1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini 8674ae819SStefano Zampini #undef __FUNCT__ 9efc2fbd9SStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushP0" 10efc2fbd9SStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushP0(PC pc, Vec v, PetscBool pop) 11efc2fbd9SStefano Zampini { 12efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13efc2fbd9SStefano Zampini PetscErrorCode ierr; 14efc2fbd9SStefano Zampini 15efc2fbd9SStefano Zampini PetscFunctionBegin; 16efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 17efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 18c9ed8603SStefano Zampini if (pcbddc->benign_p0_gidx >= 0) { 19efc2fbd9SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,1,NULL,PETSC_OWN_POINTER,&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 20c9ed8603SStefano Zampini } else { 21c9ed8603SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,0,NULL,PETSC_OWN_POINTER,&pcbddc->benign_p0_gidx);CHKERRQ(ierr); 22c9ed8603SStefano Zampini } 23efc2fbd9SStefano Zampini } 24efc2fbd9SStefano Zampini if (pop) { /* use SF to get values */ 25efc2fbd9SStefano Zampini PetscScalar *array; 26efc2fbd9SStefano Zampini 27efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 28efc2fbd9SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 29efc2fbd9SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,&pcbddc->benign_p0);CHKERRQ(ierr); 30efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 31efc2fbd9SStefano Zampini } else { /* use VecSetValue */ 32c9ed8603SStefano Zampini if (pcbddc->benign_p0_gidx >= 0) { 33efc2fbd9SStefano Zampini ierr = VecSetValue(v,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 34c9ed8603SStefano Zampini } 35efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 36efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 37efc2fbd9SStefano Zampini } 38efc2fbd9SStefano Zampini PetscFunctionReturn(0); 39efc2fbd9SStefano Zampini } 40efc2fbd9SStefano Zampini 41efc2fbd9SStefano Zampini #undef __FUNCT__ 42c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 43c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 44c263805aSStefano Zampini { 45c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 46c263805aSStefano Zampini PetscErrorCode ierr; 47c263805aSStefano Zampini 48c263805aSStefano Zampini PetscFunctionBegin; 49c263805aSStefano Zampini /* TODO: add error checking 50c263805aSStefano Zampini - avoid nested pop (or push) calls. 51c263805aSStefano Zampini - cannot push before pop. 52c263805aSStefano Zampini */ 53c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx < 0) { 54efc2fbd9SStefano Zampini PetscFunctionReturn(0); 55efc2fbd9SStefano Zampini } 56c263805aSStefano Zampini if (pop) { 57c263805aSStefano Zampini const PetscInt *cB0_cols; 58c263805aSStefano Zampini PetscInt cB0_ncol; 59c263805aSStefano Zampini const PetscScalar *cB0_vals; 60c263805aSStefano Zampini 61c263805aSStefano Zampini /* extract B_0 */ 62efc2fbd9SStefano Zampini ierr = MatGetRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 63c263805aSStefano Zampini pcbddc->B0_ncol = cB0_ncol; 64c263805aSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 65c263805aSStefano Zampini ierr = PetscMalloc2(cB0_ncol,&pcbddc->B0_cols,cB0_ncol,&pcbddc->B0_vals);CHKERRQ(ierr); 66c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_cols,cB0_cols,cB0_ncol*sizeof(PetscInt));CHKERRQ(ierr); 67c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_vals,cB0_vals,cB0_ncol*sizeof(PetscScalar));CHKERRQ(ierr); 68efc2fbd9SStefano Zampini ierr = MatRestoreRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 69c263805aSStefano Zampini /* remove rows and cols from local problem */ 70c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 71efc2fbd9SStefano Zampini ierr = MatZeroRowsColumns(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,1.,NULL,NULL);CHKERRQ(ierr); 72c263805aSStefano Zampini } else { /* push */ 73efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,pcbddc->B0_ncol,pcbddc->B0_cols,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 74efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,pcbddc->B0_ncol,pcbddc->B0_cols,1,&pcbddc->benign_p0_lidx,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 75efc2fbd9SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx,pcbddc->benign_p0_lidx,0.0,INSERT_VALUES);CHKERRQ(ierr); 76c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 77c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 78c263805aSStefano Zampini } 79c263805aSStefano Zampini PetscFunctionReturn(0); 80c263805aSStefano Zampini } 81c263805aSStefano Zampini 82c263805aSStefano Zampini #undef __FUNCT__ 83b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 8408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 85b1b3d7a2SStefano Zampini { 86b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 8808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 8908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 9008122e43SStefano Zampini PetscScalar *work,lwork; 9108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 9208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 9308122e43SStefano Zampini PetscReal *eigs,thresh; 941b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 95f6f667cfSStefano Zampini PetscBool allocated_S_St; 9608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9708122e43SStefano Zampini PetscReal *rwork; 9808122e43SStefano Zampini #endif 99b1b3d7a2SStefano Zampini PetscErrorCode ierr; 100b1b3d7a2SStefano Zampini 101b1b3d7a2SStefano Zampini PetscFunctionBegin; 10208122e43SStefano Zampini if (!sub_schurs->use_mumps) { 10308122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 10408122e43SStefano Zampini } 10508122e43SStefano Zampini 10606a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 10706a4e24aSStefano 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); 10806a4e24aSStefano Zampini } 10906a4e24aSStefano Zampini 110fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 111fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 112fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 113fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 114fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 115fd14bc51SStefano Zampini } 116fd14bc51SStefano Zampini 117e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 118e496cd5dSStefano 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); 119e496cd5dSStefano Zampini } 120e496cd5dSStefano Zampini 12108122e43SStefano Zampini /* max size of subsets */ 12208122e43SStefano Zampini mss = 0; 12308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 12408122e43SStefano Zampini PetscInt subset_size; 125862806e4SStefano Zampini 12608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 12708122e43SStefano Zampini mss = PetscMax(mss,subset_size); 12808122e43SStefano Zampini } 12908122e43SStefano Zampini 13008122e43SStefano Zampini /* min/max and threshold */ 13108122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 132f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 13308122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 134f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 135f6f667cfSStefano Zampini if (nmin) { 136f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 137f6f667cfSStefano Zampini } 13808122e43SStefano Zampini 13908122e43SStefano Zampini /* allocate lapack workspace */ 14008122e43SStefano Zampini cum = cum2 = 0; 14108122e43SStefano Zampini maxneigs = 0; 14208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 14308122e43SStefano Zampini PetscInt n,subset_size; 144f6f667cfSStefano Zampini 14508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 14608122e43SStefano Zampini n = PetscMin(subset_size,nmax); 1479162d606SStefano Zampini cum += subset_size; 1489162d606SStefano Zampini cum2 += subset_size*n; 14908122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 15008122e43SStefano Zampini } 15108122e43SStefano Zampini if (mss) { 1529ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 15308122e43SStefano Zampini PetscBLASInt B_itype = 1; 15408122e43SStefano Zampini PetscBLASInt B_N = mss; 1554c6709b3SStefano Zampini PetscReal zero = 0.0; 1564c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 15708122e43SStefano Zampini 15808122e43SStefano Zampini B_lwork = -1; 15908122e43SStefano Zampini S = NULL; 16008122e43SStefano Zampini St = NULL; 161a58a30b4SStefano Zampini eigs = NULL; 162a58a30b4SStefano Zampini eigv = NULL; 163a58a30b4SStefano Zampini B_iwork = NULL; 164a58a30b4SStefano Zampini B_ifail = NULL; 165d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 166d1710679SStefano Zampini rwork = NULL; 167d1710679SStefano Zampini #endif 1688bec7fa6SStefano Zampini thresh = 1.0; 16908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 17008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 17108122e43SStefano 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)); 17208122e43SStefano Zampini #else 17308122e43SStefano 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)); 17408122e43SStefano Zampini #endif 17508122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 17608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 17708122e43SStefano Zampini } else { 17808122e43SStefano Zampini /* TODO */ 17908122e43SStefano Zampini } 18008122e43SStefano Zampini } else { 18108122e43SStefano Zampini lwork = 0; 18208122e43SStefano Zampini } 18308122e43SStefano Zampini 18408122e43SStefano Zampini nv = 0; 185d62866d3SStefano 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) */ 186d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 18708122e43SStefano Zampini } 1884c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 189f6f667cfSStefano Zampini if (allocated_S_St) { 190f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 191f6f667cfSStefano Zampini } 192f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 19308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 19408122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 19508122e43SStefano Zampini #endif 1969162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1979162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1989162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 19908122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 2009162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 20108122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 20208122e43SStefano Zampini 20308122e43SStefano Zampini maxneigs = 0; 20408122e43SStefano Zampini cum = cum2 = cumarray = 0; 2059162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 2069162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 207d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 20808122e43SStefano Zampini const PetscInt *idxs; 20908122e43SStefano Zampini 210d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 21108122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 21208122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 21308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 21408122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 2159162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 2169162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 21708122e43SStefano Zampini } 21808122e43SStefano Zampini cum2 = cum; 219d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 22008122e43SStefano Zampini } 22108122e43SStefano Zampini 22208122e43SStefano Zampini if (mss) { /* multilevel */ 22308122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 22408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 22508122e43SStefano Zampini } 22608122e43SStefano Zampini 22708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 22808122e43SStefano Zampini 22908122e43SStefano Zampini const PetscInt *idxs; 230f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 231862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 23208122e43SStefano Zampini PetscBLASInt B_N; 233aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 23408122e43SStefano Zampini 235862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 236f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 237f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 2389ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 239aff50787SStefano Zampini PetscInt j,k; 240aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 241aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 242aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24308122e43SStefano Zampini } 24408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 245aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 246aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 247aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 248aff50787SStefano Zampini } 24908122e43SStefano Zampini } 25008122e43SStefano Zampini } else { 25108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25308122e43SStefano Zampini } 2548bec7fa6SStefano Zampini } else { 255f6f667cfSStefano Zampini S = Sarray + cumarray; 256f6f667cfSStefano Zampini St = Starray + cumarray; 2578bec7fa6SStefano Zampini } 25808122e43SStefano Zampini 259f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 260aff50787SStefano Zampini /* see if we can save some work */ 261aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 262aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 263aff50787SStefano Zampini } 264aff50787SStefano Zampini 265aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 266aff50787SStefano Zampini B_neigs = 0; 267aff50787SStefano Zampini } else { 268aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 269f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 270f6f667cfSStefano Zampini 2719ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 27208122e43SStefano Zampini PetscBLASInt B_itype = 1; 273f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 2744c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 2759552c7c7SStefano Zampini PetscInt nmin_s; 27608122e43SStefano Zampini 277f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 278fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2798bec7fa6SStefano 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]]); 280fd14bc51SStefano Zampini } 28108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 28208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 283f6f667cfSStefano 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)); 28408122e43SStefano Zampini #else 285f6f667cfSStefano 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)); 28608122e43SStefano Zampini #endif 28708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 28808122e43SStefano Zampini if (B_ierr) { 28908122e43SStefano Zampini if (B_ierr < 0 ) { 29008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 29108122e43SStefano Zampini } else if (B_ierr <= B_N) { 29208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 29308122e43SStefano Zampini } else { 2949552c7c7SStefano 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); 29508122e43SStefano Zampini } 29608122e43SStefano Zampini } 29708122e43SStefano Zampini 29808122e43SStefano Zampini if (B_neigs > nmax) { 299fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 300fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 301fd14bc51SStefano Zampini } 302f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 30308122e43SStefano Zampini B_neigs = nmax; 30408122e43SStefano Zampini } 30508122e43SStefano Zampini 3069552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 3079552c7c7SStefano Zampini if (B_neigs < nmin_s) { 30808122e43SStefano Zampini PetscBLASInt B_neigs2; 30908122e43SStefano Zampini 310f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 311f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 312fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 313fd14bc51SStefano 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); 314fd14bc51SStefano Zampini } 3159ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 31608122e43SStefano Zampini PetscInt j; 31708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 31808122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 31908122e43SStefano Zampini } 32008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 32108122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 32208122e43SStefano Zampini } 32308122e43SStefano Zampini } else { 32408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 32508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 32608122e43SStefano Zampini } 32708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 32808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 329f6f667cfSStefano 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)); 33008122e43SStefano Zampini #else 331f6f667cfSStefano 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)); 33208122e43SStefano Zampini #endif 33308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 33408122e43SStefano Zampini B_neigs += B_neigs2; 33508122e43SStefano Zampini } 33608122e43SStefano Zampini if (B_ierr) { 33708122e43SStefano Zampini if (B_ierr < 0 ) { 33808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 33908122e43SStefano Zampini } else if (B_ierr <= B_N) { 34008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 34108122e43SStefano Zampini } else { 3429552c7c7SStefano 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); 34308122e43SStefano Zampini } 34408122e43SStefano Zampini } 345fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 346ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 34708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 34808122e43SStefano Zampini if (eigs[j] == 0.0) { 349ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 35008122e43SStefano Zampini } else { 351ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 352fd14bc51SStefano Zampini } 35308122e43SStefano Zampini } 35408122e43SStefano Zampini } 35508122e43SStefano Zampini } else { 35608122e43SStefano Zampini /* TODO */ 35708122e43SStefano Zampini } 358aff50787SStefano Zampini } 3598bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 3608bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 3619162d606SStefano Zampini if (B_neigs) { 3629162d606SStefano 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); 363fd14bc51SStefano Zampini 364fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 3659552c7c7SStefano Zampini PetscInt ii; 3669552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 367ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 3689552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 369ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 370ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 371ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 372ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 373ac47001eSStefano Zampini #else 374ac47001eSStefano 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); 375ac47001eSStefano Zampini #endif 3769552c7c7SStefano Zampini } 3779552c7c7SStefano Zampini } 378fd14bc51SStefano Zampini } 37908122e43SStefano Zampini #if 0 3809162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 38108122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 38208122e43SStefano Zampini PetscScalar norm; 38308122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3849162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3859162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 38608122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 38708122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 38808122e43SStefano Zampini } else { 38908122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 39008122e43SStefano Zampini } 3919162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 392b1b3d7a2SStefano Zampini } 393b1b3d7a2SStefano Zampini #endif 3949162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3959162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3969162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3979162d606SStefano Zampini cum++; 39808122e43SStefano Zampini } 39908122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 40008122e43SStefano Zampini /* shift for next computation */ 40108122e43SStefano Zampini cumarray += subset_size*subset_size; 40208122e43SStefano Zampini } 403fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 404fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 405fd14bc51SStefano Zampini } 40608122e43SStefano Zampini 40708122e43SStefano Zampini if (mss) { 40808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 40908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 410f6f667cfSStefano Zampini /* destroy matrices (junk) */ 411f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 412f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 41308122e43SStefano Zampini } 414f6f667cfSStefano Zampini if (allocated_S_St) { 415f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 416f6f667cfSStefano Zampini } 417f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 41808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 41908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 42008122e43SStefano Zampini #endif 42108122e43SStefano Zampini if (pcbddc->dbg_flag) { 4221b968477SStefano Zampini PetscInt maxneigs_r; 42308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4249b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 42508122e43SStefano Zampini } 42608122e43SStefano Zampini PetscFunctionReturn(0); 42708122e43SStefano Zampini } 428b1b3d7a2SStefano Zampini 429674ae819SStefano Zampini #undef __FUNCT__ 430c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 431c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 432c8587f34SStefano Zampini { 433c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4348629588bSStefano Zampini PetscScalar *coarse_submat_vals; 435c8587f34SStefano Zampini PetscErrorCode ierr; 436c8587f34SStefano Zampini 437c8587f34SStefano Zampini PetscFunctionBegin; 438f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 4395e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 440c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 441c8587f34SStefano Zampini 442684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 4430fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 444684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 445c8587f34SStefano Zampini 446c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 447b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 448c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 449c8587f34SStefano Zampini } 450c8587f34SStefano Zampini 4518629588bSStefano Zampini /* 4528629588bSStefano Zampini Setup local correction and local part of coarse basis. 4538629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 4548629588bSStefano Zampini */ 45547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 4568629588bSStefano Zampini 4578629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 4588629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 4598629588bSStefano Zampini 4608629588bSStefano Zampini /* free */ 4618629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 462c8587f34SStefano Zampini PetscFunctionReturn(0); 463c8587f34SStefano Zampini } 464c8587f34SStefano Zampini 465c8587f34SStefano Zampini #undef __FUNCT__ 466674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 467674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 468674ae819SStefano Zampini { 469674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 470674ae819SStefano Zampini PetscErrorCode ierr; 471674ae819SStefano Zampini 472674ae819SStefano Zampini PetscFunctionBegin; 473674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 474674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 475674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 476674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 477785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 478674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 479f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 480f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 481785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 48263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 48363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 484674ae819SStefano Zampini PetscFunctionReturn(0); 485674ae819SStefano Zampini } 486674ae819SStefano Zampini 487674ae819SStefano Zampini #undef __FUNCT__ 488674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 489674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 490674ae819SStefano Zampini { 491674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 492674ae819SStefano Zampini PetscErrorCode ierr; 493674ae819SStefano Zampini 494674ae819SStefano Zampini PetscFunctionBegin; 495b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 496674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 497674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 498674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 499b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 500674ae819SStefano Zampini PetscFunctionReturn(0); 501674ae819SStefano Zampini } 502674ae819SStefano Zampini 503674ae819SStefano Zampini #undef __FUNCT__ 504674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 505674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 506674ae819SStefano Zampini { 507674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 50806656605SStefano Zampini PetscScalar *array; 509674ae819SStefano Zampini PetscErrorCode ierr; 510674ae819SStefano Zampini 511674ae819SStefano Zampini PetscFunctionBegin; 512674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 51358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 51406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 51506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 51658da7f69SStefano Zampini } 517674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 518674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 51915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 52015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 521674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 522674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 523674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 52406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 525674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 526674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5278ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 528674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 529674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 530674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 531f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 532f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 533f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 534f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 535727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5360e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 537f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 53870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5396e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54081d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 5410369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 5428b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 5438b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 54481d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 545674ae819SStefano Zampini PetscFunctionReturn(0); 546674ae819SStefano Zampini } 547674ae819SStefano Zampini 548674ae819SStefano Zampini #undef __FUNCT__ 549f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 550f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 5516bfb1811SStefano Zampini { 5526bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5536bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 5546bfb1811SStefano Zampini VecType impVecType; 555019a44ceSStefano Zampini PetscInt n_constraints,n_R,old_size,n_benign; 5566bfb1811SStefano Zampini PetscErrorCode ierr; 5576bfb1811SStefano Zampini 5586bfb1811SStefano Zampini PetscFunctionBegin; 559f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 560019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 561f4ddd8eeSStefano Zampini } 562e7b262bdSStefano Zampini /* get sizes */ 563c9ed8603SStefano Zampini n_benign = 0; 564c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) n_benign = 1; 565019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - pcbddc->n_vertices; 566019a44ceSStefano Zampini if (n_benign && n_constraints) { 567019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot handle benign susbspace trick without change of basis"); 568019a44ceSStefano Zampini } 569b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 5706bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 571e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 572e7b262bdSStefano Zampini /* R nodes */ 573e7b262bdSStefano Zampini old_size = -1; 574e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 575e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 576e7b262bdSStefano Zampini } 577e7b262bdSStefano Zampini if (n_R != old_size) { 578e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 579e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5806bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5816bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5826bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5836bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 584e7b262bdSStefano Zampini } 585e7b262bdSStefano Zampini /* local primal dofs */ 586e7b262bdSStefano Zampini old_size = -1; 587e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 588e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 589e7b262bdSStefano Zampini } 590e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 591e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 59283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 593e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5946bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 595e7b262bdSStefano Zampini } 596e7b262bdSStefano Zampini /* local explicit constraints */ 597e7b262bdSStefano Zampini old_size = -1; 598e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 599e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 600e7b262bdSStefano Zampini } 601e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 602e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 60383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 60483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 60583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 60683b7ccabSStefano Zampini } 6076bfb1811SStefano Zampini PetscFunctionReturn(0); 6086bfb1811SStefano Zampini } 6096bfb1811SStefano Zampini 6106bfb1811SStefano Zampini #undef __FUNCT__ 61147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 61247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 61388ebb749SStefano Zampini { 61425084f0cSStefano Zampini PetscErrorCode ierr; 61525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 61688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 61788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 618d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 61925084f0cSStefano Zampini /* submatrices of local problem */ 62080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 621019a44ceSStefano Zampini /* submatrices of benign trick */ 622019a44ceSStefano Zampini Mat B0_R,B0_V; 62306656605SStefano Zampini /* submatrices of local coarse problem */ 62406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 62525084f0cSStefano Zampini /* working matrices */ 62606656605SStefano Zampini Mat C_CR; 62725084f0cSStefano Zampini /* additional working stuff */ 62806656605SStefano Zampini PC pc_R; 629d12edf2fSStefano Zampini Mat F,B0 = NULL; 630d12edf2fSStefano Zampini Vec vec_R_benign,vec_R_benign_2; 63106656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 63206656605SStefano Zampini 63325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 63406656605SStefano Zampini PetscScalar *work; 63506656605SStefano Zampini PetscInt *idx_V_B; 636d12edf2fSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign,p0_lidx_I = 0; 63706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 638b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 63945a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 64088ebb749SStefano Zampini MatType impMatType; 64125084f0cSStefano Zampini /* some shortcuts to scalars */ 64206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 64388ebb749SStefano Zampini 64488ebb749SStefano Zampini PetscFunctionBegin; 645c9ed8603SStefano Zampini n_benign = 0; 646c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) n_benign = 1; 647b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 648019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - n_vertices; 64988ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 650b371cd4fSStefano Zampini n_B = pcis->n_B; 651b371cd4fSStefano Zampini n_D = pcis->n - n_B; 65288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 65388ebb749SStefano Zampini 65488ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 65588ebb749SStefano Zampini impMatType = MATSEQDENSE; 65688ebb749SStefano Zampini 65788ebb749SStefano Zampini /* vertices in boundary numbering */ 658785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 6590e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 66088ebb749SStefano Zampini if (i != n_vertices) { 66122d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 66288ebb749SStefano Zampini } 66388ebb749SStefano Zampini 66406656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 665019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 666019a44ceSStefano Zampini if (!n_benign) { 66706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 66806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 66906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 67006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 67106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 67206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 67306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 67406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 675019a44ceSStefano Zampini } else { /* these wrappers are not needed with the benign trick */ 676019a44ceSStefano Zampini S_VV = S_CV = S_VC = S_CC = NULL; 677019a44ceSStefano Zampini } 67806656605SStefano Zampini 67906656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 68006656605SStefano Zampini /* allocate workspace */ 68106656605SStefano Zampini n = 0; 68206656605SStefano Zampini if (n_constraints) { 68306656605SStefano Zampini n += n_R*n_constraints; 68406656605SStefano Zampini } 68506656605SStefano Zampini if (n_vertices) { 68606656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 68780677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 68806656605SStefano Zampini } 6893301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 69006656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 69106656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 69206656605SStefano Zampini } 69306656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 69406656605SStefano Zampini 69506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 69606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 69706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 69806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 69906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 70006656605SStefano Zampini if (isLU || isILU || isCHOL) { 70106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 702d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 703d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 704d62866d3SStefano Zampini MatFactorType type; 705d62866d3SStefano Zampini 7066816873aSStefano Zampini F = reuse_mumps->F; 7076816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 708d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 70906656605SStefano Zampini } else { 71006656605SStefano Zampini F = NULL; 71106656605SStefano Zampini } 71206656605SStefano Zampini 71388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 71488ebb749SStefano Zampini if (n_constraints) { 71506656605SStefano Zampini Mat M1,M2,M3; 71680677318SStefano Zampini Mat auxmat; 71706656605SStefano Zampini IS is_aux; 71880677318SStefano Zampini PetscScalar *array,*array2; 71906656605SStefano Zampini 720f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 72180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 72288ebb749SStefano Zampini 72325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 72425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 7258ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 72680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 72788ebb749SStefano Zampini 72880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 72980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 73006656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 73188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 73206656605SStefano Zampini const PetscScalar *row_cmat_values; 73306656605SStefano Zampini const PetscInt *row_cmat_indices; 73406656605SStefano Zampini PetscInt size_of_constraint,j; 73588ebb749SStefano Zampini 73606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 73706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 73806656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 73906656605SStefano Zampini } 74006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 74106656605SStefano Zampini } 74280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 74306656605SStefano Zampini if (F) { 74406656605SStefano Zampini Mat B; 74506656605SStefano Zampini 74606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 74780677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 74806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 74906656605SStefano Zampini } else { 75080677318SStefano Zampini PetscScalar *marr; 75180677318SStefano Zampini 75280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 75306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 75406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 75580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 75606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 75706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 75806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 75906656605SStefano Zampini } 76080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 76106656605SStefano Zampini } 76280677318SStefano Zampini if (!pcbddc->switch_static) { 76380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 76480677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 76580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 76680677318SStefano Zampini for (i=0;i<n_constraints;i++) { 76780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 76880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 76980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 77080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 77180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 77280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 77380677318SStefano Zampini } 77480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 77580677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 77680677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 77780677318SStefano Zampini } else { 77880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 77980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 78025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 78180677318SStefano Zampini } 78280677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 78380677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 78480677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 78506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 78606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 78780677318SStefano Zampini if (isCHOL) { 78880677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 78980677318SStefano Zampini } else { 79025084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 79180677318SStefano Zampini } 79280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 79306656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 79425084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 79525084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 79625084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 79780677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 79880677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 79980677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 80006656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 80106656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 802f4ddd8eeSStefano Zampini } 80388ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 80488ebb749SStefano Zampini if (n_vertices) { 80506656605SStefano Zampini IS is_aux; 8063a50541eSStefano Zampini 8076816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 8086816873aSStefano Zampini IS tis; 8096816873aSStefano Zampini 8106816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 8116816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 8126816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 8136816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 8146816873aSStefano Zampini } else { 8153a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 8166816873aSStefano Zampini } 8179577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 8189577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 81904708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 820019a44ceSStefano Zampini if (n_benign) { 821019a44ceSStefano Zampini IS dummy; 822019a44ceSStefano Zampini PetscInt ii[2]; 823019a44ceSStefano Zampini 824019a44ceSStefano Zampini ii[0] = 0; 825019a44ceSStefano Zampini ii[1] = pcbddc->B0_ncol; 826019a44ceSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 827019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 828019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 829019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 830019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 831019a44ceSStefano Zampini } 83225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 83388ebb749SStefano Zampini } 83488ebb749SStefano Zampini 83588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 836f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 83706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 83806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 83906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 84006656605SStefano Zampini } 841f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 84206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 84306656605SStefano Zampini PetscScalar *marray; 84406656605SStefano Zampini 84506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 847f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 848f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 849f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 850f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 851f4ddd8eeSStefano Zampini } 852f4ddd8eeSStefano Zampini } 85306656605SStefano Zampini 854f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 85506656605SStefano Zampini PetscScalar *marray; 85688ebb749SStefano Zampini 85706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8588eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 85906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 86088ebb749SStefano Zampini } 8613301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 86206656605SStefano Zampini n *= 2; 86388ebb749SStefano Zampini } 86406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 86606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8678eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 86906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 87088ebb749SStefano Zampini } 8713301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 87206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 8738eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 87406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 87506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 87688ebb749SStefano Zampini } 87788ebb749SStefano Zampini } else { 878c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 879c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8801b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 881c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 882c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 883c0553b1fSStefano Zampini } 88488ebb749SStefano Zampini } 88506656605SStefano Zampini } 886019a44ceSStefano Zampini 88706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 888d12edf2fSStefano Zampini if (n_benign && (pcbddc->switch_static || pcbddc->dbg_flag)) { 889d12edf2fSStefano Zampini const PetscInt *idxs; 890d12edf2fSStefano Zampini 891d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 892d12edf2fSStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx,pcis->n-pcis->n_B,idxs,&p0_lidx_I);CHKERRQ(ierr); 893d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 894d12edf2fSStefano Zampini } 89506656605SStefano Zampini /* vertices */ 89606656605SStefano Zampini if (n_vertices) { 89716f15bc4SStefano Zampini 89804708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 89904708bb6SStefano Zampini 90016f15bc4SStefano Zampini if (n_R) { 90106656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 902019a44ceSStefano Zampini Mat S_V_benign; 90306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 90416f15bc4SStefano Zampini PetscScalar *x,*y; 90504708bb6SStefano Zampini PetscBool isseqaij; 90606656605SStefano Zampini 90721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 90904708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 9106816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 91106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 91206656605SStefano Zampini } else { 91306656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 91406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 91506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 91806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 92006656605SStefano Zampini } 92106656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 92206656605SStefano Zampini } 923019a44ceSStefano Zampini if (n_benign) { 924019a44ceSStefano Zampini const PetscScalar *vals; 925019a44ceSStefano Zampini const PetscInt *idxs; 926019a44ceSStefano Zampini PetscInt n; 927019a44ceSStefano Zampini 928019a44ceSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign);CHKERRQ(ierr); 929d12edf2fSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&vec_R_benign_2);CHKERRQ(ierr); 930d12edf2fSStefano Zampini ierr = VecSet(vec_R_benign_2,0.);CHKERRQ(ierr); 931019a44ceSStefano Zampini ierr = MatGetRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 932d12edf2fSStefano Zampini ierr = VecSetValues(vec_R_benign_2,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 933d12edf2fSStefano Zampini ierr = VecAssemblyBegin(vec_R_benign_2);CHKERRQ(ierr); 934d12edf2fSStefano Zampini ierr = VecAssemblyEnd(vec_R_benign_2);CHKERRQ(ierr); 935019a44ceSStefano Zampini ierr = MatRestoreRow(B0_R,0,&n,&idxs,&vals);CHKERRQ(ierr); 936d12edf2fSStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,vec_R_benign_2,vec_R_benign);CHKERRQ(ierr); 937019a44ceSStefano Zampini ierr = VecScale(vec_R_benign,m_one);CHKERRQ(ierr); 938019a44ceSStefano Zampini } 93980677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 94006656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 94106656605SStefano Zampini if (n_constraints) { 94206656605SStefano Zampini Mat B; 94380677318SStefano Zampini 944b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 94580677318SStefano Zampini for (i=0;i<n_vertices;i++) { 94680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 94780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 94880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 94980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 95180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95280677318SStefano Zampini } 95380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 95480677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 95580677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 95780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 96006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 96106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 96206656605SStefano Zampini } 96304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 96404708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 96504708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 96604708bb6SStefano Zampini } 96706656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 968019a44ceSStefano Zampini if (n_benign) { 969019a44ceSStefano Zampini ierr = MatMatMult(B0_R,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_V_benign);CHKERRQ(ierr); 970019a44ceSStefano Zampini } 97180677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 97206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 97306656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 97406656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 97506656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 97606656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 978019a44ceSStefano Zampini if (!n_benign) { 97906656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 980019a44ceSStefano Zampini } else { 981019a44ceSStefano Zampini const PetscScalar *vals; 982019a44ceSStefano Zampini PetscScalar val; 983019a44ceSStefano Zampini const PetscInt *idxs; 984019a44ceSStefano Zampini PetscInt n,j; 985019a44ceSStefano Zampini 986019a44ceSStefano Zampini /* Use explicit loops to populate the local coarse matrix */ 987019a44ceSStefano Zampini ierr = MatDenseGetArray(S_VVt,&x);CHKERRQ(ierr); 988019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 989019a44ceSStefano Zampini for (i=0;i<n_vertices;i++) { 990019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+i] = x[j*n_vertices+i]; 991019a44ceSStefano Zampini } 992019a44ceSStefano Zampini } 993019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&x);CHKERRQ(ierr); 994019a44ceSStefano Zampini ierr = MatDenseGetArray(S_V_benign,&x);CHKERRQ(ierr); 995019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 996019a44ceSStefano Zampini for (j=0;j<n_vertices;j++) { 997019a44ceSStefano Zampini coarse_submat_vals[n_vertices*pcbddc->local_primal_size+j] = x[j]+vals[j]; 998019a44ceSStefano Zampini coarse_submat_vals[j*pcbddc->local_primal_size+n_vertices] = x[j]+vals[j]; 999019a44ceSStefano Zampini } 1000019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 1001019a44ceSStefano Zampini ierr = MatDenseRestoreArray(S_V_benign,&x);CHKERRQ(ierr); 1002019a44ceSStefano Zampini ierr = MatDestroy(&S_V_benign);CHKERRQ(ierr); 1003d12edf2fSStefano Zampini ierr = VecDot(vec_R_benign_2,vec_R_benign,&val);CHKERRQ(ierr); 1004d12edf2fSStefano Zampini coarse_submat_vals[pcbddc->local_primal_size*pcbddc->local_primal_size-1] = val; 1005019a44ceSStefano Zampini } 100606656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 100716f15bc4SStefano Zampini } else { 1008019a44ceSStefano Zampini /* TODO benign case also here */ 1009019a44ceSStefano Zampini if (n_benign) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not yet implemented"); 101016f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 101116f15bc4SStefano Zampini } 101221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 101306656605SStefano Zampini /* coarse basis functions */ 101406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 101516f15bc4SStefano Zampini PetscScalar *y; 101616f15bc4SStefano Zampini 101706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 101806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 101906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 102006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 102106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 102206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 102306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 102406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 102506656605SStefano Zampini 102606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 102706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 102806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 102906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 103006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 103106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1032a0aff763SStefano Zampini if (n_benign) y[n_D*i+p0_lidx_I] = 0.0; 103306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 103406656605SStefano Zampini } 103506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 103606656605SStefano Zampini } 1037019a44ceSStefano Zampini if (n_benign) { 1038019a44ceSStefano Zampini PetscScalar *y; 1039019a44ceSStefano Zampini 1040019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1041019a44ceSStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*n_vertices);CHKERRQ(ierr); 1042019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1043019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,vec_R_benign,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1044019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 1045019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 1046019a44ceSStefano Zampini 1047019a44ceSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1048019a44ceSStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1049d12edf2fSStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*n_vertices);CHKERRQ(ierr); 1050019a44ceSStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1051019a44ceSStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,vec_R_benign,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1052019a44ceSStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1053d12edf2fSStefano Zampini y[n_D*n_vertices+p0_lidx_I] = 0.0; 1054019a44ceSStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 1055019a44ceSStefano Zampini } 1056019a44ceSStefano Zampini } 105704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 105804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 105906656605SStefano Zampini } 106006656605SStefano Zampini 106106656605SStefano Zampini if (n_constraints) { 106206656605SStefano Zampini Mat B; 106306656605SStefano Zampini 106406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 106506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 106680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 106706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 106806656605SStefano Zampini if (n_vertices) { 106980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 107080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 107180677318SStefano Zampini } else { 107280677318SStefano Zampini Mat S_VCt; 107380677318SStefano Zampini 107480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 107580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 107680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 107780677318SStefano Zampini } 107806656605SStefano Zampini } 107906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 108006656605SStefano Zampini /* coarse basis functions */ 108106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 108206656605SStefano Zampini PetscScalar *y; 108306656605SStefano Zampini 108406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 108506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 108606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 108706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 108906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 109006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 109106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 109206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 109306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 109406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 109506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 109606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 109706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 109806656605SStefano Zampini } 109906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 110006656605SStefano Zampini } 110106656605SStefano Zampini } 110280677318SStefano Zampini if (n_constraints) { 110380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 110480677318SStefano Zampini } 110506656605SStefano Zampini 1106019a44ceSStefano Zampini if (n_benign) { 1107019a44ceSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1108019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1109019a44ceSStefano Zampini ierr = VecDestroy(&vec_R_benign);CHKERRQ(ierr); 1110b1ecc7b1SStefano Zampini ierr = VecDestroy(&vec_R_benign_2);CHKERRQ(ierr); 1111019a44ceSStefano Zampini } 1112019a44ceSStefano Zampini 111306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 11143301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 111506656605SStefano Zampini 111606656605SStefano Zampini if (n_constraints) { 111716f15bc4SStefano Zampini Mat S_CCT,B_C; 111806656605SStefano Zampini 111980677318SStefano Zampini /* this is a lazy thing */ 112080677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 112106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 112206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 112306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 112416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 112506656605SStefano Zampini if (n_vertices) { 112616f15bc4SStefano Zampini Mat B_V,S_VCT; 112706656605SStefano Zampini 112806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 112906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 113006656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 113106656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 113216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 113306656605SStefano Zampini } 113406656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 113580677318SStefano Zampini } else { /* if there are no constraints, reset work */ 113680677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 113706656605SStefano Zampini } 113816f15bc4SStefano Zampini if (n_vertices && n_R) { 113906656605SStefano Zampini Mat A_VRT; 114080677318SStefano Zampini PetscScalar *marray; 114106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 114206656605SStefano Zampini 114380677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 114480677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 114580677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 114780677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 114880677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 114916f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 115006656605SStefano Zampini } 115106656605SStefano Zampini 115206656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 115306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 115406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 115506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 115606656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 115806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 115906656605SStefano Zampini } 116006656605SStefano Zampini } else { 116106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 116206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 116306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 116406656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 116506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 116606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 116706656605SStefano Zampini } 116806656605SStefano Zampini } 116906656605SStefano Zampini /* coarse basis functions */ 117006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 117106656605SStefano Zampini PetscScalar *y; 117206656605SStefano Zampini 117306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 117406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 117506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 117606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 117806656605SStefano Zampini if (i<n_vertices) { 117906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 118006656605SStefano Zampini } 118106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 118206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 118306656605SStefano Zampini 118406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 118506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 118606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 118706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 118806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 118906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 119006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 119106656605SStefano Zampini } 119206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 119306656605SStefano Zampini } 119406656605SStefano Zampini } 1195d62866d3SStefano Zampini /* free memory */ 119688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 119706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 119806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 119906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 120006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1201d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1202d62866d3SStefano Zampini if (n_vertices) { 1203d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1204d62866d3SStefano Zampini } 1205d62866d3SStefano Zampini if (n_constraints) { 1206d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1207d62866d3SStefano Zampini } 120888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 120988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 121088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1211d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 121288ebb749SStefano Zampini Mat coarse_sub_mat; 121325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 121488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 121588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 121688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 12178bec7fa6SStefano Zampini Mat C_B,CPHI; 12188bec7fa6SStefano Zampini IS is_dummy; 12198bec7fa6SStefano Zampini Vec mones; 122088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 122188ebb749SStefano Zampini PetscReal real_value; 122288ebb749SStefano Zampini 122388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 122488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 122588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 122688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 122788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 122888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1229c0553b1fSStefano Zampini if (unsymmetric_check) { 123088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 123188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 123288ebb749SStefano Zampini } 123388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 123488ebb749SStefano Zampini 123525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 12363301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 123725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1238c0553b1fSStefano Zampini if (unsymmetric_check) { 123988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 124188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 124488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 124788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 124888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 124988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 125088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 125188ebb749SStefano Zampini } else { 125288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 125388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 125488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 125588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 125688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 125788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 125888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 125988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 126088ebb749SStefano Zampini } 126188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 126288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 126388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 126488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 1265d12edf2fSStefano Zampini if (n_benign) { 1266d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1267d12edf2fSStefano Zampini PetscScalar *data,*data2; 1268d12edf2fSStefano Zampini 1269d12edf2fSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&is_dummy);CHKERRQ(ierr); 1270d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 1271d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1272d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1273d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1274d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1275d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 1276d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1277d12edf2fSStefano Zampini data[n_vertices*pcbddc->local_primal_size+i] += data2[i]; 1278d12edf2fSStefano Zampini data[i*pcbddc->local_primal_size+n_vertices] += data2[i]; 1279d12edf2fSStefano Zampini } 1280d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1281d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1282d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1283d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1284d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1285d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1286d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1287d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1288d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1289d12edf2fSStefano Zampini } 1290d12edf2fSStefano Zampini #if 0 1291d12edf2fSStefano Zampini { 1292d12edf2fSStefano Zampini PetscViewer viewer; 1293d12edf2fSStefano Zampini char filename[256]; 1294d12edf2fSStefano Zampini sprintf(filename,"proj_local_coarse_mat%d.m",PetscGlobalRank); 1295d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1296d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1297d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1298d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1299d12edf2fSStefano Zampini } 1300d12edf2fSStefano Zampini #endif 130181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 13028bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 13030fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 130406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 13058bec7fa6SStefano Zampini 13068bec7fa6SStefano Zampini /* check constraints */ 1307d12edf2fSStefano Zampini if (!n_benign) { 13088bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 13098bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 13108bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 13118bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 13128bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 13138bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 13148bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1315bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1316c0553b1fSStefano Zampini if (unsymmetric_check) { 1317bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1318bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1319bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1320bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1321bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 132288ebb749SStefano Zampini } 13238bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 13248bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 13258bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 13268bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1327d12edf2fSStefano Zampini } 132825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 132988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 133088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 133188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 133288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 133388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 133488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 133588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 133688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 133788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 133888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1339c0553b1fSStefano Zampini if (unsymmetric_check) { 134088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 134188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 134288ebb749SStefano Zampini } 134388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 134488ebb749SStefano Zampini } 1345d12edf2fSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 13468629588bSStefano Zampini /* get back data */ 13478629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 134888ebb749SStefano Zampini PetscFunctionReturn(0); 134988ebb749SStefano Zampini } 135088ebb749SStefano Zampini 135188ebb749SStefano Zampini #undef __FUNCT__ 1352d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1353d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1354aa0d41d4SStefano Zampini { 1355d65f70fdSStefano Zampini Mat *work_mat; 1356d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1357d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1358d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1359aa0d41d4SStefano Zampini PetscErrorCode ierr; 1360aa0d41d4SStefano Zampini 1361aa0d41d4SStefano Zampini PetscFunctionBegin; 1362d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1363d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1364d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1365d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1366aa0d41d4SStefano Zampini 1367d65f70fdSStefano Zampini if (!rsorted) { 1368906d46d4SStefano Zampini const PetscInt *idxs; 1369906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1370aa0d41d4SStefano Zampini 1371d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1372d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1373d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1374d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1375aa0d41d4SStefano Zampini } 1376d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1377d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1378d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1379d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1380aa0d41d4SStefano Zampini } 1381d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1382d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1383d65f70fdSStefano Zampini } else { 1384d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1385d65f70fdSStefano Zampini isrow_s = isrow; 1386aa0d41d4SStefano Zampini } 1387906d46d4SStefano Zampini 1388d65f70fdSStefano Zampini if (!csorted) { 1389d65f70fdSStefano Zampini if (isrow == iscol) { 1390d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1391d65f70fdSStefano Zampini iscol_s = isrow_s; 1392d65f70fdSStefano Zampini } else { 1393d65f70fdSStefano Zampini const PetscInt *idxs; 1394d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1395906d46d4SStefano Zampini 1396d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1397d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1398d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1399d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1400d65f70fdSStefano Zampini } 1401d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1402d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1403d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1404d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1405d65f70fdSStefano Zampini } 1406d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1407d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1408d65f70fdSStefano Zampini } 1409d65f70fdSStefano Zampini } else { 1410d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1411d65f70fdSStefano Zampini iscol_s = iscol; 1412d65f70fdSStefano Zampini } 1413d65f70fdSStefano Zampini 1414d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1415d65f70fdSStefano Zampini 1416d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1417906d46d4SStefano Zampini Mat new_mat; 1418d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1419906d46d4SStefano Zampini 1420d65f70fdSStefano Zampini if (!rsorted) { 1421d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1422d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1423d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1424d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1425906d46d4SStefano Zampini } 1426d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1427d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1428d65f70fdSStefano Zampini } else { 1429d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1430906d46d4SStefano Zampini } 1431d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1432d65f70fdSStefano Zampini 1433d65f70fdSStefano Zampini if (!csorted) { 1434d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1435d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1436d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1437d65f70fdSStefano Zampini } else { 1438d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1439d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1440d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1441d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1442d65f70fdSStefano Zampini } 1443d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1444d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1445d65f70fdSStefano Zampini } 1446d65f70fdSStefano Zampini } else { 1447d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1448d65f70fdSStefano Zampini } 1449d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1450d65f70fdSStefano Zampini 1451d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1452d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1453d65f70fdSStefano Zampini work_mat[0] = new_mat; 1454d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1455d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1456d65f70fdSStefano Zampini } 1457d65f70fdSStefano Zampini 1458d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1459d65f70fdSStefano Zampini *B = work_mat[0]; 1460d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1461d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1462d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1463d65f70fdSStefano Zampini PetscFunctionReturn(0); 1464d65f70fdSStefano Zampini } 1465d65f70fdSStefano Zampini 1466d65f70fdSStefano Zampini #undef __FUNCT__ 14675e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 14685e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1469aa0d41d4SStefano Zampini { 1470aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 14713bbff08aSStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 14725e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1473d65f70fdSStefano Zampini Mat new_mat; 14745e8657edSStefano Zampini IS is_local,is_global; 1475d65f70fdSStefano Zampini PetscInt local_size; 1476d65f70fdSStefano Zampini PetscBool isseqaij; 1477aa0d41d4SStefano Zampini PetscErrorCode ierr; 1478aa0d41d4SStefano Zampini 1479aa0d41d4SStefano Zampini PetscFunctionBegin; 1480aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 14815e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 14825e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1483*b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1484aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1485d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1486aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1487906d46d4SStefano Zampini 1488906d46d4SStefano Zampini /* check */ 1489906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1490906d46d4SStefano Zampini Vec x,x_change; 1491906d46d4SStefano Zampini PetscReal error; 1492906d46d4SStefano Zampini 14935e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1494906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 14955e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1496e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1497e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1498d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1499e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1500e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1501906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1502906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1503906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1504906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1505906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1506906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1507906d46d4SStefano Zampini } 1508906d46d4SStefano Zampini 150922d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 15109b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 151122d5777bSStefano Zampini if (isseqaij) { 1512d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1513aa0d41d4SStefano Zampini } else { 1514aa0d41d4SStefano Zampini Mat work_mat; 1515aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1516d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1517aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1518aa0d41d4SStefano Zampini } 15193301b35fSStefano Zampini if (matis->A->symmetric_set) { 15203301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1521e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 15223301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1523e496cd5dSStefano Zampini #endif 15243301b35fSStefano Zampini } 152545a1bb75SStefano Zampini /* 152645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1527d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 152845a1bb75SStefano Zampini */ 1529d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1530aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1531aa0d41d4SStefano Zampini } 1532aa0d41d4SStefano Zampini 1533aa0d41d4SStefano Zampini #undef __FUNCT__ 1534a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 15358ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1536a64d13efSStefano Zampini { 1537a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1538a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1539d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 154053892102SStefano Zampini PetscInt *idx_R_local=NULL; 15413a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 15423a50541eSStefano Zampini PetscInt vbs,bs; 15436816873aSStefano Zampini PetscBT bitmask=NULL; 1544a64d13efSStefano Zampini PetscErrorCode ierr; 1545a64d13efSStefano Zampini 1546a64d13efSStefano Zampini PetscFunctionBegin; 1547b23d619eSStefano Zampini /* 1548b23d619eSStefano Zampini No need to setup local scatters if 1549b23d619eSStefano Zampini - primal space is unchanged 1550b23d619eSStefano Zampini AND 1551b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1552b23d619eSStefano Zampini AND 1553b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1554b23d619eSStefano Zampini */ 1555b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1556f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1557f4ddd8eeSStefano Zampini } 1558f4ddd8eeSStefano Zampini /* destroy old objects */ 1559f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1560f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1561f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1562a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1563b371cd4fSStefano Zampini n_B = pcis->n_B; 1564b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1565b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15663a50541eSStefano Zampini 1567a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 15686816873aSStefano Zampini 156953892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 15706816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1571854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1572a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1573a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 15740e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1575a64d13efSStefano Zampini } 1576a64d13efSStefano Zampini 1577a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 15784641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 15796816873aSStefano Zampini idx_R_local[n_R++] = i; 1580a64d13efSStefano Zampini } 1581a64d13efSStefano Zampini } 158253892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 15836816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15846816873aSStefano Zampini 158553892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 158653892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 15876816873aSStefano Zampini } 15883a50541eSStefano Zampini 15893a50541eSStefano Zampini /* Block code */ 15903a50541eSStefano Zampini vbs = 1; 15913a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 15923a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 15933a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 15943a50541eSStefano Zampini PetscInt *vary; 1595d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1596785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 15973a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1598d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1599d3df7717SStefano 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 */ 16000e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1601d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 16023a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 16033a50541eSStefano Zampini is_blocked = PETSC_FALSE; 16043a50541eSStefano Zampini break; 16053a50541eSStefano Zampini } 16063a50541eSStefano Zampini } 1607d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1608d3df7717SStefano Zampini } else { 1609d3df7717SStefano Zampini /* Verify directly the R set */ 1610d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1611d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1612d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1613d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1614d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1615d3df7717SStefano Zampini break; 1616d3df7717SStefano Zampini } 1617d3df7717SStefano Zampini } 1618d3df7717SStefano Zampini } 1619d3df7717SStefano Zampini } 16203a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 16213a50541eSStefano Zampini vbs = bs; 16223a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 16233a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 16243a50541eSStefano Zampini } 16253a50541eSStefano Zampini } 16263a50541eSStefano Zampini } 16273a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 162853892102SStefano Zampini if (sub_schurs->reuse_mumps) { 162953892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 163053892102SStefano Zampini 163153892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 163253892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 163353892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 163453892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 163553892102SStefano Zampini } else { 16363a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 163753892102SStefano Zampini } 1638a64d13efSStefano Zampini 1639a64d13efSStefano Zampini /* print some info if requested */ 1640a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1641c9ed8603SStefano Zampini PetscInt benign = 0; 1642019a44ceSStefano Zampini 1643c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) benign = 1; 1644a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1645a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16460fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1647a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1648a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1649019a44ceSStefano 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); 1650a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1651a64d13efSStefano Zampini } 1652a64d13efSStefano Zampini 1653a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 16546816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16556816873aSStefano Zampini IS is_aux1,is_aux2; 16566816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 16576816873aSStefano Zampini 16583a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1659854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1660854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1661a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 16624641a718SStefano Zampini for (i=0; i<n_D; i++) { 16634641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 16644641a718SStefano Zampini } 1665a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1666a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16674641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 16684641a718SStefano Zampini aux_array1[j++] = i; 1669a64d13efSStefano Zampini } 1670a64d13efSStefano Zampini } 1671a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1672a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1673a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 16744641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 16754641a718SStefano Zampini aux_array2[j++] = i; 1676a64d13efSStefano Zampini } 1677a64d13efSStefano Zampini } 1678a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1679a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1680a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1681a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1682a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1683a64d13efSStefano Zampini 16848eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1685785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1686a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16874641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 16884641a718SStefano Zampini aux_array1[j++] = i; 1689a64d13efSStefano Zampini } 1690a64d13efSStefano Zampini } 1691a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1692a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1693a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1694a64d13efSStefano Zampini } 16954641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 16963a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1697d62866d3SStefano Zampini } else { 169853892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16996816873aSStefano Zampini IS tis; 17006816873aSStefano Zampini PetscInt schur_size; 17016816873aSStefano Zampini 170253892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 17036816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 170453892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 17056816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17066816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 17076816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 17086816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 17096816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1710d62866d3SStefano Zampini } 1711d62866d3SStefano Zampini } 1712a64d13efSStefano Zampini PetscFunctionReturn(0); 1713a64d13efSStefano Zampini } 1714a64d13efSStefano Zampini 1715304d26faSStefano Zampini 1716304d26faSStefano Zampini #undef __FUNCT__ 1717304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1718684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1719304d26faSStefano Zampini { 1720304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1721304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1722304d26faSStefano Zampini PC pc_temp; 1723304d26faSStefano Zampini Mat A_RR; 1724f4ddd8eeSStefano Zampini MatReuse reuse; 1725304d26faSStefano Zampini PetscScalar m_one = -1.0; 1726304d26faSStefano Zampini PetscReal value; 172704708bb6SStefano Zampini PetscInt n_D,n_R; 17289577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1729304d26faSStefano Zampini PetscErrorCode ierr; 1730e604994aSStefano Zampini /* prefixes stuff */ 1731312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1732e604994aSStefano Zampini size_t len; 1733304d26faSStefano Zampini 1734304d26faSStefano Zampini PetscFunctionBegin; 1735304d26faSStefano Zampini 1736e604994aSStefano Zampini /* compute prefixes */ 1737e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1738e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1739e604994aSStefano Zampini if (!pcbddc->current_level) { 1740e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1741e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1742e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1743e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1744e604994aSStefano Zampini } else { 1745e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1746312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1747e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1748e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1749312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1750312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 175134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 175234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1753e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1754e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1755e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1756e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1757e604994aSStefano Zampini } 1758e604994aSStefano Zampini 1759304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1760684f6988SStefano Zampini if (dirichlet) { 1761d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 17623301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 17633301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1764964fefecSStefano Zampini } 1765ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1766964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1767304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1768304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1769304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1770304d26faSStefano Zampini /* default */ 1771304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1772e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 17739577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1774304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 17759577ea80SStefano Zampini if (issbaij) { 17769577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 17779577ea80SStefano Zampini } else { 1778304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 17799577ea80SStefano Zampini } 1780304d26faSStefano Zampini /* Allow user's customization */ 1781304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1782304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1783304d26faSStefano Zampini } 1784d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1785d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1786d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1787d62866d3SStefano Zampini 1788d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1789d5574798SStefano Zampini } 1790304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1791304d26faSStefano Zampini if (!n_D) { 1792304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1793304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1794304d26faSStefano Zampini } 1795304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1796304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1797304d26faSStefano Zampini /* set ksp_D into pcis data */ 1798304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1799304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1800304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1801684f6988SStefano Zampini } 1802304d26faSStefano Zampini 1803304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1804684f6988SStefano Zampini A_RR = 0; 1805684f6988SStefano Zampini if (neumann) { 1806d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 180704708bb6SStefano Zampini PetscInt ibs,mbs; 180804708bb6SStefano Zampini PetscBool issbaij; 180904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1810f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 18118ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1812f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1813f4ddd8eeSStefano Zampini PetscInt nn_R; 181481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1815f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1816f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1817f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1818f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1819f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1820f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1821f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1822727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1823f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1824f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1825f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1826f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1827f4ddd8eeSStefano Zampini } 1828f4ddd8eeSStefano Zampini } 1829f4ddd8eeSStefano Zampini /* last check */ 1830d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1831f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1832f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1833f4ddd8eeSStefano Zampini } 1834f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1835f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1836f4ddd8eeSStefano Zampini } 1837f4ddd8eeSStefano Zampini /* extract A_RR */ 1838af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1839af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 184004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 184104708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 184204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 184304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 184404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1845af732b37SStefano Zampini } else { 184604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 18476816873aSStefano Zampini } 184804708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 184904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 185004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 185104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 185204708bb6SStefano Zampini } else { 185304708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 185404708bb6SStefano Zampini } 185504708bb6SStefano Zampini } 185604708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1857f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 18583301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 18593301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 18606816873aSStefano Zampini } 18616816873aSStefano Zampini } else { 18626816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 18636816873aSStefano Zampini 18646816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 18656816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 18666816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1867af732b37SStefano Zampini } 1868f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1869304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1870304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1871304d26faSStefano Zampini /* default */ 1872304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1873e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1874304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 18759577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 18769577ea80SStefano Zampini if (issbaij) { 18779577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 18789577ea80SStefano Zampini } else { 1879304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 18809577ea80SStefano Zampini } 1881304d26faSStefano Zampini /* Allow user's customization */ 1882304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1883304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1884304d26faSStefano Zampini } 1885d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1886304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1887304d26faSStefano Zampini if (!n_R) { 1888304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1889304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1890304d26faSStefano Zampini } 1891d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1892d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1893d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1894d62866d3SStefano Zampini 1895d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1896d62866d3SStefano Zampini } 1897304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1898304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1899684f6988SStefano Zampini } 19006816873aSStefano Zampini /* free Neumann problem's matrix */ 19016816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1902304d26faSStefano Zampini 1903304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 19040fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1905684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1906684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1907684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1908684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1909684f6988SStefano Zampini } 1910684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 19110fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 19120fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 19130fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 19140fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 19150fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1916304d26faSStefano Zampini /* need to be adapted? */ 1917b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1918b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1919b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1920304d26faSStefano Zampini /* print info */ 1921304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1922e604994aSStefano 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); 1923304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1924304d26faSStefano Zampini } 1925b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1926298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1927304d26faSStefano Zampini } 1928684f6988SStefano Zampini } 1929684f6988SStefano Zampini if (neumann) { /* Neumann */ 19306816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 19310fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 19320fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 19330fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 19340fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 19350fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1936304d26faSStefano Zampini /* need to be adapted? */ 1937b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1938b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1939304d26faSStefano Zampini /* print info */ 1940304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1941e604994aSStefano 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); 1942304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1943304d26faSStefano Zampini } 1944b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1945298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1946304d26faSStefano Zampini } 19470fccc4e9SStefano Zampini } 1948684f6988SStefano Zampini } 1949304d26faSStefano Zampini PetscFunctionReturn(0); 1950304d26faSStefano Zampini } 1951304d26faSStefano Zampini 1952304d26faSStefano Zampini #undef __FUNCT__ 1953ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 195480677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1955674ae819SStefano Zampini { 1956674ae819SStefano Zampini PetscErrorCode ierr; 1957674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1958be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1959674ae819SStefano Zampini 1960674ae819SStefano Zampini PetscFunctionBegin; 1961be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 196280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 196320c7b377SStefano Zampini } 196480677318SStefano Zampini if (!pcbddc->switch_static) { 196580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 196680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 196780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 196820c7b377SStefano Zampini } 1969be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 197080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197220c7b377SStefano Zampini } else { 1973be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1974be83ff47SStefano Zampini 197553892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197653892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197720c7b377SStefano Zampini } 1978be83ff47SStefano Zampini } else { 197980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 198080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 198180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 198280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 198380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 198480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 198580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 198680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 198780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1988674ae819SStefano Zampini } 1989674ae819SStefano Zampini } 1990be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 199180677318SStefano Zampini if (applytranspose) { 199280677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 199380677318SStefano Zampini } else { 199480677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 199580677318SStefano Zampini } 199653892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1997be83ff47SStefano Zampini } else { 1998be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1999be83ff47SStefano Zampini 2000be83ff47SStefano Zampini if (applytranspose) { 200153892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2002be83ff47SStefano Zampini } else { 200353892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 2004be83ff47SStefano Zampini } 200553892102SStefano Zampini #endif 2006be83ff47SStefano Zampini } 200780677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 200880677318SStefano Zampini if (!pcbddc->switch_static) { 2009be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 201080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2012be83ff47SStefano Zampini } else { 2013be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 2014be83ff47SStefano Zampini 201553892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 201653892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2017be83ff47SStefano Zampini } 201880677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 201980677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 202080677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 202180677318SStefano Zampini } 202280677318SStefano Zampini } else { 202380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 202880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 202980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 203080677318SStefano Zampini } 203180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2035674ae819SStefano Zampini } 2036674ae819SStefano Zampini PetscFunctionReturn(0); 2037674ae819SStefano Zampini } 2038674ae819SStefano Zampini 2039dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2040674ae819SStefano Zampini #undef __FUNCT__ 2041674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2042dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2043674ae819SStefano Zampini { 2044674ae819SStefano Zampini PetscErrorCode ierr; 2045674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2046674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2047674ae819SStefano Zampini const PetscScalar zero = 0.0; 2048674ae819SStefano Zampini 2049674ae819SStefano Zampini PetscFunctionBegin; 2050dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2051dc359a40SStefano Zampini if (applytranspose) { 2052674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 20538eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2054dc359a40SStefano Zampini } else { 2055674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2056674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 205715aaf578SStefano Zampini } 2058efc2fbd9SStefano Zampini 2059efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 2060c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 2061efc2fbd9SStefano Zampini PetscScalar *array; 2062efc2fbd9SStefano Zampini 2063efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2064efc2fbd9SStefano Zampini array[pcbddc->local_primal_size-1] += pcbddc->benign_p0; 2065efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2066efc2fbd9SStefano Zampini } 2067efc2fbd9SStefano Zampini 206812edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 206912edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 207012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207212edc857SStefano Zampini 20739f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 207412edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 207512edc857SStefano Zampini if (pcbddc->coarse_ksp) { 2076964fefecSStefano Zampini Vec rhs,sol; 2077964fefecSStefano Zampini 2078964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2079964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 208012edc857SStefano Zampini if (applytranspose) { 2081964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 208212edc857SStefano Zampini } else { 2083964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 208412edc857SStefano Zampini } 208512edc857SStefano Zampini } 2086674ae819SStefano Zampini 2087674ae819SStefano Zampini /* Local solution on R nodes */ 208880677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 208980677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 20909f00e9b4SStefano Zampini } 2091674ae819SStefano Zampini 20929f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 20939f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 209412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2095674ae819SStefano Zampini 2096674ae819SStefano Zampini /* Sum contributions from two levels */ 2097dc359a40SStefano Zampini if (applytranspose) { 2098dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2099dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2100dc359a40SStefano Zampini } else { 2101674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 21028eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2103dc359a40SStefano Zampini } 2104efc2fbd9SStefano Zampini /* store p0 */ 2105c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 2106efc2fbd9SStefano Zampini PetscScalar *array; 2107efc2fbd9SStefano Zampini 2108efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2109efc2fbd9SStefano Zampini pcbddc->benign_p0 = array[pcbddc->local_primal_size-1]; 2110efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2111efc2fbd9SStefano Zampini } 2112674ae819SStefano Zampini PetscFunctionReturn(0); 2113674ae819SStefano Zampini } 2114674ae819SStefano Zampini 2115674ae819SStefano Zampini #undef __FUNCT__ 2116674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 211712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2118674ae819SStefano Zampini { 2119674ae819SStefano Zampini PetscErrorCode ierr; 2120674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 212158da7f69SStefano Zampini PetscScalar *array; 212212edc857SStefano Zampini Vec from,to; 2123674ae819SStefano Zampini 2124674ae819SStefano Zampini PetscFunctionBegin; 212512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 212612edc857SStefano Zampini from = pcbddc->coarse_vec; 212712edc857SStefano Zampini to = pcbddc->vec1_P; 212812edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 212912edc857SStefano Zampini Vec tvec; 213058da7f69SStefano Zampini 213158da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 213258da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 213312edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 213458da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 213558da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 213658da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 213712edc857SStefano Zampini } 213812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 213912edc857SStefano Zampini from = pcbddc->vec1_P; 214012edc857SStefano Zampini to = pcbddc->coarse_vec; 214112edc857SStefano Zampini } 214212edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2143674ae819SStefano Zampini PetscFunctionReturn(0); 2144674ae819SStefano Zampini } 2145674ae819SStefano Zampini 2146674ae819SStefano Zampini #undef __FUNCT__ 2147674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 214812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2149674ae819SStefano Zampini { 2150674ae819SStefano Zampini PetscErrorCode ierr; 2151674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 215258da7f69SStefano Zampini PetscScalar *array; 215312edc857SStefano Zampini Vec from,to; 2154674ae819SStefano Zampini 2155674ae819SStefano Zampini PetscFunctionBegin; 215612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 215712edc857SStefano Zampini from = pcbddc->coarse_vec; 215812edc857SStefano Zampini to = pcbddc->vec1_P; 215912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 216012edc857SStefano Zampini from = pcbddc->vec1_P; 216112edc857SStefano Zampini to = pcbddc->coarse_vec; 216212edc857SStefano Zampini } 216312edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 216412edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 216512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 216612edc857SStefano Zampini Vec tvec; 216758da7f69SStefano Zampini 216812edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 216958da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 217058da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 217158da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 217258da7f69SStefano Zampini } 217358da7f69SStefano Zampini } else { 217458da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 217558da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 217612edc857SStefano Zampini } 217712edc857SStefano Zampini } 2178674ae819SStefano Zampini PetscFunctionReturn(0); 2179674ae819SStefano Zampini } 2180674ae819SStefano Zampini 2181984c4197SStefano Zampini /* uncomment for testing purposes */ 2182984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2183674ae819SStefano Zampini #undef __FUNCT__ 2184674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2185674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2186674ae819SStefano Zampini { 2187674ae819SStefano Zampini PetscErrorCode ierr; 2188674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2189674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2190674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2191984c4197SStefano Zampini /* one and zero */ 2192984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2193984c4197SStefano Zampini /* space to store constraints and their local indices */ 21949162d606SStefano Zampini PetscScalar *constraints_data; 21959162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 21969162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 21979162d606SStefano Zampini PetscInt *constraints_n; 2198984c4197SStefano Zampini /* iterators */ 2199b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2200984c4197SStefano Zampini /* BLAS integers */ 2201e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2202e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2203c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2204727cdba6SStefano Zampini /* reuse */ 22050e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 22060e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2207984c4197SStefano Zampini /* change of basis */ 2208b3d85658SStefano Zampini PetscBool qr_needed; 22099162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2210984c4197SStefano Zampini /* auxiliary stuff */ 221164efe560SStefano Zampini PetscInt *nnz,*is_indices; 22128a0068c3SStefano Zampini PetscInt ncc; 2213984c4197SStefano Zampini /* some quantities */ 221445a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2215a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2216984c4197SStefano Zampini 2217674ae819SStefano Zampini PetscFunctionBegin; 22188e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 22198e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 22208e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2221088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2222088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 22230e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 22240e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 22250e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 22260e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 22270e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2228088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2229cf5a6209SStefano Zampini 2230cf5a6209SStefano Zampini /* print some info */ 2231cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2232cf5a6209SStefano Zampini IS vertices; 2233cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2234cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2235cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2236cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2237cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2238cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2239cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2240fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2241fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2242cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2243cf5a6209SStefano Zampini } 2244cf5a6209SStefano Zampini 2245cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 22469162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2247cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2248cf5a6209SStefano Zampini const Vec *nearnullvecs; 2249cf5a6209SStefano Zampini Vec *localnearnullsp; 2250cf5a6209SStefano Zampini PetscScalar *array; 2251cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2252cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2253674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2254b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2255674ae819SStefano Zampini PetscScalar *work; 2256674ae819SStefano Zampini PetscReal *singular_vals; 2257674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2258674ae819SStefano Zampini PetscReal *rwork; 2259674ae819SStefano Zampini #endif 2260674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2261674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2262674ae819SStefano Zampini #else 2263964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2264964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2265674ae819SStefano Zampini #endif 2266674ae819SStefano Zampini 2267674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2268d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2269d06fc5fdSStefano Zampini /* free unneeded index sets */ 2270d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2271d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2272674ae819SStefano Zampini } 2273d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2274d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2275d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2276d06fc5fdSStefano Zampini } 2277d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2278d06fc5fdSStefano Zampini n_ISForEdges = 0; 2279d06fc5fdSStefano Zampini } 2280d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2281d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2282d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2283d06fc5fdSStefano Zampini } 2284d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2285d06fc5fdSStefano Zampini n_ISForFaces = 0; 2286d06fc5fdSStefano Zampini } 228770022509SStefano Zampini 228870022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 228970022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 229070022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 229170022509SStefano Zampini if (pcbddc->NullSpace) { 229270022509SStefano Zampini PetscBool tbool[2],gbool[2]; 229370022509SStefano Zampini 229470022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2295b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2296d06fc5fdSStefano Zampini if (!ISForEdges) { 2297d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2298d06fc5fdSStefano Zampini } 2299b8ffe317SStefano Zampini } 2300d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2301d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2302d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2303d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2304d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 230598a51de6SStefano Zampini } 230670022509SStefano Zampini #endif 230708122e43SStefano Zampini 2308674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2309674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2310674ae819SStefano Zampini if (nearnullsp) { 2311674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2312f4ddd8eeSStefano Zampini /* remove any stored info */ 2313f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2314f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2315f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2316f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2317f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2318473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2319f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2320f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2321f4ddd8eeSStefano Zampini } 2322984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2323984c4197SStefano Zampini nnsp_size = 0; 2324674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2325674ae819SStefano Zampini } 2326984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2327984c4197SStefano Zampini max_constraints = nnsp_size; 2328984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2329984c4197SStefano Zampini 2330674ae819SStefano Zampini /* 2331674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 23329162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 23339162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 23349162d606SStefano Zampini There can be multiple constraints per connected component 2335674ae819SStefano Zampini */ 2336674ae819SStefano Zampini n_vertices = 0; 2337674ae819SStefano Zampini if (ISForVertices) { 2338674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2339674ae819SStefano Zampini } 23409162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 23419162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 23429162d606SStefano Zampini 23439162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 23449162d606SStefano Zampini total_counts *= max_constraints; 2345674ae819SStefano Zampini total_counts += n_vertices; 23464641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 23479162d606SStefano Zampini 2348674ae819SStefano Zampini total_counts = 0; 2349674ae819SStefano Zampini max_size_of_constraint = 0; 2350674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 23519162d606SStefano Zampini IS used_is; 2352674ae819SStefano Zampini if (i<n_ISForEdges) { 23539162d606SStefano Zampini used_is = ISForEdges[i]; 2354674ae819SStefano Zampini } else { 23559162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2356674ae819SStefano Zampini } 23579162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2358674ae819SStefano Zampini total_counts += j; 2359674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2360674ae819SStefano Zampini } 23619162d606SStefano 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); 23629162d606SStefano Zampini 2363984c4197SStefano Zampini /* get local part of global near null space vectors */ 2364785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2365984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2366984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2367e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2368e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2369984c4197SStefano Zampini } 2370674ae819SStefano Zampini 2371242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2372242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2373a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2374242a89d7SStefano Zampini 2375984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2376a773dcb8SStefano Zampini if (!skip_lapack) { 2377674ae819SStefano Zampini PetscScalar temp_work; 2378911cabfeSStefano Zampini 2379674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2380984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2381785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2382785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2383785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2384674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2385785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2386674ae819SStefano Zampini #endif 2387674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2388c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2389c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2390674ae819SStefano Zampini lwork = -1; 2391674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2392674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2393c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2394674ae819SStefano Zampini #else 2395c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2396674ae819SStefano Zampini #endif 2397674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2398984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2399674ae819SStefano Zampini #else /* on missing GESVD */ 2400674ae819SStefano Zampini /* SVD */ 2401674ae819SStefano Zampini PetscInt max_n,min_n; 2402674ae819SStefano Zampini max_n = max_size_of_constraint; 2403984c4197SStefano Zampini min_n = max_constraints; 2404984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2405674ae819SStefano Zampini min_n = max_size_of_constraint; 2406984c4197SStefano Zampini max_n = max_constraints; 2407674ae819SStefano Zampini } 2408785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2409674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2410785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2411674ae819SStefano Zampini #endif 2412674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2413674ae819SStefano Zampini lwork = -1; 2414e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2415e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2416b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2417674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2418674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24199162d606SStefano 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)); 2420674ae819SStefano Zampini #else 24219162d606SStefano 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)); 2422674ae819SStefano Zampini #endif 2423674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2424984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2425984c4197SStefano Zampini #endif /* on missing GESVD */ 2426674ae819SStefano Zampini /* Allocate optimal workspace */ 2427674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2428854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2429674ae819SStefano Zampini } 2430674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2431674ae819SStefano Zampini total_counts = 0; 24329162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 24339162d606SStefano Zampini constraints_data_ptr[0] = 0; 2434674ae819SStefano Zampini /* vertices */ 24359162d606SStefano Zampini if (n_vertices) { 2436674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 24379162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 24389162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2439674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 24409162d606SStefano Zampini constraints_n[total_counts] = 1; 24419162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24429162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24439162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2444674ae819SStefano Zampini total_counts++; 2445674ae819SStefano Zampini } 2446674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2447984c4197SStefano Zampini PetscBool used_vertex; 2448674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2449674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2450674ae819SStefano Zampini k = 0; 2451674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2452984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2453984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 24549162d606SStefano Zampini constraints_n[total_counts] = 1; 24559162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 24569162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24579162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24589162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2459674ae819SStefano Zampini total_counts++; 2460674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2461674ae819SStefano Zampini } 2462984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2463674ae819SStefano Zampini k++; 2464674ae819SStefano Zampini } 2465674ae819SStefano Zampini } 2466674ae819SStefano Zampini } 2467674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2468674ae819SStefano Zampini n_vertices = total_counts; 2469674ae819SStefano Zampini } 2470984c4197SStefano Zampini 2471674ae819SStefano Zampini /* edges and faces */ 24729162d606SStefano Zampini total_counts_cc = total_counts; 2473911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 24749162d606SStefano Zampini IS used_is; 24759162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 24769162d606SStefano Zampini 2477911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 24789162d606SStefano Zampini used_is = ISForEdges[ncc]; 2479984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2480674ae819SStefano Zampini } else { 24819162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2482984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2483674ae819SStefano Zampini } 2484674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 24859162d606SStefano Zampini 24869162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 24879162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2488984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2489984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2490674ae819SStefano Zampini if (nnsp_has_cnst) { 24915b08dc53SStefano Zampini PetscScalar quad_value; 24929162d606SStefano Zampini 24939162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24949162d606SStefano Zampini idxs_copied = PETSC_TRUE; 24959162d606SStefano Zampini 2496a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2497674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2498a773dcb8SStefano Zampini } else { 2499a773dcb8SStefano Zampini quad_value = 1.0; 2500a773dcb8SStefano Zampini } 2501674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 25029162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2503674ae819SStefano Zampini } 25049162d606SStefano Zampini temp_constraints++; 2505674ae819SStefano Zampini total_counts++; 2506674ae819SStefano Zampini } 2507674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2508984c4197SStefano Zampini PetscReal real_value; 25099162d606SStefano Zampini PetscScalar *ptr_to_data; 25109162d606SStefano Zampini 2511984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 25129162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2513674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 25149162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2515674ae819SStefano Zampini } 2516984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2517984c4197SStefano Zampini /* check if array is null on the connected component */ 2518e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 25199162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 25205b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2521674ae819SStefano Zampini temp_constraints++; 2522674ae819SStefano Zampini total_counts++; 25239162d606SStefano Zampini if (!idxs_copied) { 25249162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 25259162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2526674ae819SStefano Zampini } 2527674ae819SStefano Zampini } 25289162d606SStefano Zampini } 25299162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 253045a1bb75SStefano Zampini valid_constraints = temp_constraints; 2531eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2532a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 25339162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 25349162d606SStefano Zampini 25359162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2536a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 25379162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2538a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 25399162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2540a773dcb8SStefano Zampini } else { /* perform SVD */ 2541984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 25429162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2543674ae819SStefano Zampini 2544674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2545984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2546984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2547984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2548984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2549984c4197SStefano Zampini from that computed using LAPACKgesvd 2550984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2551984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2552984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2553674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2554e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2555984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2556674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2557674ae819SStefano Zampini for (k=0;k<j+1;k++) { 25589162d606SStefano 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)); 2559674ae819SStefano Zampini } 2560674ae819SStefano Zampini } 2561e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2562e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2563e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2564674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2565c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2566674ae819SStefano Zampini #else 2567c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2568674ae819SStefano Zampini #endif 2569674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2570984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2571984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2572674ae819SStefano Zampini j = 0; 2573984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2574674ae819SStefano Zampini total_counts = total_counts-j; 257545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2576e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2577c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2578c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2579c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2580c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2581c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2582c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2583674ae819SStefano Zampini if (j<temp_constraints) { 2584984c4197SStefano Zampini PetscInt ii; 2585984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2586674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25879162d606SStefano 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)); 2588674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2589984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2590674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 25919162d606SStefano 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]; 2592674ae819SStefano Zampini } 2593674ae819SStefano Zampini } 2594674ae819SStefano Zampini } 2595674ae819SStefano Zampini #else /* on missing GESVD */ 2596e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2597e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2598b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2599674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2600674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 26019162d606SStefano 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)); 2602674ae819SStefano Zampini #else 26039162d606SStefano 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)); 2604674ae819SStefano Zampini #endif 2605984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2606674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2607984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2608e310c8b4SStefano Zampini k = temp_constraints; 2609e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2610674ae819SStefano Zampini j = 0; 2611e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 261245a1bb75SStefano Zampini valid_constraints = k-j; 2613911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2614984c4197SStefano Zampini #endif /* on missing GESVD */ 2615674ae819SStefano Zampini } 2616a773dcb8SStefano Zampini } 26179162d606SStefano Zampini /* update pointers information */ 26189162d606SStefano Zampini if (valid_constraints) { 26199162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 26209162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 26219162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 26229162d606SStefano Zampini /* set change_of_basis flag */ 262345a1bb75SStefano Zampini if (boolforchange) { 2624b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 26259162d606SStefano Zampini } 2626b3d85658SStefano Zampini total_counts_cc++; 262745a1bb75SStefano Zampini } 262845a1bb75SStefano Zampini } 2629984c4197SStefano Zampini /* free workspace */ 26308f1c130eSStefano Zampini if (!skip_lapack) { 2631984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2632984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2633984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2634984c4197SStefano Zampini #endif 2635984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2636984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2637984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2638984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2639984c4197SStefano Zampini #endif 2640984c4197SStefano Zampini } 2641984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2642984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2643984c4197SStefano Zampini } 2644984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2645cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2646cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2647cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2648cf5a6209SStefano Zampini } 2649cf5a6209SStefano Zampini if (n_ISForFaces) { 2650cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2651cf5a6209SStefano Zampini } 2652cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2653cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2654cf5a6209SStefano Zampini } 2655cf5a6209SStefano Zampini if (n_ISForEdges) { 2656cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2657cf5a6209SStefano Zampini } 2658cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 265908122e43SStefano Zampini } else { 266008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2661984c4197SStefano Zampini 266208122e43SStefano Zampini total_counts = 0; 266308122e43SStefano Zampini n_vertices = 0; 2664d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2665d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 266608122e43SStefano Zampini } 266708122e43SStefano Zampini max_constraints = 0; 26689162d606SStefano Zampini total_counts_cc = 0; 266908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 267008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 26719162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 267208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 267308122e43SStefano Zampini } 26749162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 26759162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 26769162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 26779162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 267874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 26799162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 26809162d606SStefano Zampini total_counts_cc = 0; 26819162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 26829162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 26839162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 268408122e43SStefano Zampini } 268508122e43SStefano Zampini } 26869162d606SStefano Zampini #if 0 26879162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 26889162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26899162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 26909162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 26919162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 26929162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 26939162d606SStefano Zampini } 26949162d606SStefano Zampini printf("\n"); 26959162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 26969162d606SStefano Zampini } 26971b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 26988bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 26991b968477SStefano Zampini } 27001b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 27018bec7fa6SStefano 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]); 27021b968477SStefano Zampini } 270308122e43SStefano Zampini #endif 270408122e43SStefano Zampini 27058bec7fa6SStefano Zampini max_size_of_constraint = 0; 27069162d606SStefano 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]); 27079162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 270808122e43SStefano Zampini /* Change of basis */ 2709b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 271008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 271108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 271208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2713b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 271408122e43SStefano Zampini } 271508122e43SStefano Zampini } 271608122e43SStefano Zampini } 271708122e43SStefano Zampini } 2718984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2719019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2720019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 272108122e43SStefano Zampini 27229162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 27239162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 27249162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 27259162d606SStefano 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); 272608122e43SStefano Zampini } 2727674ae819SStefano Zampini 2728674ae819SStefano Zampini /* Create constraint matrix */ 2729674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 273016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2731984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2732984c4197SStefano Zampini 2733984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2734a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2735a717540cSStefano Zampini qr_needed = PETSC_FALSE; 273674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2737984c4197SStefano Zampini total_primal_vertices=0; 2738b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 27399162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 27409162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2741984c4197SStefano Zampini if (size_of_constraint == 1) { 27429162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2743b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 274464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 27459162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27469162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2747a717540cSStefano Zampini } 2748b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 274974d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2750a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2751a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2752a717540cSStefano Zampini } 2753fa434743SStefano Zampini } else { 2754b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2755fa434743SStefano Zampini } 2756a717540cSStefano Zampini } 2757b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2758b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2759674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 276070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2761b3d85658SStefano Zampini 2762019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2763019a44ceSStefano 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); 27640e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 27650e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2766984c4197SStefano Zampini 2767984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 276874d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2769785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2770984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 277174d5cdf7SStefano Zampini 2772984c4197SStefano Zampini j = total_primal_vertices; 277374d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2774b3d85658SStefano Zampini cum = total_primal_vertices; 27759162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27764641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2777b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2778b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2779b3d85658SStefano Zampini cum++; 27809162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 278174d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 278274d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 278374d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 278474d5cdf7SStefano Zampini } 27859162d606SStefano Zampini j += constraints_n[i]; 2786674ae819SStefano Zampini } 2787674ae819SStefano Zampini } 2788674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2789674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2790088faed8SStefano Zampini 2791674ae819SStefano Zampini /* set values in constraint matrix */ 2792984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 27930e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2794674ae819SStefano Zampini } 2795984c4197SStefano Zampini total_counts = total_primal_vertices; 27969162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27974641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 27989162d606SStefano Zampini PetscInt *cols; 27999162d606SStefano Zampini 28009162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 28019162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 28029162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 28039162d606SStefano Zampini PetscInt row = total_counts+k; 28049162d606SStefano Zampini PetscScalar *vals; 28059162d606SStefano Zampini 28069162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 28079162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 28089162d606SStefano Zampini } 28099162d606SStefano Zampini total_counts += constraints_n[i]; 2810674ae819SStefano Zampini } 2811674ae819SStefano Zampini } 2812674ae819SStefano Zampini /* assembling */ 2813674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2814674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2815088faed8SStefano Zampini 2816984c4197SStefano Zampini /* 281745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2818984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2819984c4197SStefano Zampini */ 2820674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2821674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2822026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2823984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2824984c4197SStefano Zampini /* working stuff for GEQRF */ 282581d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2826984c4197SStefano Zampini PetscBLASInt lqr_work; 2827984c4197SStefano Zampini /* working stuff for UNGQR */ 2828984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2829984c4197SStefano Zampini PetscBLASInt lgqr_work; 2830984c4197SStefano Zampini /* working stuff for TRTRS */ 2831984c4197SStefano Zampini PetscScalar *trs_rhs; 28323f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2833984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2834984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2835984c4197SStefano Zampini PetscScalar *start_vals; 2836984c4197SStefano Zampini /* working stuff for values insertion */ 28374641a718SStefano Zampini PetscBT is_primal; 283864efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2839906d46d4SStefano Zampini /* matrix sizes */ 2840906d46d4SStefano Zampini PetscInt global_size,local_size; 2841906d46d4SStefano Zampini /* temporary change of basis */ 2842906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2843cf5a6209SStefano Zampini /* extra space for debugging */ 2844cf5a6209SStefano Zampini PetscScalar *dbg_work; 2845984c4197SStefano Zampini 2846906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2847906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 284816f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2849bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2850906d46d4SStefano Zampini /* nonzeros for local mat */ 2851bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2852bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 28539162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2854a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 28559162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2856a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 28579162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2858a717540cSStefano Zampini } else { 28599162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 28609162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2861a717540cSStefano Zampini } 2862a717540cSStefano Zampini } 2863a717540cSStefano Zampini } 2864906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2866a717540cSStefano Zampini /* Set initial identity in the matrix */ 2867bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2868906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2869a717540cSStefano Zampini } 2870a717540cSStefano Zampini 2871a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2872a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2873a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2874a717540cSStefano Zampini } 2875a717540cSStefano Zampini 2876a717540cSStefano Zampini 2877a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2878a717540cSStefano Zampini /* 2879a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2880a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2881a717540cSStefano Zampini 2882a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2883a717540cSStefano Zampini 2884a6b551f4SStefano 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) 2885a6b551f4SStefano Zampini 2886a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2887a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2888a717540cSStefano Zampini | ... | 2889a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2890a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2891a717540cSStefano Zampini 2892a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2893a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2894a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2895a6b551f4SStefano Zampini 2896a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2897a717540cSStefano Zampini */ 2898a717540cSStefano Zampini if (qr_needed) { 2899984c4197SStefano Zampini /* space to store Q */ 2900854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2901984c4197SStefano Zampini /* first we issue queries for optimal work */ 29023f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 29033f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 29043f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2905984c4197SStefano Zampini lqr_work = -1; 29063f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2907984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2908984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2909785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2910984c4197SStefano Zampini lgqr_work = -1; 29113f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 29123f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 29133f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 29143f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29153f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 29163f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2917984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2918984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2919785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2920984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2921785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2922984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2923785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2924a717540cSStefano Zampini /* allocating workspace for check */ 2925a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2926cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2927a717540cSStefano Zampini } 2928a717540cSStefano Zampini } 2929984c4197SStefano Zampini /* array to store whether a node is primal or not */ 29304641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2931473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 29320e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 293339e2fb2aSStefano Zampini if (i != total_primal_vertices) { 293439e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 29354641a718SStefano Zampini } 293639e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 293739e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 293839e2fb2aSStefano Zampini } 293939e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2940984c4197SStefano Zampini 2941a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 29429162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 29439162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 29444641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2945984c4197SStefano Zampini /* get constraint info */ 29469162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2947984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2948984c4197SStefano Zampini 2949984c4197SStefano Zampini if (pcbddc->dbg_flag) { 29509162d606SStefano 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); 2951674ae819SStefano Zampini } 2952984c4197SStefano Zampini 2953fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2954a717540cSStefano Zampini 2955a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2956a717540cSStefano Zampini if (pcbddc->dbg_flag) { 29579162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2958a717540cSStefano Zampini } 2959984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 29609162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2961984c4197SStefano Zampini 2962984c4197SStefano Zampini /* compute QR decomposition of constraints */ 29633f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29643f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29653f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2966674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29673f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2968984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2969674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2970984c4197SStefano Zampini 2971984c4197SStefano Zampini /* explictly compute R^-T */ 2972984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2973984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 29743f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29753f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 29763f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29773f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2978984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29793f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2980984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2981984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2982984c4197SStefano Zampini 2983a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 29843f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29853f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29863f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29873f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2988984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29893f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2990984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2991984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2992984c4197SStefano Zampini 2993984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2994984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2995984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 29963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29973f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29983f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 30003f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 30013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3002984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 30039162d606SStefano 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)); 3004984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 30059162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3006984c4197SStefano Zampini 3007984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 30089162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3009984c4197SStefano Zampini /* insert cols for primal dofs */ 3010984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3011984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 30129162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3013906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3014984c4197SStefano Zampini } 3015984c4197SStefano Zampini /* insert cols for dual dofs */ 3016984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 30179162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3018984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 30199162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3020906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3021984c4197SStefano Zampini j++; 3022674ae819SStefano Zampini } 3023674ae819SStefano Zampini } 3024984c4197SStefano Zampini 3025984c4197SStefano Zampini /* check change of basis */ 3026984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3027984c4197SStefano Zampini PetscInt ii,jj; 3028984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3029c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3030c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3031c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3032c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3033c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3034c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3035984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3036cf5a6209SStefano 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)); 3037984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3038984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3039984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3040cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3041cf5a6209SStefano 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; 3042674ae819SStefano Zampini } 3043674ae819SStefano Zampini } 3044984c4197SStefano Zampini if (!valid_qr) { 304522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3046984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3047984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3048cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3049cf5a6209SStefano 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])); 3050674ae819SStefano Zampini } 3051cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3052cf5a6209SStefano 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])); 3053984c4197SStefano Zampini } 3054984c4197SStefano Zampini } 3055984c4197SStefano Zampini } 3056674ae819SStefano Zampini } else { 305722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3058674ae819SStefano Zampini } 3059674ae819SStefano Zampini } 3060a717540cSStefano Zampini } else { /* simple transformation block */ 3061a717540cSStefano Zampini PetscInt row,col; 3062a6b551f4SStefano Zampini PetscScalar val,norm; 3063a6b551f4SStefano Zampini 3064a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30659162d606SStefano 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)); 3066a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 30679162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 30689162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3069bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 30709162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3071906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 30729162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3073a717540cSStefano Zampini } else { 3074a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 30759162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3076a717540cSStefano Zampini if (row != col) { 30779162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3078a717540cSStefano Zampini } else { 30799162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3080a717540cSStefano Zampini } 3081906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3082a717540cSStefano Zampini } 3083a717540cSStefano Zampini } 3084a717540cSStefano Zampini } 308598a51de6SStefano Zampini if (pcbddc->dbg_flag) { 308622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3087a717540cSStefano Zampini } 3088674ae819SStefano Zampini } 3089984c4197SStefano Zampini } else { 3090984c4197SStefano Zampini if (pcbddc->dbg_flag) { 30919162d606SStefano 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); 3092674ae819SStefano Zampini } 3093674ae819SStefano Zampini } 3094674ae819SStefano Zampini } 3095a717540cSStefano Zampini 3096a717540cSStefano Zampini /* free workspace */ 3097a717540cSStefano Zampini if (qr_needed) { 3098984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3099cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3100984c4197SStefano Zampini } 3101984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3102984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3103984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3104984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3105984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3106674ae819SStefano Zampini } 3107a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3108906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3109906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3110906d46d4SStefano Zampini 3111906d46d4SStefano Zampini /* assembling of global change of variable */ 3112bbb9e6c6SStefano Zampini { 3113bbb9e6c6SStefano Zampini Mat tmat; 311416f15bc4SStefano Zampini PetscInt bs; 311516f15bc4SStefano Zampini 3116906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3117906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3118bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3119bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3120bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3121bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 312216f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 312316f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3124906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3125bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3126bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3127bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3128bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3129bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3130e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3131e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3132bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3133bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3134906d46d4SStefano Zampini } 3135906d46d4SStefano Zampini /* check */ 3136906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3137906d46d4SStefano Zampini PetscReal error; 3138906d46d4SStefano Zampini Vec x,x_change; 3139906d46d4SStefano Zampini 3140906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3141906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3142906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3143906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3144e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3145e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3146bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3147e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3148e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3149906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3150906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3151906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3152906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3153bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3154906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3155906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3156906d46d4SStefano Zampini } 3157b96c3477SStefano Zampini 3158b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3159b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3160b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3161b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3162ac632422SStefano Zampini Mat S_new,tmat; 3163*b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3164*b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3165*b087196eSStefano Zampini const PetscScalar *array; 3166*b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3167*b087196eSStefano Zampini PetscInt i,n_V; 3168bbb9e6c6SStefano Zampini 3169bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 31706816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3171*b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3172*b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3173*b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3174*b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3175bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3176*b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3177ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3178b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3179ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3180*b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3181*b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3182*b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3183*b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3184*b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3185*b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3186*b087196eSStefano Zampini PetscScalar val; 3187*b087196eSStefano Zampini PetscInt idx; 3188*b087196eSStefano Zampini 3189*b087196eSStefano Zampini idx = idxs_V[i]; 3190*b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3191*b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3192*b087196eSStefano Zampini } 3193*b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3194*b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3195ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3196ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3197ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3198ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3199b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3200ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3201*b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3202ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3203ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3204ac632422SStefano Zampini } 3205*b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3206*b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3207*b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3208b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3209*b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3210b96c3477SStefano Zampini } 3211b96c3477SStefano Zampini } 3212906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3213906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3214b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3215b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3216b9b85e73SStefano Zampini } 3217906d46d4SStefano Zampini 3218906d46d4SStefano Zampini /* set up change of basis context */ 3219906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3220906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3221906d46d4SStefano Zampini 3222906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3223906d46d4SStefano Zampini PetscInt global_size,local_size; 3224906d46d4SStefano Zampini 3225906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3226906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3227906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3228906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3229906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3230906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3231906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3232906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3233906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3234906d46d4SStefano Zampini } else { 3235906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3236906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3237906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3238906d46d4SStefano Zampini } 3239906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3240906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3241906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3242906d46d4SStefano Zampini } else { 3243906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3244906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3245906d46d4SStefano Zampini } 3246906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3247906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3248906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3249906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3250b9b85e73SStefano Zampini } 3251a717540cSStefano Zampini 3252019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3253c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 3254c9ed8603SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx; 3255c9ed8603SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx; 3256019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3257019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3258019a44ceSStefano Zampini pcbddc->local_primal_size++; 3259019a44ceSStefano Zampini } 3260019a44ceSStefano Zampini 3261019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3262727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3263727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3264aff50787SStefano 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); 3265c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 32660e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3267aff50787SStefano 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); 3268727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3269727cdba6SStefano Zampini } 32700e6343abSStefano Zampini } 32710e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3272727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3273727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3274727cdba6SStefano Zampini 3275a717540cSStefano Zampini /* flush dbg viewer */ 3276b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3277b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3278b8ffe317SStefano Zampini } 3279a717540cSStefano Zampini 3280e310c8b4SStefano Zampini /* free workspace */ 3281a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 32824641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 328308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 32849162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 32859162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 328608122e43SStefano Zampini } else { 32879162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 32889162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 32899162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 329008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 329108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 32929162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 32939162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 329408122e43SStefano Zampini } 3295674ae819SStefano Zampini PetscFunctionReturn(0); 3296674ae819SStefano Zampini } 3297674ae819SStefano Zampini 3298674ae819SStefano Zampini #undef __FUNCT__ 3299674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3300674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3301674ae819SStefano Zampini { 3302674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3303674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3304674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 33057fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3306674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3307674ae819SStefano Zampini 3308674ae819SStefano Zampini PetscFunctionBegin; 33098e61c736SStefano Zampini /* Reset previously computed graph */ 33108e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3311674ae819SStefano Zampini /* Init local Graph struct */ 33127fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 33133bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3314674ae819SStefano Zampini 3315575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3316575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3317575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3318575ad6abSStefano Zampini } 33199577ea80SStefano Zampini 3320674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 33214d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 33224d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 33234d379d7bSStefano Zampini PetscInt nvtxs; 3324e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3325674ae819SStefano Zampini 33264d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 33272fffb893SStefano Zampini 33282fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 33292fffb893SStefano Zampini if (flg_row) { 33304d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3331b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 33322fffb893SStefano Zampini } 33332fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 33349b28b941SStefano 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 */ 33354d379d7bSStefano Zampini IS is_dummy; 33364d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 33374d379d7bSStefano Zampini PetscInt j,sum; 33384d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 33394d379d7bSStefano Zampini const PetscInt *idxs; 33404d379d7bSStefano Zampini PCBDDCGraph graph; 33414d379d7bSStefano Zampini PetscBT is_on_boundary; 33424d379d7bSStefano Zampini 33434d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 33444d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 33454d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 33464d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 33477fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 33484d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3349e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3350e496cd5dSStefano Zampini if (flg_row) { 33514d379d7bSStefano Zampini graph->xadj = xadj; 33524d379d7bSStefano Zampini graph->adjncy = adjncy; 3353e496cd5dSStefano Zampini } 33544d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 33554d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3356e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 33574d379d7bSStefano Zampini 33584d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 33599b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 33604d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33614d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 33624d379d7bSStefano Zampini } 33634d379d7bSStefano Zampini } 33644d379d7bSStefano Zampini 3365e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 33664d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 33674d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 33684d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 33694d379d7bSStefano Zampini } 33704d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 33714d379d7bSStefano Zampini 3372e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 33734d379d7bSStefano Zampini sum = 0; 33744d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33754d379d7bSStefano Zampini PetscInt sizecc = 0; 33764d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33774d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33784d379d7bSStefano Zampini sizecc++; 33794d379d7bSStefano Zampini } 33804d379d7bSStefano Zampini } 33814d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33824d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33834d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 33844d379d7bSStefano Zampini } 33854d379d7bSStefano Zampini } 33864d379d7bSStefano Zampini sum += sizecc*sizecc; 33874d379d7bSStefano Zampini } 33884d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 33894d379d7bSStefano Zampini sum = 0; 3390e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 33914d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 33924d379d7bSStefano Zampini cxadj[i] = sum; 33934d379d7bSStefano Zampini sum += temp; 33944d379d7bSStefano Zampini } 3395e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 33964d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33974d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33984d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33994d379d7bSStefano Zampini PetscInt k,sizecc = 0; 34004d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 34014d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 34024d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 34034d379d7bSStefano Zampini sizecc++; 34044d379d7bSStefano Zampini } 34054d379d7bSStefano Zampini } 34064d379d7bSStefano Zampini } 34074d379d7bSStefano Zampini } 34084d379d7bSStefano Zampini } 34099b28b941SStefano Zampini if (sum) { 3410e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 34114d379d7bSStefano Zampini } else { 34124d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 34134d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 34144d379d7bSStefano Zampini } 34154d379d7bSStefano Zampini graph->xadj = 0; 34164d379d7bSStefano Zampini graph->adjncy = 0; 34174d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 34184d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 34194d379d7bSStefano Zampini } 3420674ae819SStefano Zampini } 34219b28b941SStefano Zampini if (pcbddc->dbg_flag) { 34229b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3423674ae819SStefano Zampini } 3424674ae819SStefano Zampini 342563602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3426674ae819SStefano Zampini vertex_size = 1; 342763602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 342863602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 342995ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 343063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3431e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 343263602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3433674ae819SStefano Zampini } 343463602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 343563602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 343663602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3437674ae819SStefano Zampini } 343863602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3439674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 344063602bcaSStefano Zampini } else { 344163602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 344263602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3443854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 344463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 344563602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 344663602bcaSStefano Zampini } 344763602bcaSStefano Zampini } 3448674ae819SStefano Zampini } 3449674ae819SStefano Zampini 3450674ae819SStefano Zampini /* Setup of Graph */ 3451785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3452e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3453785d1243SStefano Zampini } 3454785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3455e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3456785d1243SStefano Zampini } 3457302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3458674ae819SStefano Zampini 3459674ae819SStefano Zampini /* Graph's connected components analysis */ 3460674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3461674ae819SStefano Zampini 3462674ae819SStefano Zampini /* print some info to stdout */ 3463674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3464302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3465674ae819SStefano Zampini } 3466fb180af4SStefano Zampini 3467fb180af4SStefano Zampini /* mark topography has done */ 3468fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3469674ae819SStefano Zampini PetscFunctionReturn(0); 3470674ae819SStefano Zampini } 3471674ae819SStefano Zampini 3472dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3473674ae819SStefano Zampini #undef __FUNCT__ 3474674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3475dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3476674ae819SStefano Zampini { 3477dc456d91SStefano Zampini PetscSF sf; 3478dc456d91SStefano Zampini PetscLayout map; 3479dc456d91SStefano Zampini const PetscInt *idxs; 3480dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3481dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3482dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3483dc456d91SStefano Zampini PetscMPIInt commsize; 3484674ae819SStefano Zampini PetscBool first_found; 3485674ae819SStefano Zampini PetscErrorCode ierr; 3486674ae819SStefano Zampini 3487674ae819SStefano Zampini PetscFunctionBegin; 3488dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3489dc456d91SStefano Zampini if (subset_mult) { 3490dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3491dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3492dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3493674ae819SStefano Zampini } 3494dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3495dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3496dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3497dc456d91SStefano Zampini for (i=0;i<n;i++) { 3498dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3499dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3500674ae819SStefano Zampini } 3501dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3502dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3503dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3504dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3505dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3506dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3507dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3508dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3509dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3510dc456d91SStefano Zampini 3511dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3512dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3513dc456d91SStefano Zampini if (subset_mult) { 3514dc456d91SStefano Zampini const PetscInt* idxs_mult; 3515dc456d91SStefano Zampini 3516dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3517dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3518dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3519674ae819SStefano Zampini } else { 3520dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3521674ae819SStefano Zampini } 3522dc456d91SStefano Zampini /* local size of new subset */ 3523dc456d91SStefano Zampini n_n = 0; 3524dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3525dc456d91SStefano Zampini 3526dc456d91SStefano Zampini /* global indexes in layout */ 3527dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3528dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3529dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3530dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3531dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3532dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3533dc456d91SStefano Zampini 3534dc456d91SStefano Zampini /* reduce from leaves to roots */ 3535dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 353664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 353764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3538dc456d91SStefano Zampini 3539dc456d91SStefano Zampini /* count indexes in local part of layout */ 3540674ae819SStefano Zampini nlocals = 0; 3541674ae819SStefano Zampini first_index = -1; 3542674ae819SStefano Zampini first_found = PETSC_FALSE; 3543dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3544dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3545674ae819SStefano Zampini first_found = PETSC_TRUE; 3546674ae819SStefano Zampini first_index = i; 3547674ae819SStefano Zampini } 3548dc456d91SStefano Zampini nlocals += root_data[i]; 3549674ae819SStefano Zampini } 3550dc456d91SStefano Zampini 3551dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 35525fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3553dc456d91SStefano Zampini start = 0; 355464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 35555fa240b1SStefano Zampini #else 355664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 35575fa240b1SStefano Zampini start = start-nlocals; 35585fa240b1SStefano Zampini #endif 35595fa240b1SStefano Zampini 3560dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3561dc456d91SStefano Zampini *N_n = start + nlocals; 3562dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3563dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3564674ae819SStefano Zampini } 35655fa240b1SStefano Zampini 35665fa240b1SStefano Zampini /* adapt root data with cumulative */ 3567674ae819SStefano Zampini if (first_found) { 3568dc456d91SStefano Zampini PetscInt old_index; 3569dc456d91SStefano Zampini 3570dc456d91SStefano Zampini root_data[first_index] += start; 3571674ae819SStefano Zampini old_index = first_index; 3572dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3573dc456d91SStefano Zampini if (root_data[i]) { 3574dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3575674ae819SStefano Zampini old_index = i; 3576674ae819SStefano Zampini } 3577674ae819SStefano Zampini } 3578674ae819SStefano Zampini } 3579dc456d91SStefano Zampini 3580dc456d91SStefano Zampini /* from roots to leaves */ 3581dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3582dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3583dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3584dc456d91SStefano Zampini 3585dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3586dc456d91SStefano Zampini if (subset_mult) { 3587dc456d91SStefano Zampini const PetscInt* idxs_mult; 3588dc456d91SStefano Zampini PetscInt cum; 3589dc456d91SStefano Zampini 3590dc456d91SStefano Zampini cum = 0; 3591dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3592dc456d91SStefano Zampini for (i=0;i<n;i++) { 3593dc456d91SStefano Zampini PetscInt j; 3594dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3595674ae819SStefano Zampini } 3596dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3597674ae819SStefano Zampini } else { 3598dc456d91SStefano Zampini for (i=0;i<n;i++) { 3599dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3600674ae819SStefano Zampini } 3601674ae819SStefano Zampini } 3602dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3603dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3604674ae819SStefano Zampini PetscFunctionReturn(0); 3605674ae819SStefano Zampini } 36069a7d3425SStefano Zampini 36079a7d3425SStefano Zampini #undef __FUNCT__ 36089a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 36099a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 36109a7d3425SStefano Zampini { 36119a7d3425SStefano Zampini PetscInt i,j; 36129a7d3425SStefano Zampini PetscScalar *alphas; 36139a7d3425SStefano Zampini PetscErrorCode ierr; 36149a7d3425SStefano Zampini 36159a7d3425SStefano Zampini PetscFunctionBegin; 36169a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3617785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 36189a7d3425SStefano Zampini for (i=0;i<n;i++) { 36199a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 36209a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 36219a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 36229a7d3425SStefano Zampini } 36239a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 36249a7d3425SStefano Zampini PetscFunctionReturn(0); 36259a7d3425SStefano Zampini } 36269a7d3425SStefano Zampini 3627e7931f94SStefano Zampini #undef __FUNCT__ 362870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3629b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3630e7931f94SStefano Zampini { 363152e5ac9dSStefano Zampini IS ranks_send_to; 3632e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3633e7931f94SStefano Zampini PetscMPIInt size,rank,color; 363452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 363552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 36363837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 36372b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3638e7931f94SStefano Zampini PetscSubcomm subcomm; 363952e5ac9dSStefano Zampini PetscErrorCode ierr; 3640a57a6d2fSStefano Zampini 3641e7931f94SStefano Zampini PetscFunctionBegin; 36422b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 36432b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 36442b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3645e7931f94SStefano Zampini 3646e7931f94SStefano Zampini /* Get info on mapping */ 36473bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 36483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3649e7931f94SStefano Zampini 3650e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3651785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3652e7931f94SStefano Zampini xadj[0] = 0; 3653e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3654785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3655785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3656e7931f94SStefano Zampini 36572b510759SStefano Zampini if (threshold) { 3658d023bfaeSStefano Zampini PetscInt xadj_count = 0; 36592b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3660d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3661d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3662d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3663d023bfaeSStefano Zampini xadj_count++; 3664e7931f94SStefano Zampini } 3665e7931f94SStefano Zampini } 3666d023bfaeSStefano Zampini xadj[1] = xadj_count; 3667c8587f34SStefano Zampini } else { 3668e7931f94SStefano Zampini if (xadj[1]) { 3669e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3670e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3671c8587f34SStefano Zampini } 3672e7931f94SStefano Zampini } 36733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3674e7931f94SStefano Zampini if (use_square) { 3675e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3676e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3677e7931f94SStefano Zampini } 3678e7931f94SStefano Zampini } 3679e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3680e7931f94SStefano Zampini 36813837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3682e7931f94SStefano Zampini 3683e7931f94SStefano Zampini /* 3684e7931f94SStefano Zampini Restrict work on active processes only. 3685e7931f94SStefano Zampini */ 3686e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3687e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3688e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 36892b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3690d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3691e7931f94SStefano Zampini if (color) { 3692e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3693e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3694e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3695c8587f34SStefano Zampini } else { 369652e5ac9dSStefano Zampini Mat subdomain_adj; 369752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 369852e5ac9dSStefano Zampini MatPartitioning partitioner; 369952e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 370052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3701b0c7d250SStefano Zampini PetscBool aggregate; 3702b0c7d250SStefano Zampini 3703306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3704785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3705e7931f94SStefano Zampini prank = rank; 3706306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 37078002ef2cSStefano Zampini /* 3708e7931f94SStefano Zampini for (i=0;i<size;i++) { 3709e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3710c8587f34SStefano Zampini } 37118002ef2cSStefano Zampini */ 3712e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3713e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3714c8587f34SStefano Zampini } 3715e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3716b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3717b0c7d250SStefano Zampini if (aggregate) { 3718b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3719b0c7d250SStefano Zampini PetscMPIInt nrank; 3720b0c7d250SStefano Zampini PetscScalar *vals; 3721b0c7d250SStefano Zampini 3722b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3723b0c7d250SStefano Zampini lrows = 0; 3724b0c7d250SStefano Zampini if (nrank<redprocs) { 3725b0c7d250SStefano Zampini lrows = size/redprocs; 3726b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3727b0c7d250SStefano Zampini } 37285fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3729b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3730b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3731b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3732b0c7d250SStefano Zampini row = nrank; 3733b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3734b0c7d250SStefano Zampini cols = adjncy; 3735b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3736b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3737b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3738b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3739b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 374052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 374152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 374252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3743b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3744b0c7d250SStefano Zampini } else { 3745306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3746b0c7d250SStefano Zampini } 374722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3748e7931f94SStefano Zampini 3749e7931f94SStefano Zampini /* Partition */ 3750306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3751e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3752e7931f94SStefano Zampini if (use_vwgt) { 37533837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3754e7931f94SStefano Zampini v_wgt[0] = local_size; 3755e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3756c8587f34SStefano Zampini } 375728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 375828143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3759e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3760e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 376122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3762e7931f94SStefano Zampini 376352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 376452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 376552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 376652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3767b0c7d250SStefano Zampini if (!redprocs) { 3768b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 376928143c3dSStefano Zampini } else { 3770b0c7d250SStefano Zampini PetscInt idxs[1]; 3771b0c7d250SStefano Zampini PetscMPIInt tag; 3772b0c7d250SStefano Zampini MPI_Request *reqs; 3773b0c7d250SStefano Zampini 3774b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3775b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3776b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3777b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 377828143c3dSStefano Zampini } 3779b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3780b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3781b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3782b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3783e7931f94SStefano Zampini } 378452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3785e7931f94SStefano Zampini /* clean up */ 3786e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 378752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3788e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3789e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3790e7931f94SStefano Zampini } 3791e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3792e7931f94SStefano Zampini 3793e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3794e7931f94SStefano Zampini i = 1; 3795e7931f94SStefano Zampini if (color) i=0; 3796e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3797e7931f94SStefano Zampini /* get back IS */ 3798e7931f94SStefano Zampini *is_sends = ranks_send_to; 3799e7931f94SStefano Zampini PetscFunctionReturn(0); 3800e7931f94SStefano Zampini } 3801e7931f94SStefano Zampini 3802e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3803e7931f94SStefano Zampini 3804e7931f94SStefano Zampini #undef __FUNCT__ 3805e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 380653a05cb3SStefano 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[]) 3807e7931f94SStefano Zampini { 380870cf5478SStefano Zampini Mat local_mat; 3809e7931f94SStefano Zampini IS is_sends_internal; 38109d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 381128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 38129d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3813e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3814e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3815e7931f94SStefano Zampini const PetscInt* is_indices; 3816e7931f94SStefano Zampini MatType new_local_type; 3817e7931f94SStefano Zampini /* buffers */ 3818e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 381928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 38209d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3821e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3822e7931f94SStefano Zampini /* MPI */ 382328143c3dSStefano Zampini MPI_Comm comm,comm_n; 382428143c3dSStefano Zampini PetscSubcomm subcomm; 3825e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 382628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 382728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 382828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 382928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 383028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3831e7931f94SStefano Zampini PetscErrorCode ierr; 3832e7931f94SStefano Zampini 3833e7931f94SStefano Zampini PetscFunctionBegin; 383428143c3dSStefano Zampini /* TODO: add missing checks */ 383528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 383628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 383728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 383828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3839e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 384028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3841e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3842e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3843e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3844e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3845e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 384628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 384770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 384870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 384928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 385070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 385170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 385270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 385370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 385470cf5478SStefano Zampini } 3855e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3856e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3857e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3858e7931f94SStefano Zampini if (!is_sends) { 385928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3860b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3861c8587f34SStefano Zampini } else { 3862e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3863e7931f94SStefano Zampini is_sends_internal = is_sends; 3864c8587f34SStefano Zampini } 3865e7931f94SStefano Zampini 3866e7931f94SStefano Zampini /* get comm */ 3867a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3868e7931f94SStefano Zampini 3869e7931f94SStefano Zampini /* compute number of sends */ 3870e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3871e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3872e7931f94SStefano Zampini 3873e7931f94SStefano Zampini /* compute number of receives */ 3874e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3875785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3876e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3877e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3878e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3879e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3880e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3881e7931f94SStefano Zampini 388228143c3dSStefano Zampini /* restrict comm if requested */ 388328143c3dSStefano Zampini subcomm = 0; 388428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 388528143c3dSStefano Zampini if (restrict_comm) { 3886779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3887779c1cceSStefano Zampini 388828143c3dSStefano Zampini color = 0; 388953a05cb3SStefano Zampini if (restrict_full) { 389053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 389153a05cb3SStefano Zampini } else { 389253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 389353a05cb3SStefano Zampini } 389428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 389528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 389628143c3dSStefano Zampini /* check if reuse has been requested */ 389728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 389828143c3dSStefano Zampini if (*mat_n) { 389928143c3dSStefano Zampini PetscMPIInt subcommsize2; 390028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 390128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 390228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 390328143c3dSStefano Zampini } else { 390428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 390528143c3dSStefano Zampini } 390628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3907779c1cceSStefano Zampini PetscMPIInt rank; 3908779c1cceSStefano Zampini 3909779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 391028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 391128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 391228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3913306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 391428143c3dSStefano Zampini } 391528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 391628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 391728143c3dSStefano Zampini } else { 391828143c3dSStefano Zampini comm_n = comm; 391928143c3dSStefano Zampini } 392028143c3dSStefano Zampini 3921e7931f94SStefano Zampini /* prepare send/receive buffers */ 3922785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3923e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3924785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3925e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 392628143c3dSStefano Zampini if (nis) { 3927854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 392828143c3dSStefano Zampini } 3929e7931f94SStefano Zampini 393028143c3dSStefano Zampini /* Get data from local matrices */ 3931e7931f94SStefano Zampini if (!isdense) { 3932a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3933e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3934e7931f94SStefano Zampini /* 3935e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3936e7931f94SStefano Zampini send_buffer_idxs should contain: 3937e7931f94SStefano Zampini - MatType_PRIVATE type 3938e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3939e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3940e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3941e7931f94SStefano Zampini */ 3942e7931f94SStefano Zampini } else { 3943e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 39443bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3945854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3946e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3947e7931f94SStefano Zampini send_buffer_idxs[1] = i; 39483bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3949e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 39503bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3951e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3952e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3953e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3954e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3955c8587f34SStefano Zampini } 3956c8587f34SStefano Zampini } 3957e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 395828143c3dSStefano Zampini /* additional is (if any) */ 395928143c3dSStefano Zampini if (nis) { 396028143c3dSStefano Zampini PetscMPIInt psum; 396128143c3dSStefano Zampini PetscInt j; 396228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 396328143c3dSStefano Zampini PetscInt plen; 396428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 396528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 396628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 396728143c3dSStefano Zampini } 3968854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 396928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 397028143c3dSStefano Zampini PetscInt plen; 397128143c3dSStefano Zampini const PetscInt *is_array_idxs; 397228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 397328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 397428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 397528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 397628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 397728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 397828143c3dSStefano Zampini } 397928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 398028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 398128143c3dSStefano Zampini } 398228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 398328143c3dSStefano Zampini } 398428143c3dSStefano Zampini 3985e7931f94SStefano Zampini buf_size_idxs = 0; 3986e7931f94SStefano Zampini buf_size_vals = 0; 398728143c3dSStefano Zampini buf_size_idxs_is = 0; 3988e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3989e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3990e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 399128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3992e7931f94SStefano Zampini } 3993785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3994785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 399595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3996e7931f94SStefano Zampini 3997e7931f94SStefano Zampini /* get new tags for clean communications */ 3998e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3999e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 400028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4001e7931f94SStefano Zampini 4002e7931f94SStefano Zampini /* allocate for requests */ 4003785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4004785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 400595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4006785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4007785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 400895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4009e7931f94SStefano Zampini 4010e7931f94SStefano Zampini /* communications */ 4011e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4012e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 401328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4014e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4015e7931f94SStefano Zampini source_dest = onodes[i]; 4016e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4017e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4018e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4019e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 402028143c3dSStefano Zampini if (nis) { 402128143c3dSStefano 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); 402228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 402328143c3dSStefano Zampini } 4024e7931f94SStefano Zampini } 4025e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4026e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4027e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4028e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 402928143c3dSStefano Zampini if (nis) { 403028143c3dSStefano 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); 403128143c3dSStefano Zampini } 4032e7931f94SStefano Zampini } 4033e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4034e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4035e7931f94SStefano Zampini 4036e7931f94SStefano Zampini /* assemble new l2g map */ 4037e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4038e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 40399d30be91SStefano Zampini new_local_rows = 0; 4040e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 40419d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4042e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4043e7931f94SStefano Zampini } 40449d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4045e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 40469d30be91SStefano Zampini new_local_rows = 0; 4047e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 40489d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 40499d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4050e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4051e7931f94SStefano Zampini } 40529d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 40539d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4054e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4055e7931f94SStefano Zampini 4056e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4057e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4058e7931f94SStefano 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) */ 4059e7931f94SStefano Zampini if (n_recvs) { 406028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4061e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4062e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4063e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4064e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4065e7931f94SStefano Zampini break; 4066e7931f94SStefano Zampini } 4067e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4068e7931f94SStefano Zampini } 4069e7931f94SStefano Zampini switch (new_local_type_private) { 407028143c3dSStefano Zampini case MATDENSE_PRIVATE: 407128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4072e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4073e7931f94SStefano Zampini bs = 1; 407428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 407528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 407628143c3dSStefano Zampini bs = 1; 407728143c3dSStefano Zampini } 4078e7931f94SStefano Zampini break; 4079e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4080e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4081e7931f94SStefano Zampini bs = 1; 4082e7931f94SStefano Zampini break; 4083e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4084e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4085e7931f94SStefano Zampini break; 4086e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4087e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4088e7931f94SStefano Zampini break; 4089e7931f94SStefano Zampini default: 40909d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4091e7931f94SStefano Zampini break; 4092e7931f94SStefano Zampini } 409328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 409428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 409528143c3dSStefano Zampini bs = 1; 4096e7931f94SStefano Zampini } 4097e7931f94SStefano Zampini 409870cf5478SStefano Zampini /* create MATIS object if needed */ 409970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4100e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4101e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 410270cf5478SStefano Zampini } else { 410370cf5478SStefano Zampini /* it also destroys the local matrices */ 410470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 410570cf5478SStefano Zampini } 410670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4107e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 41089d30be91SStefano Zampini 41099d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 41109d30be91SStefano Zampini 41119d30be91SStefano Zampini /* Global to local map of received indices */ 41129d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 41139d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 41149d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 41159d30be91SStefano Zampini 41169d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 41179d30be91SStefano Zampini buf_size_idxs = 0; 41189d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 41199d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 41209d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 41219d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 41229d30be91SStefano Zampini } 41239d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 41249d30be91SStefano Zampini 41259d30be91SStefano Zampini /* set preallocation */ 41269d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 41279d30be91SStefano Zampini if (!newisdense) { 41289d30be91SStefano Zampini PetscInt *new_local_nnz=0; 41299d30be91SStefano Zampini 41309d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 41319d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 41329d30be91SStefano Zampini if (n_recvs) { 41339d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 41349d30be91SStefano Zampini } 41359d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 41369d30be91SStefano Zampini PetscInt j; 41379d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 41389d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 41399d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 41409d30be91SStefano Zampini } 41419d30be91SStefano Zampini } else { 41429d30be91SStefano Zampini /* TODO */ 41439d30be91SStefano Zampini } 41449d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 41459d30be91SStefano Zampini } 41469d30be91SStefano Zampini if (new_local_nnz) { 41479d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 41489d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 41499d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 41509d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 41519d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 41529d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 41539d30be91SStefano Zampini } else { 41549d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 41559d30be91SStefano Zampini } 41569d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 41579d30be91SStefano Zampini } else { 41589d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 41599d30be91SStefano Zampini } 4160e7931f94SStefano Zampini 4161e7931f94SStefano Zampini /* set values */ 4162e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 41639d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4164e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4165e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4166e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 41679d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4168e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4169e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4170e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 417128143c3dSStefano Zampini } else { 417228143c3dSStefano Zampini /* TODO */ 4173e7931f94SStefano Zampini } 4174e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4175e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4176e7931f94SStefano Zampini } 4177e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4178e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 417970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 418070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41819d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 41829d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4183e7931f94SStefano Zampini 4184dfd14d43SStefano Zampini #if 0 418528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4186e7931f94SStefano Zampini Vec lvec,rvec; 4187e7931f94SStefano Zampini PetscReal infty_error; 4188e7931f94SStefano Zampini 41892a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4190e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4191e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4192e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 419370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4194e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4195e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4196e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4197e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4198e7931f94SStefano Zampini } 419928143c3dSStefano Zampini #endif 4200e7931f94SStefano Zampini 420128143c3dSStefano Zampini /* assemble new additional is (if any) */ 420228143c3dSStefano Zampini if (nis) { 420328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 420428143c3dSStefano Zampini 420528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4206854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 420728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 420828143c3dSStefano Zampini psum = 0; 420928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 421028143c3dSStefano Zampini for (j=0;j<nis;j++) { 421128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 421228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 421328143c3dSStefano Zampini psum += plen; 421428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 421528143c3dSStefano Zampini } 421628143c3dSStefano Zampini } 4217854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4218854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 421928143c3dSStefano Zampini for (i=1;i<nis;i++) { 422028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 422128143c3dSStefano Zampini } 422228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 422328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 422428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 422528143c3dSStefano Zampini for (j=0;j<nis;j++) { 422628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 422728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 422828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 422928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 423028143c3dSStefano Zampini } 423128143c3dSStefano Zampini } 423228143c3dSStefano Zampini for (i=0;i<nis;i++) { 423328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 423428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 423528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 423628143c3dSStefano Zampini } 423728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 423828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 423928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 424028143c3dSStefano Zampini } 4241e7931f94SStefano Zampini /* free workspace */ 424228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4243e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4244e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4245e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4246e7931f94SStefano Zampini if (isdense) { 4247e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4248e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4249e7931f94SStefano Zampini } else { 4250e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4251e7931f94SStefano Zampini } 425228143c3dSStefano Zampini if (nis) { 425328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 425428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 425528143c3dSStefano Zampini } 4256e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4257e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 425828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4259e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4260e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 426128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4262e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4263e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4264e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4265e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4266e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 426728143c3dSStefano Zampini if (nis) { 426828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 426928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 427028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 427128143c3dSStefano Zampini } 427228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 427328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 427428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 427528143c3dSStefano Zampini for (i=0;i<nis;i++) { 427628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 427728143c3dSStefano Zampini } 427853a05cb3SStefano Zampini *mat_n = NULL; 427928143c3dSStefano Zampini } 4280e7931f94SStefano Zampini PetscFunctionReturn(0); 4281e7931f94SStefano Zampini } 4282a57a6d2fSStefano Zampini 428312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4284af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 428512edc857SStefano Zampini 4286c8587f34SStefano Zampini #undef __FUNCT__ 4287c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4288c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4289c8587f34SStefano Zampini { 4290c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4291c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 429220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 42939881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 429420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 42956e683305SStefano Zampini IS coarse_is,*isarray; 42966e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 42976e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4298f9eb5b7dSStefano Zampini PC pc_temp; 4299c8587f34SStefano Zampini PCType coarse_pc_type; 4300c8587f34SStefano Zampini KSPType coarse_ksp_type; 4301f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 43024f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 43036e683305SStefano Zampini Mat t_coarse_mat_is; 43046e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 43056e683305SStefano Zampini PetscMPIInt all_procs; 430674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 430768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 430822bc73bbSStefano Zampini PetscScalar *array; 43099881197aSStefano Zampini PetscErrorCode ierr; 4310fdc09c96SStefano Zampini 4311c8587f34SStefano Zampini PetscFunctionBegin; 4312c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 431368457ee5SStefano 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 */ 4314fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 43155a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4316fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4317f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4318f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4319f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4320fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 432151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 432251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4323dc4bcba2SStefano Zampini PC pc; 4324dc4bcba2SStefano Zampini PetscBool isbddc; 4325dc4bcba2SStefano Zampini 4326dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4327dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4328dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4329dc4bcba2SStefano Zampini if (isbddc) { 4330dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4331dc4bcba2SStefano Zampini } 4332727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4333fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4334fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4335fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4336f4ddd8eeSStefano Zampini } 4337fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4338fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4339f4ddd8eeSStefano Zampini } 434070cf5478SStefano Zampini /* reset any subassembling information */ 434170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 43426e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 43436e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4344fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4345f4ddd8eeSStefano Zampini } 4346c8587f34SStefano Zampini 43476e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 43482b510759SStefano Zampini im_active = !!(pcis->n); 43492b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43506e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 43516e683305SStefano Zampini void_procs = all_procs-active_procs; 43526e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 435374e2c79eSStefano Zampini redist = PETSC_FALSE; 435422bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 43556e683305SStefano Zampini csin_ml = PETSC_TRUE; 43566e683305SStefano Zampini ncoarse_ml = void_procs; 4357779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4358779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 43596e683305SStefano Zampini csin_ds = PETSC_TRUE; 436018a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 436118a45a71SStefano Zampini redist = PETSC_TRUE; 436218a45a71SStefano Zampini } else { 43636e683305SStefano Zampini csin_ds = PETSC_TRUE; 4364779c1cceSStefano Zampini ncoarse_ds = active_procs; 4365779c1cceSStefano Zampini redist = PETSC_TRUE; 436618a45a71SStefano Zampini } 43676e683305SStefano Zampini } else { 43686e683305SStefano Zampini csin_ml = PETSC_FALSE; 43696e683305SStefano Zampini ncoarse_ml = all_procs; 43706e683305SStefano Zampini if (void_procs) { 43716e683305SStefano Zampini csin_ds = PETSC_TRUE; 43726e683305SStefano Zampini ncoarse_ds = void_procs; 43736e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 43746e683305SStefano Zampini } else { 4375779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 437674e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 437774e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 437874e2c79eSStefano Zampini redist = PETSC_TRUE; 437974e2c79eSStefano Zampini } else { 43806e683305SStefano Zampini csin_ds = PETSC_FALSE; 43816e683305SStefano Zampini ncoarse_ds = all_procs; 43826e683305SStefano Zampini } 43836e683305SStefano Zampini } 438474e2c79eSStefano Zampini } 43856e683305SStefano Zampini 43866e683305SStefano Zampini /* 43876e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 43886e683305SStefano Zampini - we have not exceeded the number of levels requested 43896e683305SStefano Zampini - we can actually subassemble the active processes 43906e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 43916e683305SStefano Zampini */ 43926e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 43936e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 43946e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 43956e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 43966e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4397f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 43982b510759SStefano Zampini } else { 4399f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4400c8587f34SStefano Zampini } 4401c8587f34SStefano Zampini } 44026e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 44036e683305SStefano Zampini if (multilevel_allowed) { 44046e683305SStefano Zampini ncoarse = ncoarse_ml; 44056e683305SStefano Zampini csin = csin_ml; 440658da7f69SStefano Zampini redist = PETSC_FALSE; 44076e683305SStefano Zampini } else { 44086e683305SStefano Zampini ncoarse = ncoarse_ds; 44096e683305SStefano Zampini csin = csin_ds; 44106e683305SStefano Zampini } 4411e7931f94SStefano Zampini 4412abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4413abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4414abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4415abbbba34SStefano Zampini 4416abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 441722bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 441822bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 441922bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 442022bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4421b1ecc7b1SStefano Zampini #if 0 4422b9b85e73SStefano Zampini { 4423b9b85e73SStefano Zampini PetscViewer viewer; 4424b9b85e73SStefano Zampini char filename[256]; 4425b1ecc7b1SStefano Zampini sprintf(filename,"local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4426b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4427b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4428b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4429b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4430b9b85e73SStefano Zampini } 4431b9b85e73SStefano Zampini #endif 4432e176bc59SStefano 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); 44336e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 44346e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 44356e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4436abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4437abbbba34SStefano Zampini 44386e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 44396e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 44406e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 44416e683305SStefano Zampini const PetscInt *idxs; 44426e683305SStefano Zampini ISLocalToGlobalMapping tmap; 44436e683305SStefano Zampini 44446e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 44450be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 44466e683305SStefano Zampini /* allocate space for temporary storage */ 4447854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4448854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 44496e683305SStefano Zampini /* allocate for IS array */ 44506e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 44516e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 44526e683305SStefano Zampini nis = nisdofs + nisneu; 4453854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 44546e683305SStefano Zampini /* dofs splitting */ 44556e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 44566e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 44576e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 44586e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 44596e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 44606e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 44616e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 44626e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 44636e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 44646e683305SStefano Zampini } 44656e683305SStefano Zampini /* neumann boundaries */ 44666e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 44676e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 44686e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 44696e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 44706e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 44716e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 44726e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 44736e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 44746e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 44756e683305SStefano Zampini } 44766e683305SStefano Zampini /* free memory */ 44776e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 44786e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 44796e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 44806e683305SStefano Zampini } else { 44816e683305SStefano Zampini nis = 0; 44826e683305SStefano Zampini nisdofs = 0; 44836e683305SStefano Zampini nisneu = 0; 44846e683305SStefano Zampini isarray = NULL; 44856e683305SStefano Zampini } 44866e683305SStefano Zampini /* destroy no longer needed map */ 44876e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 44886e683305SStefano Zampini 44896e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 44906e683305SStefano Zampini coarse_mat_is = NULL; 44916e683305SStefano Zampini if (csin) { 44926e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 449374e2c79eSStefano Zampini if (redist) { 449474e2c79eSStefano Zampini PetscMPIInt rank; 4495779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 449674e2c79eSStefano Zampini 449774e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 449858da7f69SStefano Zampini spc = active_procs/ncoarse; 449958da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4500779c1cceSStefano Zampini if (im_active) { 4501779c1cceSStefano Zampini destsize = 1; 450274e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 450374e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 450474e2c79eSStefano Zampini } else { 450574e2c79eSStefano Zampini dest[0] = rank/(spc+1); 450674e2c79eSStefano Zampini } 450774e2c79eSStefano Zampini } else { 4508779c1cceSStefano Zampini destsize = 0; 45096e683305SStefano Zampini } 4510779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4511779c1cceSStefano Zampini } else if (csin_type_simple) { 45126e683305SStefano Zampini PetscMPIInt rank; 45136e683305SStefano Zampini PetscInt issize,isidx; 4514779c1cceSStefano Zampini 45156e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 45166e683305SStefano Zampini if (im_active) { 45176e683305SStefano Zampini issize = 1; 45186e683305SStefano Zampini isidx = (PetscInt)rank; 45196e683305SStefano Zampini } else { 45206e683305SStefano Zampini issize = 0; 45216e683305SStefano Zampini isidx = -1; 45226e683305SStefano Zampini } 45236e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4524779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4525b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 45266e683305SStefano Zampini } 4527779c1cceSStefano Zampini 4528779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4529779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4530779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4531779c1cceSStefano Zampini PetscInt *coarse_candidates; 4532779c1cceSStefano Zampini const PetscInt* tisindices; 4533779c1cceSStefano Zampini 4534779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4535779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4536779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4537779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4538779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4539779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4540779c1cceSStefano Zampini } 4541779c1cceSStefano Zampini } 4542779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4543779c1cceSStefano Zampini 4544779c1cceSStefano Zampini 45456e683305SStefano Zampini if (pcbddc->dbg_flag) { 45466e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 45486e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 45496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4550779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 45516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 45526e683305SStefano Zampini } 45536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 45546e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45556e683305SStefano Zampini } 45566e683305SStefano Zampini /* shift the pattern on coarse candidates */ 45576e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 45586e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4559854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 45606e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 45616e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 45626e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 45636e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 45646e683305SStefano Zampini } 45656e683305SStefano Zampini if (pcbddc->dbg_flag) { 45666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 45686e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 45696e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45706e683305SStefano Zampini } 4571779c1cceSStefano Zampini } 45726e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 457353a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 457453a05cb3SStefano 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); 457553a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 457653a05cb3SStefano 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); 457753a05cb3SStefano Zampini } 45786e683305SStefano Zampini } else { 45796e683305SStefano Zampini if (pcbddc->dbg_flag) { 45806e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 45826e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45836e683305SStefano Zampini } 45846e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 45856e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 45866e683305SStefano Zampini } 45876e683305SStefano Zampini 45886e683305SStefano Zampini /* create local to global scatters for coarse problem */ 458968457ee5SStefano Zampini if (compute_vecs) { 45906e683305SStefano Zampini PetscInt lrows; 45916e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 45926e683305SStefano Zampini if (coarse_mat_is) { 45936e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 45946e683305SStefano Zampini } else { 45956e683305SStefano Zampini lrows = 0; 45966e683305SStefano Zampini } 45976e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 45986e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 45996e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 46006e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 46016e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 46026e683305SStefano Zampini } 46036e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 46046e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4605c8587f34SStefano Zampini 4606f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4607f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4608f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4609f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4610f9eb5b7dSStefano Zampini } else { 4611f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4612f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4613c8587f34SStefano Zampini } 4614c8587f34SStefano Zampini 46156e683305SStefano Zampini /* print some info if requested */ 46166e683305SStefano Zampini if (pcbddc->dbg_flag) { 46176e683305SStefano Zampini if (!multilevel_allowed) { 46186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46196e683305SStefano Zampini if (multilevel_requested) { 46206e683305SStefano 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); 46216e683305SStefano Zampini } else if (pcbddc->max_levels) { 46226e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 46236e683305SStefano Zampini } 46246e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 46256e683305SStefano Zampini } 46266e683305SStefano Zampini } 46276e683305SStefano Zampini 4628f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 46296e683305SStefano Zampini if (coarse_mat_is) { 46306e683305SStefano Zampini MatReuse coarse_mat_reuse; 46316a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 46326e683305SStefano Zampini if (pcbddc->dbg_flag) { 46336e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 46346e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 46356e683305SStefano Zampini } 4636f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4637312be037SStefano Zampini char prefix[256],str_level[16]; 4638e604994aSStefano Zampini size_t len; 46396e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4640422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4641c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4642f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 46435f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4644c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 46456e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4646c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4647c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4648e604994aSStefano Zampini /* prefix */ 4649e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4650e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4651e604994aSStefano Zampini if (!pcbddc->current_level) { 4652e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4653e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4654c8587f34SStefano Zampini } else { 4655e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4656312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4657312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 465834d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4659312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4660e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4661e604994aSStefano Zampini } 4662e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 46633e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 46643e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 46653e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 46663e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4667f9eb5b7dSStefano Zampini /* allow user customization */ 4668f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 46693e3c6dadSStefano Zampini } 46703e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 467151bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 46723e3c6dadSStefano Zampini if (nisdofs) { 46733e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 46743e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 46753e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 46763e3c6dadSStefano Zampini } 46773e3c6dadSStefano Zampini } 46783e3c6dadSStefano Zampini if (nisneu) { 46793e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 46803e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4681312be037SStefano Zampini } 4682f9eb5b7dSStefano Zampini 4683f9eb5b7dSStefano Zampini /* get some info after set from options */ 4684f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4685f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 46864f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 46876e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4688f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4689f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4690f9eb5b7dSStefano Zampini } 469139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 46924f3a063dSStefano Zampini if (isredundant) { 46934f3a063dSStefano Zampini KSP inner_ksp; 46944f3a063dSStefano Zampini PC inner_pc; 46954f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 46964f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 46974f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 46984f3a063dSStefano Zampini } 4699f9eb5b7dSStefano Zampini 4700f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4701fa7f1dd8SStefano Zampini if (coarse_reuse) { 470281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4703fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 47046e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4705fa7f1dd8SStefano Zampini } else { 47066e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4707fa7f1dd8SStefano Zampini } 4708c8587f34SStefano Zampini if (isbddc || isnn) { 470922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 471070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4711b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 471222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 47136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 47146e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 47156e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 47166e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 471722b6e8a2SStefano Zampini } 471870cf5478SStefano Zampini } 471953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 472070cf5478SStefano Zampini } else { 472122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 472222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 472322bc73bbSStefano Zampini } 472422bc73bbSStefano Zampini } else { 47252e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4726c8587f34SStefano Zampini } 4727c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4728c8587f34SStefano Zampini 47293301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 47305a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 47313301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 47323301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 47333301b35fSStefano Zampini } 47343301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 47353301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 47363301b35fSStefano Zampini } 47373301b35fSStefano Zampini if (pc->pmat->spd_set) { 47383301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 47393301b35fSStefano Zampini } 47406e683305SStefano Zampini /* set operators */ 47415f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 47426e683305SStefano Zampini if (pcbddc->dbg_flag) { 47436e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 47446e683305SStefano Zampini } 47456e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 47466e683305SStefano Zampini coarse_mat = 0; 47476e683305SStefano Zampini } 47486e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4749b1ecc7b1SStefano Zampini #if 0 4750b9b85e73SStefano Zampini { 4751b9b85e73SStefano Zampini PetscViewer viewer; 4752b9b85e73SStefano Zampini char filename[256]; 4753b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 4754b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 4755b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4756b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4757b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4758b9b85e73SStefano Zampini } 4759b9b85e73SStefano Zampini #endif 4760c8587f34SStefano Zampini 4761c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4762298c0119SStefano Zampini #if 0 4763c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4764c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 476598a51de6SStefano Zampini } 4766298c0119SStefano Zampini #endif 476798a51de6SStefano Zampini 476898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 476998a51de6SStefano Zampini Vec crhs,csol; 477098a51de6SStefano Zampini PetscBool ispreonly; 477104708bb6SStefano Zampini 477298a51de6SStefano Zampini if (CoarseNullSpace) { 4773c8587f34SStefano Zampini if (isbddc) { 4774c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4775c8587f34SStefano Zampini } else { 47765fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4777c8587f34SStefano Zampini } 4778c8587f34SStefano Zampini } 4779f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4780f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4781f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4782f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 47836e683305SStefano Zampini /* hack */ 4784f347579bSStefano Zampini if (!csol) { 47852a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4786f9eb5b7dSStefano Zampini } 4787f347579bSStefano Zampini if (!crhs) { 47882a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4789f347579bSStefano Zampini } 4790cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4791cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 47926e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4793c8587f34SStefano Zampini KSP check_ksp; 47942b510759SStefano Zampini KSPType check_ksp_type; 4795c8587f34SStefano Zampini PC check_pc; 47966e683305SStefano Zampini Vec check_vec,coarse_vec; 47976a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 47982b510759SStefano Zampini PetscInt its; 47996e683305SStefano Zampini PetscBool compute_eigs; 48006e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 48016e683305SStefano Zampini PetscInt neigs; 48028e185a42SStefano Zampini const char *prefix; 4803c8587f34SStefano Zampini 48042b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 48056e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4806422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 480723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4808f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 48092b510759SStefano Zampini if (ispreonly) { 48102b510759SStefano Zampini check_ksp_type = KSPPREONLY; 48116e683305SStefano Zampini compute_eigs = PETSC_FALSE; 48122b510759SStefano Zampini } else { 4813cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 48146e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4815c8587f34SStefano Zampini } 4816c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 48176e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 48186e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 48196e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4820a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4821a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4822a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4823a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4824c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4825c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4826c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4827c8587f34SStefano Zampini /* create random vec */ 48286e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 48296e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4830c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4831c8587f34SStefano Zampini if (CoarseNullSpace) { 4832c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4833c8587f34SStefano Zampini } 48346e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4835c8587f34SStefano Zampini /* solve coarse problem */ 48366e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4837c8587f34SStefano Zampini if (CoarseNullSpace) { 48386e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4839c8587f34SStefano Zampini } 4840cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 48416e683305SStefano Zampini if (compute_eigs) { 4842854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4843854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 48446e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 48456e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 48466e683305SStefano Zampini lambda_min = eigs_r[0]; 48476e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 48486e683305SStefano Zampini if (lambda_max>lambda_min) { 4849cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4850cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4851cbcc2c2aSStefano Zampini } 4852c8587f34SStefano Zampini } 4853c8587f34SStefano Zampini } 4854cbcc2c2aSStefano Zampini 4855c8587f34SStefano Zampini /* check coarse problem residual error */ 48566e683305SStefano Zampini if (pcbddc->dbg_flag) { 48576e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 48586e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48596e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4860c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 48616e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 48626e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4863c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4864779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 48656e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 48666e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 48676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 48686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 48696e683305SStefano Zampini if (compute_eigs) { 48706e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4871deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4872c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 48736e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 48746e683305SStefano 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); 48756e683305SStefano Zampini for (i=0;i<neigs;i++) { 48766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4877c8587f34SStefano Zampini } 48786e683305SStefano Zampini } 48796e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 48806e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48816e683305SStefano Zampini } 4882c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 48836e683305SStefano Zampini if (compute_eigs) { 48846e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 48856e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4886c8587f34SStefano Zampini } 48876e683305SStefano Zampini } 48886e683305SStefano Zampini } 4889cbcc2c2aSStefano Zampini /* print additional info */ 4890cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 48916e683305SStefano Zampini /* waits until all processes reaches this point */ 48926e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4893cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4894cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4895cbcc2c2aSStefano Zampini } 4896cbcc2c2aSStefano Zampini 48972b510759SStefano Zampini /* free memory */ 4898c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4899fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4900c8587f34SStefano Zampini PetscFunctionReturn(0); 4901c8587f34SStefano Zampini } 4902674ae819SStefano Zampini 4903f34684f1SStefano Zampini #undef __FUNCT__ 4904f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4905f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4906f34684f1SStefano Zampini { 4907f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4908f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4909f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4910dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4911dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 491273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4913dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4914f34684f1SStefano Zampini PetscErrorCode ierr; 4915f34684f1SStefano Zampini 4916f34684f1SStefano Zampini PetscFunctionBegin; 4917f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 49180e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 49190e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4920727cdba6SStefano Zampini } 4921dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 49223bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4923dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4924dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4925dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4926dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4927dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4928dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 49290e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 49300e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 49310e6343abSStefano Zampini } 4932dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4933dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4934dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4935dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4936dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4937f34684f1SStefano Zampini 4938f34684f1SStefano Zampini /* check numbering */ 4939f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4940019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 4941dc456d91SStefano Zampini PetscInt i; 4942b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4943f34684f1SStefano Zampini 4944f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4945f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4946f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 49470fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4948019a44ceSStefano Zampini /* counter */ 4949019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4950019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4951019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4952019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4953019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4954019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4955019a44ceSStefano Zampini 4956f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4957f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4958727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4959f34684f1SStefano Zampini } 4960f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4961f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4962f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4963e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4964e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4965e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4966e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4967f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4968019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4969f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4970019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 4971019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 4972019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 4973b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4974019a44ceSStefano 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); 4975f34684f1SStefano Zampini } 4976f34684f1SStefano Zampini } 4977019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4978b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4979f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4980f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4981f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4982f34684f1SStefano Zampini } 4983f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4984f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4985e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4986e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4987f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4988f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4989b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4990ca8b9ea9SStefano Zampini PetscInt *gidxs; 4991ca8b9ea9SStefano Zampini 4992ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 49933bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4994f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4995f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4996f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4997f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 49984bc2dc4bSStefano 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); 4999f34684f1SStefano Zampini } 5000f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5001ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5002f34684f1SStefano Zampini } 5003f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5004302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5005f34684f1SStefano Zampini } 50068bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5007f34684f1SStefano Zampini /* get back data */ 5008f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5009f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5010674ae819SStefano Zampini PetscFunctionReturn(0); 5011674ae819SStefano Zampini } 5012674ae819SStefano Zampini 5013e456f2a8SStefano Zampini #undef __FUNCT__ 5014e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5015a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5016e456f2a8SStefano Zampini { 5017e456f2a8SStefano Zampini IS localis_t; 5018a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5019e456f2a8SStefano Zampini PetscScalar *vals; 5020e456f2a8SStefano Zampini PetscErrorCode ierr; 5021e456f2a8SStefano Zampini 5022e456f2a8SStefano Zampini PetscFunctionBegin; 5023a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5024e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5025854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5026e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5027e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5028a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5029a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 50301035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5031a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 50321035eff8SStefano Zampini } 5033a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5034e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5035e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5036a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5037a7dc3881SStefano Zampini /* now compute set in local ordering */ 5038a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5039a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5040a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5041a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5042a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5043ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5044e456f2a8SStefano Zampini lsize++; 5045e456f2a8SStefano Zampini } 5046e456f2a8SStefano Zampini } 5047854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5048a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5049ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5050e456f2a8SStefano Zampini idxs[lsize++] = i; 5051e456f2a8SStefano Zampini } 5052e456f2a8SStefano Zampini } 5053a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5054a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5055e456f2a8SStefano Zampini *localis = localis_t; 5056e456f2a8SStefano Zampini PetscFunctionReturn(0); 5057e456f2a8SStefano Zampini } 5058906d46d4SStefano Zampini 5059906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5060906d46d4SStefano Zampini #undef __FUNCT__ 5061906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5062906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5063906d46d4SStefano Zampini { 5064906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5065906d46d4SStefano Zampini PetscErrorCode ierr; 5066906d46d4SStefano Zampini 5067906d46d4SStefano Zampini PetscFunctionBegin; 5068906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5069906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5070906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5071906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5072906d46d4SStefano Zampini PetscFunctionReturn(0); 5073906d46d4SStefano Zampini } 5074906d46d4SStefano Zampini 5075906d46d4SStefano Zampini #undef __FUNCT__ 5076906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5077906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5078906d46d4SStefano Zampini { 5079906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5080906d46d4SStefano Zampini PetscErrorCode ierr; 5081906d46d4SStefano Zampini 5082906d46d4SStefano Zampini PetscFunctionBegin; 5083906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5084906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5085906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5086906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5087906d46d4SStefano Zampini PetscFunctionReturn(0); 5088906d46d4SStefano Zampini } 5089b96c3477SStefano Zampini 5090b96c3477SStefano Zampini #undef __FUNCT__ 5091b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 509208122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5093b96c3477SStefano Zampini { 5094a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5095b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5096b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5097a64f4aa4SStefano Zampini Mat S_j; 5098b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5099b96c3477SStefano Zampini PetscBool free_used_adj; 5100b96c3477SStefano Zampini PetscErrorCode ierr; 5101b96c3477SStefano Zampini 5102b96c3477SStefano Zampini PetscFunctionBegin; 5103b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5104b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 510508122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5106b96c3477SStefano Zampini used_xadj = NULL; 5107b96c3477SStefano Zampini used_adjncy = NULL; 5108b96c3477SStefano Zampini } else { 510908122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 511008122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 511108122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 511208122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5113b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5114b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5115b96c3477SStefano Zampini } else { 51162fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5117b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5118b96c3477SStefano Zampini PetscInt nvtxs; 5119b96c3477SStefano Zampini 51202fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 51212fffb893SStefano Zampini if (flg_row) { 5122b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5123b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5124b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5125b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 51262fffb893SStefano Zampini } else { 51272fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 51282fffb893SStefano Zampini used_xadj = NULL; 51292fffb893SStefano Zampini used_adjncy = NULL; 51302fffb893SStefano Zampini } 51312fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5132b96c3477SStefano Zampini } 5133b96c3477SStefano Zampini } 5134d5574798SStefano Zampini 5135d5574798SStefano Zampini /* setup sub_schurs data */ 5136a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5137a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5138a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5139a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 514006a4e24aSStefano 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); 5141a64f4aa4SStefano Zampini } else { 51426816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 514304708bb6SStefano Zampini PetscBool isseqaij; 51445feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 51455feab87aSStefano Zampini PetscInt n_vertices; 51465feab87aSStefano Zampini 51475feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 51482034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 51495feab87aSStefano Zampini } 515004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 515104708bb6SStefano Zampini if (!isseqaij) { 515204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 515304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 515404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 515504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 515604708bb6SStefano Zampini } else { 515704708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 515804708bb6SStefano Zampini } 515904708bb6SStefano Zampini } 516006a4e24aSStefano 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); 5161a64f4aa4SStefano Zampini } 5162a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5163b96c3477SStefano Zampini 5164b96c3477SStefano Zampini /* free adjacency */ 5165b96c3477SStefano Zampini if (free_used_adj) { 5166b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5167b96c3477SStefano Zampini } 5168b96c3477SStefano Zampini PetscFunctionReturn(0); 5169b96c3477SStefano Zampini } 5170b96c3477SStefano Zampini 5171b96c3477SStefano Zampini #undef __FUNCT__ 5172b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 517308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5174b96c3477SStefano Zampini { 5175b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5176b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5177b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5178b96c3477SStefano Zampini PCBDDCGraph graph; 5179b96c3477SStefano Zampini PetscErrorCode ierr; 5180b96c3477SStefano Zampini 5181b96c3477SStefano Zampini PetscFunctionBegin; 5182b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 518308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 51843301b35fSStefano Zampini IS verticesIS,verticescomm; 51853301b35fSStefano Zampini PetscInt vsize,*idxs; 5186b96c3477SStefano Zampini 5187b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 51883301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 51893301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51903301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 51913301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51923301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5193b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 51947fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 51953301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 51963301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5197b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5198b96c3477SStefano Zampini /* 5199b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5200b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5201b96c3477SStefano Zampini } 5202b96c3477SStefano Zampini */ 5203b96c3477SStefano Zampini } else { 5204b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5205b96c3477SStefano Zampini } 5206b96c3477SStefano Zampini 5207b96c3477SStefano Zampini /* sub_schurs init */ 5208a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5209a64f4aa4SStefano Zampini 5210b96c3477SStefano Zampini /* free graph struct */ 521108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5212b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5213b96c3477SStefano Zampini } 5214b96c3477SStefano Zampini PetscFunctionReturn(0); 5215b96c3477SStefano Zampini } 5216