStefano Zampini } 50522fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5053b96c3477SStefano Zampini } 5054b96c3477SStefano Zampini } 5055d5574798SStefano Zampini 5056d5574798SStefano Zampini /* setup sub_schurs data */ 5057a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5058a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5059a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5060a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 506106a4e24aSStefano 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); 5062a64f4aa4SStefano Zampini } else { 50636816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 506404708bb6SStefano Zampini PetscBool isseqaij; 50655feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 50665feab87aSStefano Zampini PetscInt n_vertices; 50675feab87aSStefano Zampini 50685feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 50692034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 50705feab87aSStefano Zampini } 507104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 507204708bb6SStefano Zampini if (!isseqaij) { 507304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 507404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 507504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 507604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 507704708bb6SStefano Zampini } else { 507804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 507904708bb6SStefano Zampini } 508004708bb6SStefano Zampini } 508106a4e24aSStefano 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); 5082a64f4aa4SStefano Zampini } 5083a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5084b96c3477SStefano Zampini 5085b96c3477SStefano Zampini /* free adjacency */ 5086b96c3477SStefano Zampini if (free_used_adj) { 5087b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5088b96c3477SStefano Zampini } 5089b96c3477SStefano Zampini PetscFunctionReturn(0); 5090b96c3477SStefano Zampini } 5091b96c3477SStefano Zampini 5092b96c3477SStefano Zampini #undef __FUNCT__ 5093b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 509408122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5095b96c3477SStefano Zampini { 5096b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5097b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5098b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5099b96c3477SStefano Zampini PCBDDCGraph graph; 5100b96c3477SStefano Zampini PetscErrorCode ierr; 5101b96c3477SStefano Zampini 5102b96c3477SStefano Zampini PetscFunctionBegin; 5103b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 510408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 51053301b35fSStefano Zampini IS verticesIS,verticescomm; 51063301b35fSStefano Zampini PetscInt vsize,*idxs; 5107b96c3477SStefano Zampini 5108b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 51093301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 51103301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51113301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 51123301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51133301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5114b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 51157fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 51163301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 51173301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5118b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5119b96c3477SStefano Zampini /* 5120b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5121b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5122b96c3477SStefano Zampini } 5123b96c3477SStefano Zampini */ 5124b96c3477SStefano Zampini } else { 5125b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5126b96c3477SStefano Zampini } 5127b96c3477SStefano Zampini 5128b96c3477SStefano Zampini /* sub_schurs init */ 5129a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5130a64f4aa4SStefano Zampini 5131b96c3477SStefano Zampini /* free graph struct */ 513208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5133b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5134b96c3477SStefano Zampini } 5135b96c3477SStefano Zampini PetscFunctionReturn(0); 5136b96c3477SStefano Zampini } 5137