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. 521c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 53c263805aSStefano Zampini */ 54c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx < 0) { 55efc2fbd9SStefano Zampini PetscFunctionReturn(0); 56efc2fbd9SStefano Zampini } 57c263805aSStefano Zampini if (pop) { 58c263805aSStefano Zampini const PetscInt *cB0_cols; 59c263805aSStefano Zampini PetscInt cB0_ncol; 60c263805aSStefano Zampini const PetscScalar *cB0_vals; 61c263805aSStefano Zampini 62c263805aSStefano Zampini /* extract B_0 */ 63efc2fbd9SStefano Zampini ierr = MatGetRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 64c263805aSStefano Zampini pcbddc->B0_ncol = cB0_ncol; 65c263805aSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 66c263805aSStefano Zampini ierr = PetscMalloc2(cB0_ncol,&pcbddc->B0_cols,cB0_ncol,&pcbddc->B0_vals);CHKERRQ(ierr); 67c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_cols,cB0_cols,cB0_ncol*sizeof(PetscInt));CHKERRQ(ierr); 68c263805aSStefano Zampini ierr = PetscMemcpy(pcbddc->B0_vals,cB0_vals,cB0_ncol*sizeof(PetscScalar));CHKERRQ(ierr); 69efc2fbd9SStefano Zampini ierr = MatRestoreRow(pcbddc->local_mat,pcbddc->benign_p0_lidx,&cB0_ncol,&cB0_cols,&cB0_vals);CHKERRQ(ierr); 70c263805aSStefano Zampini /* remove rows and cols from local problem */ 71c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 72efc2fbd9SStefano Zampini ierr = MatZeroRowsColumns(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,1.,NULL,NULL);CHKERRQ(ierr); 73c263805aSStefano Zampini } else { /* push */ 74efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,&pcbddc->benign_p0_lidx,pcbddc->B0_ncol,pcbddc->B0_cols,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 75efc2fbd9SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,pcbddc->B0_ncol,pcbddc->B0_cols,1,&pcbddc->benign_p0_lidx,pcbddc->B0_vals,INSERT_VALUES);CHKERRQ(ierr); 76efc2fbd9SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx,pcbddc->benign_p0_lidx,0.0,INSERT_VALUES);CHKERRQ(ierr); 77c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 78c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 79c263805aSStefano Zampini } 80c263805aSStefano Zampini PetscFunctionReturn(0); 81c263805aSStefano Zampini } 82c263805aSStefano Zampini 83c263805aSStefano Zampini #undef __FUNCT__ 84b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 8508122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 86b1b3d7a2SStefano Zampini { 87b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 8808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 8908122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 9008122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 9108122e43SStefano Zampini PetscScalar *work,lwork; 9208122e43SStefano Zampini PetscScalar *St,*S,*eigv; 9308122e43SStefano Zampini PetscScalar *Sarray,*Starray; 9408122e43SStefano Zampini PetscReal *eigs,thresh; 951b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 96f6f667cfSStefano Zampini PetscBool allocated_S_St; 9708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9808122e43SStefano Zampini PetscReal *rwork; 9908122e43SStefano Zampini #endif 100b1b3d7a2SStefano Zampini PetscErrorCode ierr; 101b1b3d7a2SStefano Zampini 102b1b3d7a2SStefano Zampini PetscFunctionBegin; 10308122e43SStefano Zampini if (!sub_schurs->use_mumps) { 10408122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 10508122e43SStefano Zampini } 10608122e43SStefano Zampini 10706a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 10806a4e24aSStefano 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); 10906a4e24aSStefano Zampini } 11006a4e24aSStefano Zampini 111fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 112fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 113fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 114fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 115fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 116fd14bc51SStefano Zampini } 117fd14bc51SStefano Zampini 118e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 119e496cd5dSStefano 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); 120e496cd5dSStefano Zampini } 121e496cd5dSStefano Zampini 12208122e43SStefano Zampini /* max size of subsets */ 12308122e43SStefano Zampini mss = 0; 12408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 12508122e43SStefano Zampini PetscInt subset_size; 126862806e4SStefano Zampini 12708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 12808122e43SStefano Zampini mss = PetscMax(mss,subset_size); 12908122e43SStefano Zampini } 13008122e43SStefano Zampini 13108122e43SStefano Zampini /* min/max and threshold */ 13208122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 133f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 13408122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 135f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 136f6f667cfSStefano Zampini if (nmin) { 137f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 138f6f667cfSStefano Zampini } 13908122e43SStefano Zampini 14008122e43SStefano Zampini /* allocate lapack workspace */ 14108122e43SStefano Zampini cum = cum2 = 0; 14208122e43SStefano Zampini maxneigs = 0; 14308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 14408122e43SStefano Zampini PetscInt n,subset_size; 145f6f667cfSStefano Zampini 14608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 14708122e43SStefano Zampini n = PetscMin(subset_size,nmax); 1489162d606SStefano Zampini cum += subset_size; 1499162d606SStefano Zampini cum2 += subset_size*n; 15008122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 15108122e43SStefano Zampini } 15208122e43SStefano Zampini if (mss) { 1539ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 15408122e43SStefano Zampini PetscBLASInt B_itype = 1; 15508122e43SStefano Zampini PetscBLASInt B_N = mss; 1564c6709b3SStefano Zampini PetscReal zero = 0.0; 1574c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 15808122e43SStefano Zampini 15908122e43SStefano Zampini B_lwork = -1; 16008122e43SStefano Zampini S = NULL; 16108122e43SStefano Zampini St = NULL; 162a58a30b4SStefano Zampini eigs = NULL; 163a58a30b4SStefano Zampini eigv = NULL; 164a58a30b4SStefano Zampini B_iwork = NULL; 165a58a30b4SStefano Zampini B_ifail = NULL; 166d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 167d1710679SStefano Zampini rwork = NULL; 168d1710679SStefano Zampini #endif 1698bec7fa6SStefano Zampini thresh = 1.0; 17008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 17108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 17208122e43SStefano 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)); 17308122e43SStefano Zampini #else 17408122e43SStefano 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)); 17508122e43SStefano Zampini #endif 17608122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 17708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 17808122e43SStefano Zampini } else { 17908122e43SStefano Zampini /* TODO */ 18008122e43SStefano Zampini } 18108122e43SStefano Zampini } else { 18208122e43SStefano Zampini lwork = 0; 18308122e43SStefano Zampini } 18408122e43SStefano Zampini 18508122e43SStefano Zampini nv = 0; 186d62866d3SStefano 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) */ 187d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 18808122e43SStefano Zampini } 1894c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 190f6f667cfSStefano Zampini if (allocated_S_St) { 191f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 192f6f667cfSStefano Zampini } 193f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 19408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 19508122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 19608122e43SStefano Zampini #endif 1979162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1989162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1999162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 20008122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 2019162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 20208122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 20308122e43SStefano Zampini 20408122e43SStefano Zampini maxneigs = 0; 20508122e43SStefano Zampini cum = cum2 = cumarray = 0; 2069162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 2079162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 208d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 20908122e43SStefano Zampini const PetscInt *idxs; 21008122e43SStefano Zampini 211d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 21208122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 21308122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 21408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 21508122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 2169162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 2179162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 21808122e43SStefano Zampini } 21908122e43SStefano Zampini cum2 = cum; 220d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 22108122e43SStefano Zampini } 22208122e43SStefano Zampini 22308122e43SStefano Zampini if (mss) { /* multilevel */ 22408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 22508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 22608122e43SStefano Zampini } 22708122e43SStefano Zampini 22808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 22908122e43SStefano Zampini 23008122e43SStefano Zampini const PetscInt *idxs; 231f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 232862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 23308122e43SStefano Zampini PetscBLASInt B_N; 234aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 23508122e43SStefano Zampini 236862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 237f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 238f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 2399ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 240aff50787SStefano Zampini PetscInt j,k; 241aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 242aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 243aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24408122e43SStefano Zampini } 24508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 246aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 247aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 248aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 249aff50787SStefano Zampini } 25008122e43SStefano Zampini } 25108122e43SStefano Zampini } else { 25208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 25408122e43SStefano Zampini } 2558bec7fa6SStefano Zampini } else { 256f6f667cfSStefano Zampini S = Sarray + cumarray; 257f6f667cfSStefano Zampini St = Starray + cumarray; 2588bec7fa6SStefano Zampini } 25908122e43SStefano Zampini 260f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 261aff50787SStefano Zampini /* see if we can save some work */ 262aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 263aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 264aff50787SStefano Zampini } 265aff50787SStefano Zampini 266aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 267aff50787SStefano Zampini B_neigs = 0; 268aff50787SStefano Zampini } else { 269aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 270f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 271f6f667cfSStefano Zampini 2729ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 27308122e43SStefano Zampini PetscBLASInt B_itype = 1; 274f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 2754c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 2769552c7c7SStefano Zampini PetscInt nmin_s; 27708122e43SStefano Zampini 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 } 281*d16cbb6bSStefano Zampini 28208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 283*d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 284*d16cbb6bSStefano Zampini 285*d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 28608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 287f6f667cfSStefano 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)); 28808122e43SStefano Zampini #else 289f6f667cfSStefano 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)); 29008122e43SStefano Zampini #endif 291*d16cbb6bSStefano Zampini } else { 292*d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 293*d16cbb6bSStefano Zampini B_IL = 1; 294*d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 295*d16cbb6bSStefano 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_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 296*d16cbb6bSStefano Zampini #else 297*d16cbb6bSStefano 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_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 298*d16cbb6bSStefano Zampini #endif 299*d16cbb6bSStefano Zampini } 30008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 30108122e43SStefano Zampini if (B_ierr) { 30208122e43SStefano Zampini if (B_ierr < 0 ) { 30308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 30408122e43SStefano Zampini } else if (B_ierr <= B_N) { 30508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 30608122e43SStefano Zampini } else { 3079552c7c7SStefano 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); 30808122e43SStefano Zampini } 30908122e43SStefano Zampini } 31008122e43SStefano Zampini 31108122e43SStefano Zampini if (B_neigs > nmax) { 312fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 313fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 314fd14bc51SStefano Zampini } 315f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 31608122e43SStefano Zampini B_neigs = nmax; 31708122e43SStefano Zampini } 31808122e43SStefano Zampini 3199552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 3209552c7c7SStefano Zampini if (B_neigs < nmin_s) { 32108122e43SStefano Zampini PetscBLASInt B_neigs2; 32208122e43SStefano Zampini 323f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 324f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 325fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 326fd14bc51SStefano 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); 327fd14bc51SStefano Zampini } 3289ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 32908122e43SStefano Zampini PetscInt j; 33008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 33108122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 33208122e43SStefano Zampini } 33308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 33408122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 33508122e43SStefano Zampini } 33608122e43SStefano Zampini } else { 33708122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 33808122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 33908122e43SStefano Zampini } 34008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 342f6f667cfSStefano 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)); 34308122e43SStefano Zampini #else 344f6f667cfSStefano 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)); 34508122e43SStefano Zampini #endif 34608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 34708122e43SStefano Zampini B_neigs += B_neigs2; 34808122e43SStefano Zampini } 34908122e43SStefano Zampini if (B_ierr) { 35008122e43SStefano Zampini if (B_ierr < 0 ) { 35108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 35208122e43SStefano Zampini } else if (B_ierr <= B_N) { 35308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 35408122e43SStefano Zampini } else { 3559552c7c7SStefano 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); 35608122e43SStefano Zampini } 35708122e43SStefano Zampini } 358fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 359ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 36008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 36108122e43SStefano Zampini if (eigs[j] == 0.0) { 362ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 36308122e43SStefano Zampini } else { 364ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 365fd14bc51SStefano Zampini } 36608122e43SStefano Zampini } 36708122e43SStefano Zampini } 36808122e43SStefano Zampini } else { 36908122e43SStefano Zampini /* TODO */ 37008122e43SStefano Zampini } 371aff50787SStefano Zampini } 3728bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 3738bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 3749162d606SStefano Zampini if (B_neigs) { 3759162d606SStefano 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); 376fd14bc51SStefano Zampini 377fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 3789552c7c7SStefano Zampini PetscInt ii; 3799552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 380ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 3819552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 382ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 383ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 384ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 385ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 386ac47001eSStefano Zampini #else 387ac47001eSStefano 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); 388ac47001eSStefano Zampini #endif 3899552c7c7SStefano Zampini } 3909552c7c7SStefano Zampini } 391fd14bc51SStefano Zampini } 39208122e43SStefano Zampini #if 0 3939162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 39408122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 39508122e43SStefano Zampini PetscScalar norm; 39608122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3979162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3989162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 39908122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 40008122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 40108122e43SStefano Zampini } else { 40208122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 40308122e43SStefano Zampini } 4049162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 405b1b3d7a2SStefano Zampini } 406b1b3d7a2SStefano Zampini #endif 4079162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 4089162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 4099162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 4109162d606SStefano Zampini cum++; 41108122e43SStefano Zampini } 41208122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 41308122e43SStefano Zampini /* shift for next computation */ 41408122e43SStefano Zampini cumarray += subset_size*subset_size; 41508122e43SStefano Zampini } 416fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 417fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 418fd14bc51SStefano Zampini } 41908122e43SStefano Zampini 42008122e43SStefano Zampini if (mss) { 42108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 42208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 423f6f667cfSStefano Zampini /* destroy matrices (junk) */ 424f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 425f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 42608122e43SStefano Zampini } 427f6f667cfSStefano Zampini if (allocated_S_St) { 428f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 429f6f667cfSStefano Zampini } 430f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 43108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 43208122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 43308122e43SStefano Zampini #endif 43408122e43SStefano Zampini if (pcbddc->dbg_flag) { 4351b968477SStefano Zampini PetscInt maxneigs_r; 43608122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4379b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 43808122e43SStefano Zampini } 43908122e43SStefano Zampini PetscFunctionReturn(0); 44008122e43SStefano Zampini } 441b1b3d7a2SStefano Zampini 442674ae819SStefano Zampini #undef __FUNCT__ 443c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 444c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 445c8587f34SStefano Zampini { 446c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4478629588bSStefano Zampini PetscScalar *coarse_submat_vals; 448c8587f34SStefano Zampini PetscErrorCode ierr; 449c8587f34SStefano Zampini 450c8587f34SStefano Zampini PetscFunctionBegin; 451f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 4525e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 453c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 454c8587f34SStefano Zampini 455684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 4560fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 457684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 458c8587f34SStefano Zampini 459c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 460b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 461c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 462c8587f34SStefano Zampini } 463c8587f34SStefano Zampini 4648629588bSStefano Zampini /* 4658629588bSStefano Zampini Setup local correction and local part of coarse basis. 4668629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 4678629588bSStefano Zampini */ 46847f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 4698629588bSStefano Zampini 4708629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 4718629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 4728629588bSStefano Zampini 4738629588bSStefano Zampini /* free */ 4748629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 475c8587f34SStefano Zampini PetscFunctionReturn(0); 476c8587f34SStefano Zampini } 477c8587f34SStefano Zampini 478c8587f34SStefano Zampini #undef __FUNCT__ 479674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 480674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 481674ae819SStefano Zampini { 482674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 483674ae819SStefano Zampini PetscErrorCode ierr; 484674ae819SStefano Zampini 485674ae819SStefano Zampini PetscFunctionBegin; 486674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 487674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 488674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 489674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 490785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 491674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 492f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 493f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 494785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 49563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 49663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 497674ae819SStefano Zampini PetscFunctionReturn(0); 498674ae819SStefano Zampini } 499674ae819SStefano Zampini 500674ae819SStefano Zampini #undef __FUNCT__ 501674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 502674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 503674ae819SStefano Zampini { 504674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 505674ae819SStefano Zampini PetscErrorCode ierr; 506674ae819SStefano Zampini 507674ae819SStefano Zampini PetscFunctionBegin; 508b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 509674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 510674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 511674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 512b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 513674ae819SStefano Zampini PetscFunctionReturn(0); 514674ae819SStefano Zampini } 515674ae819SStefano Zampini 516674ae819SStefano Zampini #undef __FUNCT__ 517674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 518674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 519674ae819SStefano Zampini { 520674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 52106656605SStefano Zampini PetscScalar *array; 522674ae819SStefano Zampini PetscErrorCode ierr; 523674ae819SStefano Zampini 524674ae819SStefano Zampini PetscFunctionBegin; 525674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 52658da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 52706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 52806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 52958da7f69SStefano Zampini } 530674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 531674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 53215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 53315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 534674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 535674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 536674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 53706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 538674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 539674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5408ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 541674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 542674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 543674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 544f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 545f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 546f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 547f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 548727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 5490e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 550f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 55170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 5526e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55381d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 5540369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 5558b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 5568b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 55781d14e9dSStefano Zampini ierr = PetscFree2(pcbddc->B0_cols,pcbddc->B0_vals);CHKERRQ(ierr); 558674ae819SStefano Zampini PetscFunctionReturn(0); 559674ae819SStefano Zampini } 560674ae819SStefano Zampini 561674ae819SStefano Zampini #undef __FUNCT__ 562f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 563f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 5646bfb1811SStefano Zampini { 5656bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5666bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 5676bfb1811SStefano Zampini VecType impVecType; 568019a44ceSStefano Zampini PetscInt n_constraints,n_R,old_size,n_benign; 5696bfb1811SStefano Zampini PetscErrorCode ierr; 5706bfb1811SStefano Zampini 5716bfb1811SStefano Zampini PetscFunctionBegin; 572f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 573019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 574f4ddd8eeSStefano Zampini } 575e7b262bdSStefano Zampini /* get sizes */ 576c9ed8603SStefano Zampini n_benign = 0; 577c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) n_benign = 1; 578019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - pcbddc->n_vertices; 579b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 5806bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 581e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 582e7b262bdSStefano Zampini /* R nodes */ 583e7b262bdSStefano Zampini old_size = -1; 584e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 585e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 586e7b262bdSStefano Zampini } 587e7b262bdSStefano Zampini if (n_R != old_size) { 588e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 589e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 5906bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 5916bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 5926bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 5936bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 594e7b262bdSStefano Zampini } 595e7b262bdSStefano Zampini /* local primal dofs */ 596e7b262bdSStefano Zampini old_size = -1; 597e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 598e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 599e7b262bdSStefano Zampini } 600e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 601e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 60283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 603e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 6046bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 605e7b262bdSStefano Zampini } 606e7b262bdSStefano Zampini /* local explicit constraints */ 607e7b262bdSStefano Zampini old_size = -1; 608e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 609e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 610e7b262bdSStefano Zampini } 611e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 612e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 61383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 61483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 61583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 61683b7ccabSStefano Zampini } 6176bfb1811SStefano Zampini PetscFunctionReturn(0); 6186bfb1811SStefano Zampini } 6196bfb1811SStefano Zampini 6206bfb1811SStefano Zampini #undef __FUNCT__ 62147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 62247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 62388ebb749SStefano Zampini { 62425084f0cSStefano Zampini PetscErrorCode ierr; 62525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 62688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 62788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 628d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 62925084f0cSStefano Zampini /* submatrices of local problem */ 63080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 631019a44ceSStefano Zampini /* submatrices of benign trick */ 632*d16cbb6bSStefano Zampini Mat B0_V = NULL; 63306656605SStefano Zampini /* submatrices of local coarse problem */ 63406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 63525084f0cSStefano Zampini /* working matrices */ 63606656605SStefano Zampini Mat C_CR; 63725084f0cSStefano Zampini /* additional working stuff */ 63806656605SStefano Zampini PC pc_R; 639d12edf2fSStefano Zampini Mat F,B0 = NULL; 64006656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 64106656605SStefano Zampini 64225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 64306656605SStefano Zampini PetscScalar *work; 64406656605SStefano Zampini PetscInt *idx_V_B; 645d12edf2fSStefano Zampini PetscInt n,n_vertices,n_constraints,n_benign,p0_lidx_I = 0; 64606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 647b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 64845a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 64988ebb749SStefano Zampini MatType impMatType; 65025084f0cSStefano Zampini /* some shortcuts to scalars */ 65106656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 65288ebb749SStefano Zampini 65388ebb749SStefano Zampini PetscFunctionBegin; 654c9ed8603SStefano Zampini n_benign = 0; 655c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) n_benign = 1; 656b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 657019a44ceSStefano Zampini n_constraints = pcbddc->local_primal_size - n_benign - n_vertices; 65888ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 659b371cd4fSStefano Zampini n_B = pcis->n_B; 660b371cd4fSStefano Zampini n_D = pcis->n - n_B; 66188ebb749SStefano Zampini n_R = pcis->n - n_vertices; 66288ebb749SStefano Zampini 66388ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 66488ebb749SStefano Zampini impMatType = MATSEQDENSE; 66588ebb749SStefano Zampini 66688ebb749SStefano Zampini /* vertices in boundary numbering */ 667785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 6680e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 66988ebb749SStefano Zampini if (i != n_vertices) { 67022d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 67188ebb749SStefano Zampini } 67288ebb749SStefano Zampini 67306656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 674019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 67506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 67706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 67806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 67906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 68006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 68106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 68206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 68306656605SStefano Zampini 68406656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 68506656605SStefano Zampini /* allocate workspace */ 68606656605SStefano Zampini n = 0; 68706656605SStefano Zampini if (n_constraints) { 68806656605SStefano Zampini n += n_R*n_constraints; 68906656605SStefano Zampini } 69006656605SStefano Zampini if (n_vertices) { 69106656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 69280677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 69306656605SStefano Zampini } 6943301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 69506656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 69606656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 69706656605SStefano Zampini } 69806656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 69906656605SStefano Zampini 70006656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 70106656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 70206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 70306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 70406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 70506656605SStefano Zampini if (isLU || isILU || isCHOL) { 70606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 707d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 708d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 709d62866d3SStefano Zampini MatFactorType type; 710d62866d3SStefano Zampini 7116816873aSStefano Zampini F = reuse_mumps->F; 7126816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 713d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 71406656605SStefano Zampini } else { 71506656605SStefano Zampini F = NULL; 71606656605SStefano Zampini } 71706656605SStefano Zampini 71888ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 71988ebb749SStefano Zampini if (n_constraints) { 72006656605SStefano Zampini Mat M1,M2,M3; 72180677318SStefano Zampini Mat auxmat; 72206656605SStefano Zampini IS is_aux; 72380677318SStefano Zampini PetscScalar *array,*array2; 72406656605SStefano Zampini 725f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 72680677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 72788ebb749SStefano Zampini 72825084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 72925084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 7308ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 73180677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 73288ebb749SStefano Zampini 73380677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 73480677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 73506656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 73688ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 73706656605SStefano Zampini const PetscScalar *row_cmat_values; 73806656605SStefano Zampini const PetscInt *row_cmat_indices; 73906656605SStefano Zampini PetscInt size_of_constraint,j; 74088ebb749SStefano Zampini 74106656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 74206656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 74306656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 74406656605SStefano Zampini } 74506656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 74606656605SStefano Zampini } 74780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 74806656605SStefano Zampini if (F) { 74906656605SStefano Zampini Mat B; 75006656605SStefano Zampini 75106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 75280677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 75306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 75406656605SStefano Zampini } else { 75580677318SStefano Zampini PetscScalar *marr; 75680677318SStefano Zampini 75780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 75806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 75906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 76080677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 76106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 76206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 76306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 76406656605SStefano Zampini } 76580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 76606656605SStefano Zampini } 76780677318SStefano Zampini if (!pcbddc->switch_static) { 76880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 76980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 77080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 77180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 77280677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 77380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 77480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 77580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 77680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 77780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 77880677318SStefano Zampini } 77980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 78080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 78180677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 78280677318SStefano Zampini } else { 78380677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 78480677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 78525084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 78680677318SStefano Zampini } 78780677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 78880677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 78980677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 79006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 79106656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 79280677318SStefano Zampini if (isCHOL) { 79380677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 79480677318SStefano Zampini } else { 79525084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 79680677318SStefano Zampini } 79780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 79806656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 79925084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 80025084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 80125084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 80280677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 80380677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 80480677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 80506656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 80606656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 807f4ddd8eeSStefano Zampini } 80888ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 809*d16cbb6bSStefano Zampini if (n_benign) { 810*d16cbb6bSStefano Zampini IS dummy; 811*d16cbb6bSStefano Zampini Mat B0_R; 812*d16cbb6bSStefano Zampini PetscReal norm; 813*d16cbb6bSStefano Zampini PetscInt ii[2]; 814*d16cbb6bSStefano Zampini 815*d16cbb6bSStefano Zampini ii[0] = 0; 816*d16cbb6bSStefano Zampini ii[1] = pcbddc->B0_ncol; 817*d16cbb6bSStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,1,pcis->n,ii,pcbddc->B0_cols,pcbddc->B0_vals,&B0);CHKERRQ(ierr); 818*d16cbb6bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 819*d16cbb6bSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 820*d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 821*d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 822*d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 823*d16cbb6bSStefano Zampini } 824*d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 825*d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 826*d16cbb6bSStefano Zampini } 827*d16cbb6bSStefano Zampini 82888ebb749SStefano Zampini if (n_vertices) { 82906656605SStefano Zampini IS is_aux; 8303a50541eSStefano Zampini 8316816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 8326816873aSStefano Zampini IS tis; 8336816873aSStefano Zampini 8346816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 8356816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 8366816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 8376816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 8386816873aSStefano Zampini } else { 8393a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 8406816873aSStefano Zampini } 8419577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 8429577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 84304708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 844019a44ceSStefano Zampini if (n_benign) { 845019a44ceSStefano Zampini IS dummy; 846019a44ceSStefano Zampini 847019a44ceSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&dummy);CHKERRQ(ierr); 848019a44ceSStefano Zampini ierr = MatGetSubMatrix(B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 849019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 850019a44ceSStefano Zampini } 85125084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 85288ebb749SStefano Zampini } 85388ebb749SStefano Zampini 85488ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 855f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 85606656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 85706656605SStefano Zampini if (pcbddc->coarse_phi_D) { 85806656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 85906656605SStefano Zampini } 860f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 86106656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 86206656605SStefano Zampini PetscScalar *marray; 86306656605SStefano Zampini 86406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 866f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 867f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 868f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 869f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 870f4ddd8eeSStefano Zampini } 871f4ddd8eeSStefano Zampini } 87206656605SStefano Zampini 873f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 87406656605SStefano Zampini PetscScalar *marray; 87588ebb749SStefano Zampini 87606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 87806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 87988ebb749SStefano Zampini } 8803301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 88106656605SStefano Zampini n *= 2; 88288ebb749SStefano Zampini } 88306656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 88406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 88506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 8868eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 88706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 88806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 88988ebb749SStefano Zampini } 8903301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 89106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 8928eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 89306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 89406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 89588ebb749SStefano Zampini } 89688ebb749SStefano Zampini } else { 897c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 898c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 8991b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 900c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 901c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 902c0553b1fSStefano Zampini } 90388ebb749SStefano Zampini } 90406656605SStefano Zampini } 905019a44ceSStefano Zampini 90606656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 907d12edf2fSStefano Zampini if (n_benign && (pcbddc->switch_static || pcbddc->dbg_flag)) { 908d12edf2fSStefano Zampini const PetscInt *idxs; 909d12edf2fSStefano Zampini 910d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 911d12edf2fSStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx,pcis->n-pcis->n_B,idxs,&p0_lidx_I);CHKERRQ(ierr); 912d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 913d12edf2fSStefano Zampini } 914*d16cbb6bSStefano Zampini 91506656605SStefano Zampini /* vertices */ 91606656605SStefano Zampini if (n_vertices) { 91716f15bc4SStefano Zampini 91804708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 91904708bb6SStefano Zampini 92016f15bc4SStefano Zampini if (n_R) { 92106656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 92206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 92316f15bc4SStefano Zampini PetscScalar *x,*y; 92404708bb6SStefano Zampini PetscBool isseqaij; 92506656605SStefano Zampini 92621eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 92706656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 92804708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 9296816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 93006656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 93106656605SStefano Zampini } else { 93206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 93306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 93406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 93606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 93706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 93806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 93906656605SStefano Zampini } 94006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 94106656605SStefano Zampini } 94280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 94306656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 94406656605SStefano Zampini if (n_constraints) { 94506656605SStefano Zampini Mat B; 94680677318SStefano Zampini 947b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 94880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 94980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95080677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 95180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 95380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 95480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95580677318SStefano Zampini } 95680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 95780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 95880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 96080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 96106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 96306656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 96406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 96506656605SStefano Zampini } 96604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 96704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 96804708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 96904708bb6SStefano Zampini } 97006656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 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); 97806656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 979*d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 980019a44ceSStefano Zampini } else { 981*d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 982*d16cbb6bSStefano Zampini } 983*d16cbb6bSStefano Zampini if (n_benign) { 984019a44ceSStefano Zampini const PetscScalar *vals; 985019a44ceSStefano Zampini const PetscInt *idxs; 986019a44ceSStefano Zampini PetscInt n,j; 987019a44ceSStefano Zampini 988019a44ceSStefano Zampini ierr = MatGetRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 989*d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 990*d16cbb6bSStefano Zampini coarse_submat_vals[(pcbddc->local_primal_size-1)*pcbddc->local_primal_size+idxs[j]] = vals[j]; 991*d16cbb6bSStefano Zampini coarse_submat_vals[(idxs[j]+1)*pcbddc->local_primal_size-1] = vals[j]; 992019a44ceSStefano Zampini } 993019a44ceSStefano Zampini ierr = MatRestoreRow(B0_V,0,&n,&idxs,&vals);CHKERRQ(ierr); 99416f15bc4SStefano Zampini } 99521eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 996*d16cbb6bSStefano Zampini 99706656605SStefano Zampini /* coarse basis functions */ 99806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 99916f15bc4SStefano Zampini PetscScalar *y; 100016f15bc4SStefano Zampini 100106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 100406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 100706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 100806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 100906656605SStefano Zampini 101006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 101106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 101506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1016a0aff763SStefano Zampini if (n_benign) y[n_D*i+p0_lidx_I] = 0.0; 101706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 101806656605SStefano Zampini } 101906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 102006656605SStefano Zampini } 102104708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 102204708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 102306656605SStefano Zampini } 102406656605SStefano Zampini 102506656605SStefano Zampini if (n_constraints) { 102606656605SStefano Zampini Mat B; 102706656605SStefano Zampini 102806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 102906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 103080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 103106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 103206656605SStefano Zampini if (n_vertices) { 103380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 103480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 103580677318SStefano Zampini } else { 103680677318SStefano Zampini Mat S_VCt; 103780677318SStefano Zampini 103880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 103980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 104080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 104180677318SStefano Zampini } 104206656605SStefano Zampini } 104306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 104406656605SStefano Zampini /* coarse basis functions */ 104506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 104606656605SStefano Zampini PetscScalar *y; 104706656605SStefano Zampini 104806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 104906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 105006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 105106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 105206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 105306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 105406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 105506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 105606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 105706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 105806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 105906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 106006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 1061*d16cbb6bSStefano Zampini if (n_benign) y[n_D*(i+n_vertices)+p0_lidx_I] = 0.0; 106206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 106306656605SStefano Zampini } 106406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 106506656605SStefano Zampini } 106606656605SStefano Zampini } 106780677318SStefano Zampini if (n_constraints) { 106880677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 106980677318SStefano Zampini } 107006656605SStefano Zampini 1071019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1072019a44ceSStefano Zampini 107306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 10743301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 107506656605SStefano Zampini 107606656605SStefano Zampini if (n_constraints) { 107716f15bc4SStefano Zampini Mat S_CCT,B_C; 107806656605SStefano Zampini 107980677318SStefano Zampini /* this is a lazy thing */ 108080677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 108206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 108306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 108416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 108506656605SStefano Zampini if (n_vertices) { 108616f15bc4SStefano Zampini Mat B_V,S_VCT; 108706656605SStefano Zampini 108806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 108906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 109006656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 109106656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 109216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 109306656605SStefano Zampini } 109406656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 109580677318SStefano Zampini } else { /* if there are no constraints, reset work */ 109680677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 109706656605SStefano Zampini } 109816f15bc4SStefano Zampini if (n_vertices && n_R) { 109906656605SStefano Zampini Mat A_VRT; 110080677318SStefano Zampini PetscScalar *marray; 110106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 110206656605SStefano Zampini 110380677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 110480677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 110580677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 110606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 110780677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 110880677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 110916f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 111006656605SStefano Zampini } 111106656605SStefano Zampini 111206656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 111306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 111406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 111506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 111606656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 111706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 111806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 111906656605SStefano Zampini } 112006656605SStefano Zampini } else { 112106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 112206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 112306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 112406656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 112506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 112606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 112706656605SStefano Zampini } 112806656605SStefano Zampini } 112906656605SStefano Zampini /* coarse basis functions */ 113006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 113106656605SStefano Zampini PetscScalar *y; 113206656605SStefano Zampini 113306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 113406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 113506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 113606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 113706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 113806656605SStefano Zampini if (i<n_vertices) { 113906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 114006656605SStefano Zampini } 114106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 114206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 114306656605SStefano Zampini 114406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 114506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 114606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 114706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 114806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 114906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 115006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 115106656605SStefano Zampini } 115206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 115306656605SStefano Zampini } 115406656605SStefano Zampini } 1155d62866d3SStefano Zampini /* free memory */ 115688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 115706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 115806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 115906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 116006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1161d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1162d62866d3SStefano Zampini if (n_vertices) { 1163d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1164d62866d3SStefano Zampini } 1165d62866d3SStefano Zampini if (n_constraints) { 1166d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1167d62866d3SStefano Zampini } 116888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 116988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 117088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 1171d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 117288ebb749SStefano Zampini Mat coarse_sub_mat; 117325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 117488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 117588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 117688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 11778bec7fa6SStefano Zampini Mat C_B,CPHI; 11788bec7fa6SStefano Zampini IS is_dummy; 11798bec7fa6SStefano Zampini Vec mones; 118088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 118188ebb749SStefano Zampini PetscReal real_value; 118288ebb749SStefano Zampini 118388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 118488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 118588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 118688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 118788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 118888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1189c0553b1fSStefano Zampini if (unsymmetric_check) { 119088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 119188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 119288ebb749SStefano Zampini } 119388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 119488ebb749SStefano Zampini 119525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 11963301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 119725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1198c0553b1fSStefano Zampini if (unsymmetric_check) { 119988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 120188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 120288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 120488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 120588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 120788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 120888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 120988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 121088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 121188ebb749SStefano Zampini } else { 121288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 121388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 121488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 121588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 121688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 121788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 121888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 121988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 122088ebb749SStefano Zampini } 122188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 122288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 122388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 122488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 1225d12edf2fSStefano Zampini if (n_benign) { 1226d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 1227d12edf2fSStefano Zampini PetscScalar *data,*data2; 1228d12edf2fSStefano Zampini 1229d12edf2fSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,1,0,1,&is_dummy);CHKERRQ(ierr); 1230d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 1231d12edf2fSStefano Zampini ierr = MatGetSubMatrix(B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 1232d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 1233d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 1234d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 1235d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 1236d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 1237*d16cbb6bSStefano Zampini data[(pcbddc->local_primal_size-1)*pcbddc->local_primal_size+i] += data2[i]; 1238*d16cbb6bSStefano Zampini data[(i+1)*pcbddc->local_primal_size-1] += data2[i]; 1239d12edf2fSStefano Zampini } 1240d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 1241d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 1242d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 1243d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 1244d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 1245d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 1246d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 1247d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1248d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 1249d12edf2fSStefano Zampini } 1250d12edf2fSStefano Zampini #if 0 1251d12edf2fSStefano Zampini { 1252d12edf2fSStefano Zampini PetscViewer viewer; 1253d12edf2fSStefano Zampini char filename[256]; 1254d12edf2fSStefano Zampini sprintf(filename,"proj_local_coarse_mat%d.m",PetscGlobalRank); 1255d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 1256d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1257d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 1258d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 1259d12edf2fSStefano Zampini } 1260d12edf2fSStefano Zampini #endif 126181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 12628bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 12630fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 126406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 12658bec7fa6SStefano Zampini 12668bec7fa6SStefano Zampini /* check constraints */ 1267d12edf2fSStefano Zampini if (!n_benign) { 12688bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 12698bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 12708bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 12718bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 12728bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 12738bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 12748bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1275bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1276c0553b1fSStefano Zampini if (unsymmetric_check) { 1277bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1278bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1279bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1280bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1281bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 128288ebb749SStefano Zampini } 12838bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 12848bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 12858bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 12868bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 1287d12edf2fSStefano Zampini } 128825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 128988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 129088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 129188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 129288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 129388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 129488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 129588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 129688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 129788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 129888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1299c0553b1fSStefano Zampini if (unsymmetric_check) { 130088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 130188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 130288ebb749SStefano Zampini } 130388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 130488ebb749SStefano Zampini } 1305d12edf2fSStefano Zampini ierr = MatDestroy(&B0);CHKERRQ(ierr); 13068629588bSStefano Zampini /* get back data */ 13078629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 130888ebb749SStefano Zampini PetscFunctionReturn(0); 130988ebb749SStefano Zampini } 131088ebb749SStefano Zampini 131188ebb749SStefano Zampini #undef __FUNCT__ 1312d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1313d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1314aa0d41d4SStefano Zampini { 1315d65f70fdSStefano Zampini Mat *work_mat; 1316d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1317d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1318d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1319aa0d41d4SStefano Zampini PetscErrorCode ierr; 1320aa0d41d4SStefano Zampini 1321aa0d41d4SStefano Zampini PetscFunctionBegin; 1322d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1323d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1324d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1325d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1326aa0d41d4SStefano Zampini 1327d65f70fdSStefano Zampini if (!rsorted) { 1328906d46d4SStefano Zampini const PetscInt *idxs; 1329906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 1330aa0d41d4SStefano Zampini 1331d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1332d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1333d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1334d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1335aa0d41d4SStefano Zampini } 1336d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1337d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1338d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1339d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1340aa0d41d4SStefano Zampini } 1341d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1342d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1343d65f70fdSStefano Zampini } else { 1344d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1345d65f70fdSStefano Zampini isrow_s = isrow; 1346aa0d41d4SStefano Zampini } 1347906d46d4SStefano Zampini 1348d65f70fdSStefano Zampini if (!csorted) { 1349d65f70fdSStefano Zampini if (isrow == iscol) { 1350d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1351d65f70fdSStefano Zampini iscol_s = isrow_s; 1352d65f70fdSStefano Zampini } else { 1353d65f70fdSStefano Zampini const PetscInt *idxs; 1354d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1355906d46d4SStefano Zampini 1356d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1357d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1358d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1359d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1360d65f70fdSStefano Zampini } 1361d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1362d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1363d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1364d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1365d65f70fdSStefano Zampini } 1366d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1367d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1368d65f70fdSStefano Zampini } 1369d65f70fdSStefano Zampini } else { 1370d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1371d65f70fdSStefano Zampini iscol_s = iscol; 1372d65f70fdSStefano Zampini } 1373d65f70fdSStefano Zampini 1374d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1375d65f70fdSStefano Zampini 1376d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1377906d46d4SStefano Zampini Mat new_mat; 1378d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1379906d46d4SStefano Zampini 1380d65f70fdSStefano Zampini if (!rsorted) { 1381d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1382d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1383d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1384d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1385906d46d4SStefano Zampini } 1386d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1387d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1388d65f70fdSStefano Zampini } else { 1389d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1390906d46d4SStefano Zampini } 1391d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1392d65f70fdSStefano Zampini 1393d65f70fdSStefano Zampini if (!csorted) { 1394d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1395d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1396d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1397d65f70fdSStefano Zampini } else { 1398d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1399d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1400d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1401d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1402d65f70fdSStefano Zampini } 1403d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1404d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1405d65f70fdSStefano Zampini } 1406d65f70fdSStefano Zampini } else { 1407d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1408d65f70fdSStefano Zampini } 1409d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1410d65f70fdSStefano Zampini 1411d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1412d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1413d65f70fdSStefano Zampini work_mat[0] = new_mat; 1414d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1415d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1416d65f70fdSStefano Zampini } 1417d65f70fdSStefano Zampini 1418d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1419d65f70fdSStefano Zampini *B = work_mat[0]; 1420d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1421d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1422d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1423d65f70fdSStefano Zampini PetscFunctionReturn(0); 1424d65f70fdSStefano Zampini } 1425d65f70fdSStefano Zampini 1426d65f70fdSStefano Zampini #undef __FUNCT__ 14275e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 14285e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1429aa0d41d4SStefano Zampini { 1430aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 14315e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1432d65f70fdSStefano Zampini Mat new_mat; 14335e8657edSStefano Zampini IS is_local,is_global; 1434d65f70fdSStefano Zampini PetscInt local_size; 1435d65f70fdSStefano Zampini PetscBool isseqaij; 1436aa0d41d4SStefano Zampini PetscErrorCode ierr; 1437aa0d41d4SStefano Zampini 1438aa0d41d4SStefano Zampini PetscFunctionBegin; 1439aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 14405e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 14415e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1442b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 1443aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1444d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1445aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1446906d46d4SStefano Zampini 1447906d46d4SStefano Zampini /* check */ 1448906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1449906d46d4SStefano Zampini Vec x,x_change; 1450906d46d4SStefano Zampini PetscReal error; 1451906d46d4SStefano Zampini 14525e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1453906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 14545e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 1455e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1456e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1457d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 1458e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1459e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1460906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1461906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1462906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1463906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1464906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1465906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1466906d46d4SStefano Zampini } 1467906d46d4SStefano Zampini 146822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 14699b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 147022d5777bSStefano Zampini if (isseqaij) { 1471d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1472aa0d41d4SStefano Zampini } else { 1473aa0d41d4SStefano Zampini Mat work_mat; 1474aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1475d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1476aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1477aa0d41d4SStefano Zampini } 14783301b35fSStefano Zampini if (matis->A->symmetric_set) { 14793301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1480e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 14813301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1482e496cd5dSStefano Zampini #endif 14833301b35fSStefano Zampini } 148445a1bb75SStefano Zampini /* 148545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1486d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 148745a1bb75SStefano Zampini */ 1488d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1489aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1490aa0d41d4SStefano Zampini } 1491aa0d41d4SStefano Zampini 1492aa0d41d4SStefano Zampini #undef __FUNCT__ 1493a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 14948ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1495a64d13efSStefano Zampini { 1496a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1497a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1498d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 149953892102SStefano Zampini PetscInt *idx_R_local=NULL; 15003a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 15013a50541eSStefano Zampini PetscInt vbs,bs; 15026816873aSStefano Zampini PetscBT bitmask=NULL; 1503a64d13efSStefano Zampini PetscErrorCode ierr; 1504a64d13efSStefano Zampini 1505a64d13efSStefano Zampini PetscFunctionBegin; 1506b23d619eSStefano Zampini /* 1507b23d619eSStefano Zampini No need to setup local scatters if 1508b23d619eSStefano Zampini - primal space is unchanged 1509b23d619eSStefano Zampini AND 1510b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1511b23d619eSStefano Zampini AND 1512b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1513b23d619eSStefano Zampini */ 1514b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1515f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1516f4ddd8eeSStefano Zampini } 1517f4ddd8eeSStefano Zampini /* destroy old objects */ 1518f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1519f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1520f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1521a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1522b371cd4fSStefano Zampini n_B = pcis->n_B; 1523b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1524b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15253a50541eSStefano Zampini 1526a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 15276816873aSStefano Zampini 152853892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 15296816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1530854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 1531a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 1532a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 15330e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 1534a64d13efSStefano Zampini } 1535a64d13efSStefano Zampini 1536a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 15374641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 15386816873aSStefano Zampini idx_R_local[n_R++] = i; 1539a64d13efSStefano Zampini } 1540a64d13efSStefano Zampini } 154153892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 15426816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15436816873aSStefano Zampini 154453892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 154553892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 15466816873aSStefano Zampini } 15473a50541eSStefano Zampini 15483a50541eSStefano Zampini /* Block code */ 15493a50541eSStefano Zampini vbs = 1; 15503a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 15513a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 15523a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 15533a50541eSStefano Zampini PetscInt *vary; 1554d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1555785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 15563a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1557d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1558d3df7717SStefano 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 */ 15590e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1560d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 15613a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 15623a50541eSStefano Zampini is_blocked = PETSC_FALSE; 15633a50541eSStefano Zampini break; 15643a50541eSStefano Zampini } 15653a50541eSStefano Zampini } 1566d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1567d3df7717SStefano Zampini } else { 1568d3df7717SStefano Zampini /* Verify directly the R set */ 1569d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1570d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1571d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1572d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1573d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1574d3df7717SStefano Zampini break; 1575d3df7717SStefano Zampini } 1576d3df7717SStefano Zampini } 1577d3df7717SStefano Zampini } 1578d3df7717SStefano Zampini } 15793a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 15803a50541eSStefano Zampini vbs = bs; 15813a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 15823a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 15833a50541eSStefano Zampini } 15843a50541eSStefano Zampini } 15853a50541eSStefano Zampini } 15863a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 158753892102SStefano Zampini if (sub_schurs->reuse_mumps) { 158853892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 158953892102SStefano Zampini 159053892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 159153892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 159253892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 159353892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 159453892102SStefano Zampini } else { 15953a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 159653892102SStefano Zampini } 1597a64d13efSStefano Zampini 1598a64d13efSStefano Zampini /* print some info if requested */ 1599a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1600c9ed8603SStefano Zampini PetscInt benign = 0; 1601019a44ceSStefano Zampini 1602c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) benign = 1; 1603a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1604a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 16050fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1606a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1607a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 1608019a44ceSStefano 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); 1609a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1610a64d13efSStefano Zampini } 1611a64d13efSStefano Zampini 1612a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 16136816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16146816873aSStefano Zampini IS is_aux1,is_aux2; 16156816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 16166816873aSStefano Zampini 16173a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1618854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1619854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1620a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 16214641a718SStefano Zampini for (i=0; i<n_D; i++) { 16224641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 16234641a718SStefano Zampini } 1624a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1625a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16264641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 16274641a718SStefano Zampini aux_array1[j++] = i; 1628a64d13efSStefano Zampini } 1629a64d13efSStefano Zampini } 1630a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1631a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1632a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 16334641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 16344641a718SStefano Zampini aux_array2[j++] = i; 1635a64d13efSStefano Zampini } 1636a64d13efSStefano Zampini } 1637a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1638a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1639a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1640a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1641a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1642a64d13efSStefano Zampini 16438eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1644785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1645a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 16464641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 16474641a718SStefano Zampini aux_array1[j++] = i; 1648a64d13efSStefano Zampini } 1649a64d13efSStefano Zampini } 1650a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1651a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1652a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1653a64d13efSStefano Zampini } 16544641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 16553a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1656d62866d3SStefano Zampini } else { 165753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16586816873aSStefano Zampini IS tis; 16596816873aSStefano Zampini PetscInt schur_size; 16606816873aSStefano Zampini 166153892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 16626816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 166353892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 16646816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16656816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 16666816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 16676816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 16686816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1669d62866d3SStefano Zampini } 1670d62866d3SStefano Zampini } 1671a64d13efSStefano Zampini PetscFunctionReturn(0); 1672a64d13efSStefano Zampini } 1673a64d13efSStefano Zampini 1674304d26faSStefano Zampini 1675304d26faSStefano Zampini #undef __FUNCT__ 1676304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1677684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1678304d26faSStefano Zampini { 1679304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1680304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1681304d26faSStefano Zampini PC pc_temp; 1682304d26faSStefano Zampini Mat A_RR; 1683f4ddd8eeSStefano Zampini MatReuse reuse; 1684304d26faSStefano Zampini PetscScalar m_one = -1.0; 1685304d26faSStefano Zampini PetscReal value; 168604708bb6SStefano Zampini PetscInt n_D,n_R; 16879577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1688304d26faSStefano Zampini PetscErrorCode ierr; 1689e604994aSStefano Zampini /* prefixes stuff */ 1690312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1691e604994aSStefano Zampini size_t len; 1692304d26faSStefano Zampini 1693304d26faSStefano Zampini PetscFunctionBegin; 1694304d26faSStefano Zampini 1695e604994aSStefano Zampini /* compute prefixes */ 1696e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1697e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1698e604994aSStefano Zampini if (!pcbddc->current_level) { 1699e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1700e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1701e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1702e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1703e604994aSStefano Zampini } else { 1704e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1705312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1706e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1707e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1708312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1709312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 171034d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 171134d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1712e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1713e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1714e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1715e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1716e604994aSStefano Zampini } 1717e604994aSStefano Zampini 1718304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1719684f6988SStefano Zampini if (dirichlet) { 1720d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 17213301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 17223301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1723964fefecSStefano Zampini } 1724ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1725964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1726304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1727304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1728304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1729304d26faSStefano Zampini /* default */ 1730304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1731e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 17329577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1733304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 17349577ea80SStefano Zampini if (issbaij) { 17359577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 17369577ea80SStefano Zampini } else { 1737304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 17389577ea80SStefano Zampini } 1739304d26faSStefano Zampini /* Allow user's customization */ 1740304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1741304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1742304d26faSStefano Zampini } 1743d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1744d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1745d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1746d62866d3SStefano Zampini 1747d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1748d5574798SStefano Zampini } 1749304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1750304d26faSStefano Zampini if (!n_D) { 1751304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1752304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1753304d26faSStefano Zampini } 1754304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1755304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1756304d26faSStefano Zampini /* set ksp_D into pcis data */ 1757304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1758304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1759304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1760684f6988SStefano Zampini } 1761304d26faSStefano Zampini 1762304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1763684f6988SStefano Zampini A_RR = 0; 1764684f6988SStefano Zampini if (neumann) { 1765d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 176604708bb6SStefano Zampini PetscInt ibs,mbs; 176704708bb6SStefano Zampini PetscBool issbaij; 176804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1769f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 17708ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1771f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1772f4ddd8eeSStefano Zampini PetscInt nn_R; 177381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1774f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1775f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1776f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1777f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1778f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1779f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1780f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1781727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1782f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1783f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1784f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1785f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1786f4ddd8eeSStefano Zampini } 1787f4ddd8eeSStefano Zampini } 1788f4ddd8eeSStefano Zampini /* last check */ 1789d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1790f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1791f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1792f4ddd8eeSStefano Zampini } 1793f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1794f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1795f4ddd8eeSStefano Zampini } 1796f4ddd8eeSStefano Zampini /* extract A_RR */ 1797af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 1798af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 179904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 180004708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 180104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 180204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 180304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1804af732b37SStefano Zampini } else { 180504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 18066816873aSStefano Zampini } 180704708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 180804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 180904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 181004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 181104708bb6SStefano Zampini } else { 181204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 181304708bb6SStefano Zampini } 181404708bb6SStefano Zampini } 181504708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1816f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 18173301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 18183301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 18196816873aSStefano Zampini } 18206816873aSStefano Zampini } else { 18216816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 18226816873aSStefano Zampini 18236816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 18246816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 18256816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1826af732b37SStefano Zampini } 1827f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1828304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1829304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1830304d26faSStefano Zampini /* default */ 1831304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1832e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1833304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 18349577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 18359577ea80SStefano Zampini if (issbaij) { 18369577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 18379577ea80SStefano Zampini } else { 1838304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 18399577ea80SStefano Zampini } 1840304d26faSStefano Zampini /* Allow user's customization */ 1841304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1842304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1843304d26faSStefano Zampini } 1844d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1845304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1846304d26faSStefano Zampini if (!n_R) { 1847304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1848304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1849304d26faSStefano Zampini } 1850d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1851d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1852d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1853d62866d3SStefano Zampini 1854d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1855d62866d3SStefano Zampini } 1856304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1857304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1858684f6988SStefano Zampini } 18596816873aSStefano Zampini /* free Neumann problem's matrix */ 18606816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1861304d26faSStefano Zampini 1862304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 18630fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1864684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1865684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1866684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1867684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1868684f6988SStefano Zampini } 1869684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 18700fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 18710fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 18720fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 18730fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 18740fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1875304d26faSStefano Zampini /* need to be adapted? */ 1876b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1877b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1878b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1879304d26faSStefano Zampini /* print info */ 1880304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1881e604994aSStefano 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); 1882304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1883304d26faSStefano Zampini } 1884b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 1885298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 1886304d26faSStefano Zampini } 1887684f6988SStefano Zampini } 1888684f6988SStefano Zampini if (neumann) { /* Neumann */ 18896816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 18900fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 18910fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 18920fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 18930fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 18940fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1895304d26faSStefano Zampini /* need to be adapted? */ 1896b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1897b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1898304d26faSStefano Zampini /* print info */ 1899304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1900e604994aSStefano 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); 1901304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1902304d26faSStefano Zampini } 1903b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 1904298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 1905304d26faSStefano Zampini } 19060fccc4e9SStefano Zampini } 1907684f6988SStefano Zampini } 1908304d26faSStefano Zampini PetscFunctionReturn(0); 1909304d26faSStefano Zampini } 1910304d26faSStefano Zampini 1911304d26faSStefano Zampini #undef __FUNCT__ 1912ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 191380677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1914674ae819SStefano Zampini { 1915674ae819SStefano Zampini PetscErrorCode ierr; 1916674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1917be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1918674ae819SStefano Zampini 1919674ae819SStefano Zampini PetscFunctionBegin; 1920be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 192180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 192220c7b377SStefano Zampini } 192380677318SStefano Zampini if (!pcbddc->switch_static) { 192480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 192580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 192680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 192720c7b377SStefano Zampini } 1928be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 192980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 193080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 193120c7b377SStefano Zampini } else { 1932be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1933be83ff47SStefano Zampini 193453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193620c7b377SStefano Zampini } 1937be83ff47SStefano Zampini } else { 193880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 193980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 194080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 194180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 194280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 194380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 194480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 194580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 194680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1947674ae819SStefano Zampini } 1948674ae819SStefano Zampini } 1949be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 195080677318SStefano Zampini if (applytranspose) { 195180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 195280677318SStefano Zampini } else { 195380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 195480677318SStefano Zampini } 195553892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1956be83ff47SStefano Zampini } else { 1957be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1958be83ff47SStefano Zampini 1959be83ff47SStefano Zampini if (applytranspose) { 196053892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1961be83ff47SStefano Zampini } else { 196253892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1963be83ff47SStefano Zampini } 196453892102SStefano Zampini #endif 1965be83ff47SStefano Zampini } 196680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 196780677318SStefano Zampini if (!pcbddc->switch_static) { 1968be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 196980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1971be83ff47SStefano Zampini } else { 1972be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1973be83ff47SStefano Zampini 197453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1976be83ff47SStefano Zampini } 197780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 197880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 197980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 198080677318SStefano Zampini } 198180677318SStefano Zampini } else { 198280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 198780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 198880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 198980677318SStefano Zampini } 199080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1994674ae819SStefano Zampini } 1995674ae819SStefano Zampini PetscFunctionReturn(0); 1996674ae819SStefano Zampini } 1997674ae819SStefano Zampini 1998dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1999674ae819SStefano Zampini #undef __FUNCT__ 2000674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2001dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2002674ae819SStefano Zampini { 2003674ae819SStefano Zampini PetscErrorCode ierr; 2004674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2005674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2006674ae819SStefano Zampini const PetscScalar zero = 0.0; 2007674ae819SStefano Zampini 2008674ae819SStefano Zampini PetscFunctionBegin; 2009dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2010dc359a40SStefano Zampini if (applytranspose) { 2011674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 20128eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2013dc359a40SStefano Zampini } else { 2014674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2015674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 201615aaf578SStefano Zampini } 2017efc2fbd9SStefano Zampini 2018efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 2019c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 2020efc2fbd9SStefano Zampini PetscScalar *array; 2021efc2fbd9SStefano Zampini 2022efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2023efc2fbd9SStefano Zampini array[pcbddc->local_primal_size-1] += pcbddc->benign_p0; 2024efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2025efc2fbd9SStefano Zampini } 2026efc2fbd9SStefano Zampini 202712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 202812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 202912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203112edc857SStefano Zampini 20329f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 203312edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 203412edc857SStefano Zampini if (pcbddc->coarse_ksp) { 2035964fefecSStefano Zampini Vec rhs,sol; 2036964fefecSStefano Zampini 2037964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2038964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 203912edc857SStefano Zampini if (applytranspose) { 2040964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 204112edc857SStefano Zampini } else { 2042964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 204312edc857SStefano Zampini } 204412edc857SStefano Zampini } 2045674ae819SStefano Zampini 2046674ae819SStefano Zampini /* Local solution on R nodes */ 204780677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 204880677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 20499f00e9b4SStefano Zampini } 2050674ae819SStefano Zampini 20519f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 20529f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 205312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2054674ae819SStefano Zampini 2055674ae819SStefano Zampini /* Sum contributions from two levels */ 2056dc359a40SStefano Zampini if (applytranspose) { 2057dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 2058dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2059dc359a40SStefano Zampini } else { 2060674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 20618eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 2062dc359a40SStefano Zampini } 2063efc2fbd9SStefano Zampini /* store p0 */ 2064c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 2065efc2fbd9SStefano Zampini PetscScalar *array; 2066efc2fbd9SStefano Zampini 2067efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2068efc2fbd9SStefano Zampini pcbddc->benign_p0 = array[pcbddc->local_primal_size-1]; 2069efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2070efc2fbd9SStefano Zampini } 2071674ae819SStefano Zampini PetscFunctionReturn(0); 2072674ae819SStefano Zampini } 2073674ae819SStefano Zampini 2074674ae819SStefano Zampini #undef __FUNCT__ 2075674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 207612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 2077674ae819SStefano Zampini { 2078674ae819SStefano Zampini PetscErrorCode ierr; 2079674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 208058da7f69SStefano Zampini PetscScalar *array; 208112edc857SStefano Zampini Vec from,to; 2082674ae819SStefano Zampini 2083674ae819SStefano Zampini PetscFunctionBegin; 208412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 208512edc857SStefano Zampini from = pcbddc->coarse_vec; 208612edc857SStefano Zampini to = pcbddc->vec1_P; 208712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 208812edc857SStefano Zampini Vec tvec; 208958da7f69SStefano Zampini 209058da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 209158da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 209212edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 209358da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 209458da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 209558da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 209612edc857SStefano Zampini } 209712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 209812edc857SStefano Zampini from = pcbddc->vec1_P; 209912edc857SStefano Zampini to = pcbddc->coarse_vec; 210012edc857SStefano Zampini } 210112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 2102674ae819SStefano Zampini PetscFunctionReturn(0); 2103674ae819SStefano Zampini } 2104674ae819SStefano Zampini 2105674ae819SStefano Zampini #undef __FUNCT__ 2106674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 210712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 2108674ae819SStefano Zampini { 2109674ae819SStefano Zampini PetscErrorCode ierr; 2110674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 211158da7f69SStefano Zampini PetscScalar *array; 211212edc857SStefano Zampini Vec from,to; 2113674ae819SStefano Zampini 2114674ae819SStefano Zampini PetscFunctionBegin; 211512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 211612edc857SStefano Zampini from = pcbddc->coarse_vec; 211712edc857SStefano Zampini to = pcbddc->vec1_P; 211812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 211912edc857SStefano Zampini from = pcbddc->vec1_P; 212012edc857SStefano Zampini to = pcbddc->coarse_vec; 212112edc857SStefano Zampini } 212212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 212312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 212412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 212512edc857SStefano Zampini Vec tvec; 212658da7f69SStefano Zampini 212712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 212858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 212958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 213058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 213158da7f69SStefano Zampini } 213258da7f69SStefano Zampini } else { 213358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 213458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 213512edc857SStefano Zampini } 213612edc857SStefano Zampini } 2137674ae819SStefano Zampini PetscFunctionReturn(0); 2138674ae819SStefano Zampini } 2139674ae819SStefano Zampini 2140984c4197SStefano Zampini /* uncomment for testing purposes */ 2141984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 2142674ae819SStefano Zampini #undef __FUNCT__ 2143674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 2144674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 2145674ae819SStefano Zampini { 2146674ae819SStefano Zampini PetscErrorCode ierr; 2147674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2148674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2149674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2150984c4197SStefano Zampini /* one and zero */ 2151984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 2152984c4197SStefano Zampini /* space to store constraints and their local indices */ 21539162d606SStefano Zampini PetscScalar *constraints_data; 21549162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 21559162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 21569162d606SStefano Zampini PetscInt *constraints_n; 2157984c4197SStefano Zampini /* iterators */ 2158b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 2159984c4197SStefano Zampini /* BLAS integers */ 2160e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 2161e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 2162c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 2163727cdba6SStefano Zampini /* reuse */ 21640e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 21650e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 2166984c4197SStefano Zampini /* change of basis */ 2167b3d85658SStefano Zampini PetscBool qr_needed; 21689162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 2169984c4197SStefano Zampini /* auxiliary stuff */ 217064efe560SStefano Zampini PetscInt *nnz,*is_indices; 21718a0068c3SStefano Zampini PetscInt ncc; 2172984c4197SStefano Zampini /* some quantities */ 217345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 2174a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 2175984c4197SStefano Zampini 2176674ae819SStefano Zampini PetscFunctionBegin; 21778e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 21788e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 21798e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 2180088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 2181088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 21820e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 21830e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 21840e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21850e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 21860e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 2187088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2188cf5a6209SStefano Zampini 2189cf5a6209SStefano Zampini /* print some info */ 2190cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 2191cf5a6209SStefano Zampini IS vertices; 2192cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 2193cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 2194cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 2195cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 2196cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 2197cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2198cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 2199fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 2200fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 2201cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2202cf5a6209SStefano Zampini } 2203cf5a6209SStefano Zampini 2204cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 22059162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 2206cf5a6209SStefano Zampini MatNullSpace nearnullsp; 2207cf5a6209SStefano Zampini const Vec *nearnullvecs; 2208cf5a6209SStefano Zampini Vec *localnearnullsp; 2209cf5a6209SStefano Zampini PetscScalar *array; 2210cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2211cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2212674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2213b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2214674ae819SStefano Zampini PetscScalar *work; 2215674ae819SStefano Zampini PetscReal *singular_vals; 2216674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2217674ae819SStefano Zampini PetscReal *rwork; 2218674ae819SStefano Zampini #endif 2219674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2220674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2221674ae819SStefano Zampini #else 2222964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2223964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2224674ae819SStefano Zampini #endif 2225674ae819SStefano Zampini 2226674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2227d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2228d06fc5fdSStefano Zampini /* free unneeded index sets */ 2229d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2230d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2231674ae819SStefano Zampini } 2232d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2233d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2234d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2235d06fc5fdSStefano Zampini } 2236d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2237d06fc5fdSStefano Zampini n_ISForEdges = 0; 2238d06fc5fdSStefano Zampini } 2239d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2240d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2241d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2242d06fc5fdSStefano Zampini } 2243d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2244d06fc5fdSStefano Zampini n_ISForFaces = 0; 2245d06fc5fdSStefano Zampini } 224670022509SStefano Zampini 224770022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 224870022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 224970022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 225070022509SStefano Zampini if (pcbddc->NullSpace) { 225170022509SStefano Zampini PetscBool tbool[2],gbool[2]; 225270022509SStefano Zampini 225370022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2254b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2255d06fc5fdSStefano Zampini if (!ISForEdges) { 2256d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2257d06fc5fdSStefano Zampini } 2258b8ffe317SStefano Zampini } 2259d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2260d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2261d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2262d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2263d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 226498a51de6SStefano Zampini } 226570022509SStefano Zampini #endif 226608122e43SStefano Zampini 2267674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2268674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2269674ae819SStefano Zampini if (nearnullsp) { 2270674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2271f4ddd8eeSStefano Zampini /* remove any stored info */ 2272f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2273f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2274f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2275f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2276f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2277473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2278f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2279f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2280f4ddd8eeSStefano Zampini } 2281984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2282984c4197SStefano Zampini nnsp_size = 0; 2283674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2284674ae819SStefano Zampini } 2285984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2286984c4197SStefano Zampini max_constraints = nnsp_size; 2287984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2288984c4197SStefano Zampini 2289674ae819SStefano Zampini /* 2290674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 22919162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 22929162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 22939162d606SStefano Zampini There can be multiple constraints per connected component 2294674ae819SStefano Zampini */ 2295674ae819SStefano Zampini n_vertices = 0; 2296674ae819SStefano Zampini if (ISForVertices) { 2297674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2298674ae819SStefano Zampini } 22999162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 23009162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 23019162d606SStefano Zampini 23029162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 23039162d606SStefano Zampini total_counts *= max_constraints; 2304674ae819SStefano Zampini total_counts += n_vertices; 23054641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 23069162d606SStefano Zampini 2307674ae819SStefano Zampini total_counts = 0; 2308674ae819SStefano Zampini max_size_of_constraint = 0; 2309674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 23109162d606SStefano Zampini IS used_is; 2311674ae819SStefano Zampini if (i<n_ISForEdges) { 23129162d606SStefano Zampini used_is = ISForEdges[i]; 2313674ae819SStefano Zampini } else { 23149162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2315674ae819SStefano Zampini } 23169162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2317674ae819SStefano Zampini total_counts += j; 2318674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2319674ae819SStefano Zampini } 23209162d606SStefano 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); 23219162d606SStefano Zampini 2322984c4197SStefano Zampini /* get local part of global near null space vectors */ 2323785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2324984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2325984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2326e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2327e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2328984c4197SStefano Zampini } 2329674ae819SStefano Zampini 2330242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2331242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2332a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2333242a89d7SStefano Zampini 2334984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2335a773dcb8SStefano Zampini if (!skip_lapack) { 2336674ae819SStefano Zampini PetscScalar temp_work; 2337911cabfeSStefano Zampini 2338674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2339984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2340785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2341785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2342785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2343674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2344785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2345674ae819SStefano Zampini #endif 2346674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2347c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2348c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2349674ae819SStefano Zampini lwork = -1; 2350674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2351674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2352c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2353674ae819SStefano Zampini #else 2354c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2355674ae819SStefano Zampini #endif 2356674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2357984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2358674ae819SStefano Zampini #else /* on missing GESVD */ 2359674ae819SStefano Zampini /* SVD */ 2360674ae819SStefano Zampini PetscInt max_n,min_n; 2361674ae819SStefano Zampini max_n = max_size_of_constraint; 2362984c4197SStefano Zampini min_n = max_constraints; 2363984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2364674ae819SStefano Zampini min_n = max_size_of_constraint; 2365984c4197SStefano Zampini max_n = max_constraints; 2366674ae819SStefano Zampini } 2367785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2368674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2369785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2370674ae819SStefano Zampini #endif 2371674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2372674ae819SStefano Zampini lwork = -1; 2373e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2374e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2375b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2376674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2377674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23789162d606SStefano 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)); 2379674ae819SStefano Zampini #else 23809162d606SStefano 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)); 2381674ae819SStefano Zampini #endif 2382674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2383984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2384984c4197SStefano Zampini #endif /* on missing GESVD */ 2385674ae819SStefano Zampini /* Allocate optimal workspace */ 2386674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2387854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2388674ae819SStefano Zampini } 2389674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2390674ae819SStefano Zampini total_counts = 0; 23919162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 23929162d606SStefano Zampini constraints_data_ptr[0] = 0; 2393674ae819SStefano Zampini /* vertices */ 23949162d606SStefano Zampini if (n_vertices) { 2395674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 23969162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 23979162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2398674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 23999162d606SStefano Zampini constraints_n[total_counts] = 1; 24009162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24019162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24029162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2403674ae819SStefano Zampini total_counts++; 2404674ae819SStefano Zampini } 2405674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2406984c4197SStefano Zampini PetscBool used_vertex; 2407674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2408674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2409674ae819SStefano Zampini k = 0; 2410674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2411984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2412984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 24139162d606SStefano Zampini constraints_n[total_counts] = 1; 24149162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 24159162d606SStefano Zampini constraints_data[total_counts] = 1.0; 24169162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 24179162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2418674ae819SStefano Zampini total_counts++; 2419674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2420674ae819SStefano Zampini } 2421984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2422674ae819SStefano Zampini k++; 2423674ae819SStefano Zampini } 2424674ae819SStefano Zampini } 2425674ae819SStefano Zampini } 2426674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2427674ae819SStefano Zampini n_vertices = total_counts; 2428674ae819SStefano Zampini } 2429984c4197SStefano Zampini 2430674ae819SStefano Zampini /* edges and faces */ 24319162d606SStefano Zampini total_counts_cc = total_counts; 2432911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 24339162d606SStefano Zampini IS used_is; 24349162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 24359162d606SStefano Zampini 2436911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 24379162d606SStefano Zampini used_is = ISForEdges[ncc]; 2438984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2439674ae819SStefano Zampini } else { 24409162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2441984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2442674ae819SStefano Zampini } 2443674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 24449162d606SStefano Zampini 24459162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 24469162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2447984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2448984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2449674ae819SStefano Zampini if (nnsp_has_cnst) { 24505b08dc53SStefano Zampini PetscScalar quad_value; 24519162d606SStefano Zampini 24529162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24539162d606SStefano Zampini idxs_copied = PETSC_TRUE; 24549162d606SStefano Zampini 2455a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2456674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2457a773dcb8SStefano Zampini } else { 2458a773dcb8SStefano Zampini quad_value = 1.0; 2459a773dcb8SStefano Zampini } 2460674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24619162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2462674ae819SStefano Zampini } 24639162d606SStefano Zampini temp_constraints++; 2464674ae819SStefano Zampini total_counts++; 2465674ae819SStefano Zampini } 2466674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2467984c4197SStefano Zampini PetscReal real_value; 24689162d606SStefano Zampini PetscScalar *ptr_to_data; 24699162d606SStefano Zampini 2470984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 24719162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2472674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 24739162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2474674ae819SStefano Zampini } 2475984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2476984c4197SStefano Zampini /* check if array is null on the connected component */ 2477e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24789162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 24795b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2480674ae819SStefano Zampini temp_constraints++; 2481674ae819SStefano Zampini total_counts++; 24829162d606SStefano Zampini if (!idxs_copied) { 24839162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 24849162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2485674ae819SStefano Zampini } 2486674ae819SStefano Zampini } 24879162d606SStefano Zampini } 24889162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 248945a1bb75SStefano Zampini valid_constraints = temp_constraints; 2490eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2491a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 24929162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 24939162d606SStefano Zampini 24949162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2495a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 24969162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2497a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 24989162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2499a773dcb8SStefano Zampini } else { /* perform SVD */ 2500984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 25019162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2502674ae819SStefano Zampini 2503674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2504984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2505984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2506984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2507984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2508984c4197SStefano Zampini from that computed using LAPACKgesvd 2509984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2510984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2511984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2512674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2513e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2514984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2515674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2516674ae819SStefano Zampini for (k=0;k<j+1;k++) { 25179162d606SStefano 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)); 2518674ae819SStefano Zampini } 2519674ae819SStefano Zampini } 2520e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2521e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2522e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2523674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2524c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2525674ae819SStefano Zampini #else 2526c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2527674ae819SStefano Zampini #endif 2528674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2529984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2530984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2531674ae819SStefano Zampini j = 0; 2532984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2533674ae819SStefano Zampini total_counts = total_counts-j; 253445a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2535e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2536c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2537c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2538c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2539c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2540c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2541c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2542674ae819SStefano Zampini if (j<temp_constraints) { 2543984c4197SStefano Zampini PetscInt ii; 2544984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2545674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25469162d606SStefano 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)); 2547674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2548984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2549674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 25509162d606SStefano 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]; 2551674ae819SStefano Zampini } 2552674ae819SStefano Zampini } 2553674ae819SStefano Zampini } 2554674ae819SStefano Zampini #else /* on missing GESVD */ 2555e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2556e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2557b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2558674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2559674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25609162d606SStefano 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)); 2561674ae819SStefano Zampini #else 25629162d606SStefano 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)); 2563674ae819SStefano Zampini #endif 2564984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2565674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2566984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2567e310c8b4SStefano Zampini k = temp_constraints; 2568e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2569674ae819SStefano Zampini j = 0; 2570e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 257145a1bb75SStefano Zampini valid_constraints = k-j; 2572911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2573984c4197SStefano Zampini #endif /* on missing GESVD */ 2574674ae819SStefano Zampini } 2575a773dcb8SStefano Zampini } 25769162d606SStefano Zampini /* update pointers information */ 25779162d606SStefano Zampini if (valid_constraints) { 25789162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 25799162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 25809162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 25819162d606SStefano Zampini /* set change_of_basis flag */ 258245a1bb75SStefano Zampini if (boolforchange) { 2583b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 25849162d606SStefano Zampini } 2585b3d85658SStefano Zampini total_counts_cc++; 258645a1bb75SStefano Zampini } 258745a1bb75SStefano Zampini } 2588984c4197SStefano Zampini /* free workspace */ 25898f1c130eSStefano Zampini if (!skip_lapack) { 2590984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2591984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2592984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2593984c4197SStefano Zampini #endif 2594984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2595984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2596984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2597984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2598984c4197SStefano Zampini #endif 2599984c4197SStefano Zampini } 2600984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2601984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2602984c4197SStefano Zampini } 2603984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2604cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2605cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2606cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2607cf5a6209SStefano Zampini } 2608cf5a6209SStefano Zampini if (n_ISForFaces) { 2609cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2610cf5a6209SStefano Zampini } 2611cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2612cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2613cf5a6209SStefano Zampini } 2614cf5a6209SStefano Zampini if (n_ISForEdges) { 2615cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2616cf5a6209SStefano Zampini } 2617cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 261808122e43SStefano Zampini } else { 261908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2620984c4197SStefano Zampini 262108122e43SStefano Zampini total_counts = 0; 262208122e43SStefano Zampini n_vertices = 0; 2623d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2624d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 262508122e43SStefano Zampini } 262608122e43SStefano Zampini max_constraints = 0; 26279162d606SStefano Zampini total_counts_cc = 0; 262808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 262908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 26309162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 263108122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 263208122e43SStefano Zampini } 26339162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 26349162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 26359162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 26369162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 263774d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 26389162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 26399162d606SStefano Zampini total_counts_cc = 0; 26409162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 26419162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 26429162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 264308122e43SStefano Zampini } 264408122e43SStefano Zampini } 26459162d606SStefano Zampini #if 0 26469162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 26479162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26489162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 26499162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 26509162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 26519162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 26529162d606SStefano Zampini } 26539162d606SStefano Zampini printf("\n"); 26549162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 26559162d606SStefano Zampini } 26561b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 26578bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 26581b968477SStefano Zampini } 26591b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 26608bec7fa6SStefano 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]); 26611b968477SStefano Zampini } 266208122e43SStefano Zampini #endif 266308122e43SStefano Zampini 26648bec7fa6SStefano Zampini max_size_of_constraint = 0; 26659162d606SStefano 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]); 26669162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 266708122e43SStefano Zampini /* Change of basis */ 2668b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 266908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 267008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 267108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2672b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 267308122e43SStefano Zampini } 267408122e43SStefano Zampini } 267508122e43SStefano Zampini } 267608122e43SStefano Zampini } 2677984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 2678019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2679019a44ceSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+1,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 268008122e43SStefano Zampini 26819162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 26829162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 26839162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 26849162d606SStefano 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); 268508122e43SStefano Zampini } 2686674ae819SStefano Zampini 2687674ae819SStefano Zampini /* Create constraint matrix */ 2688674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 268916f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2690984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2691984c4197SStefano Zampini 2692984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 2693a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 2694a717540cSStefano Zampini qr_needed = PETSC_FALSE; 269574d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2696984c4197SStefano Zampini total_primal_vertices=0; 2697b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 26989162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 26999162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2700984c4197SStefano Zampini if (size_of_constraint == 1) { 27019162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2702b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 270364efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 27049162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27059162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 2706a717540cSStefano Zampini } 2707b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 270874d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 2709a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 2710a717540cSStefano Zampini qr_needed = PETSC_TRUE; 2711a717540cSStefano Zampini } 2712fa434743SStefano Zampini } else { 2713b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2714fa434743SStefano Zampini } 2715a717540cSStefano Zampini } 2716b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2717b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 2718674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 271970022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2720b3d85658SStefano Zampini 2721019a44ceSStefano Zampini /* allocating one extra space (in case an extra primal dof should be stored for the benign trick */ 2722019a44ceSStefano 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); 27230e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 27240e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 2725984c4197SStefano Zampini 2726984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 272774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2728785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2729984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 273074d5cdf7SStefano Zampini 2731984c4197SStefano Zampini j = total_primal_vertices; 273274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2733b3d85658SStefano Zampini cum = total_primal_vertices; 27349162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27354641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2736b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2737b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2738b3d85658SStefano Zampini cum++; 27399162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 274074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 274174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 274274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 274374d5cdf7SStefano Zampini } 27449162d606SStefano Zampini j += constraints_n[i]; 2745674ae819SStefano Zampini } 2746674ae819SStefano Zampini } 2747674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2748674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2749088faed8SStefano Zampini 2750674ae819SStefano Zampini /* set values in constraint matrix */ 2751984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 27520e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2753674ae819SStefano Zampini } 2754984c4197SStefano Zampini total_counts = total_primal_vertices; 27559162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 27564641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 27579162d606SStefano Zampini PetscInt *cols; 27589162d606SStefano Zampini 27599162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 27609162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 27619162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 27629162d606SStefano Zampini PetscInt row = total_counts+k; 27639162d606SStefano Zampini PetscScalar *vals; 27649162d606SStefano Zampini 27659162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 27669162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 27679162d606SStefano Zampini } 27689162d606SStefano Zampini total_counts += constraints_n[i]; 2769674ae819SStefano Zampini } 2770674ae819SStefano Zampini } 2771674ae819SStefano Zampini /* assembling */ 2772674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2773674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2774088faed8SStefano Zampini 2775984c4197SStefano Zampini /* 277645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2777984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2778984c4197SStefano Zampini */ 2779674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2780674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2781026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2782984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2783984c4197SStefano Zampini /* working stuff for GEQRF */ 278481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2785984c4197SStefano Zampini PetscBLASInt lqr_work; 2786984c4197SStefano Zampini /* working stuff for UNGQR */ 2787984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2788984c4197SStefano Zampini PetscBLASInt lgqr_work; 2789984c4197SStefano Zampini /* working stuff for TRTRS */ 2790984c4197SStefano Zampini PetscScalar *trs_rhs; 27913f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2792984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2793984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2794984c4197SStefano Zampini PetscScalar *start_vals; 2795984c4197SStefano Zampini /* working stuff for values insertion */ 27964641a718SStefano Zampini PetscBT is_primal; 279764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2798906d46d4SStefano Zampini /* matrix sizes */ 2799906d46d4SStefano Zampini PetscInt global_size,local_size; 2800906d46d4SStefano Zampini /* temporary change of basis */ 2801906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2802cf5a6209SStefano Zampini /* extra space for debugging */ 2803cf5a6209SStefano Zampini PetscScalar *dbg_work; 2804984c4197SStefano Zampini 2805906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2806906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 280716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2808bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2809906d46d4SStefano Zampini /* nonzeros for local mat */ 2810bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2811bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 28129162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2813a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 28149162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2815a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 28169162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2817a717540cSStefano Zampini } else { 28189162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 28199162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2820a717540cSStefano Zampini } 2821a717540cSStefano Zampini } 2822a717540cSStefano Zampini } 2823906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2824bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2825a717540cSStefano Zampini /* Set initial identity in the matrix */ 2826bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2827906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2828a717540cSStefano Zampini } 2829a717540cSStefano Zampini 2830a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2831a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2832a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2833a717540cSStefano Zampini } 2834a717540cSStefano Zampini 2835a717540cSStefano Zampini 2836a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2837a717540cSStefano Zampini /* 2838a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2839a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2840a717540cSStefano Zampini 2841a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2842a717540cSStefano Zampini 2843a6b551f4SStefano 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) 2844a6b551f4SStefano Zampini 2845a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2846a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2847a717540cSStefano Zampini | ... | 2848a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2849a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2850a717540cSStefano Zampini 2851a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2852a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2853a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2854a6b551f4SStefano Zampini 2855a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2856a717540cSStefano Zampini */ 2857a717540cSStefano Zampini if (qr_needed) { 2858984c4197SStefano Zampini /* space to store Q */ 2859854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2860984c4197SStefano Zampini /* first we issue queries for optimal work */ 28613f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28623f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 28633f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2864984c4197SStefano Zampini lqr_work = -1; 28653f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2866984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2867984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2868785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2869984c4197SStefano Zampini lgqr_work = -1; 28703f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 28713f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 28723f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 28733f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 28743f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 28753f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2876984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2877984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2878785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2879984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2880785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2881984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2882785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2883a717540cSStefano Zampini /* allocating workspace for check */ 2884a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2885cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2886a717540cSStefano Zampini } 2887a717540cSStefano Zampini } 2888984c4197SStefano Zampini /* array to store whether a node is primal or not */ 28894641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2890473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 28910e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 289239e2fb2aSStefano Zampini if (i != total_primal_vertices) { 289339e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 28944641a718SStefano Zampini } 289539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 289639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 289739e2fb2aSStefano Zampini } 289839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2899984c4197SStefano Zampini 2900a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 29019162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 29029162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 29034641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2904984c4197SStefano Zampini /* get constraint info */ 29059162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2906984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2907984c4197SStefano Zampini 2908984c4197SStefano Zampini if (pcbddc->dbg_flag) { 29099162d606SStefano 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); 2910674ae819SStefano Zampini } 2911984c4197SStefano Zampini 2912fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2913a717540cSStefano Zampini 2914a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2915a717540cSStefano Zampini if (pcbddc->dbg_flag) { 29169162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2917a717540cSStefano Zampini } 2918984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 29199162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2920984c4197SStefano Zampini 2921984c4197SStefano Zampini /* compute QR decomposition of constraints */ 29223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2925674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29263f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2927984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2928674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2929984c4197SStefano Zampini 2930984c4197SStefano Zampini /* explictly compute R^-T */ 2931984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2932984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 29333f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29343f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 29353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2937984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29383f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2939984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2940984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2941984c4197SStefano Zampini 2942a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 29433f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 29453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2947984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29483f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2949984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2950984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2951984c4197SStefano Zampini 2952984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2953984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2954984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 29553f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 29563f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 29573f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 29583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 29593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 29603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2961984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 29629162d606SStefano 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)); 2963984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 29649162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2965984c4197SStefano Zampini 2966984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 29679162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2968984c4197SStefano Zampini /* insert cols for primal dofs */ 2969984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2970984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 29719162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2972906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2973984c4197SStefano Zampini } 2974984c4197SStefano Zampini /* insert cols for dual dofs */ 2975984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 29769162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2977984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 29789162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2979906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2980984c4197SStefano Zampini j++; 2981674ae819SStefano Zampini } 2982674ae819SStefano Zampini } 2983984c4197SStefano Zampini 2984984c4197SStefano Zampini /* check change of basis */ 2985984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2986984c4197SStefano Zampini PetscInt ii,jj; 2987984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2988c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2989c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2990c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2991c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2992c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2993c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2994984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2995cf5a6209SStefano 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)); 2996984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2997984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2998984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2999cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3000cf5a6209SStefano 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; 3001674ae819SStefano Zampini } 3002674ae819SStefano Zampini } 3003984c4197SStefano Zampini if (!valid_qr) { 300422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3005984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3006984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3007cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3008cf5a6209SStefano 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])); 3009674ae819SStefano Zampini } 3010cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3011cf5a6209SStefano 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])); 3012984c4197SStefano Zampini } 3013984c4197SStefano Zampini } 3014984c4197SStefano Zampini } 3015674ae819SStefano Zampini } else { 301622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3017674ae819SStefano Zampini } 3018674ae819SStefano Zampini } 3019a717540cSStefano Zampini } else { /* simple transformation block */ 3020a717540cSStefano Zampini PetscInt row,col; 3021a6b551f4SStefano Zampini PetscScalar val,norm; 3022a6b551f4SStefano Zampini 3023a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 30249162d606SStefano 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)); 3025a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 30269162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 30279162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3028bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 30299162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3030906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 30319162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3032a717540cSStefano Zampini } else { 3033a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 30349162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3035a717540cSStefano Zampini if (row != col) { 30369162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3037a717540cSStefano Zampini } else { 30389162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3039a717540cSStefano Zampini } 3040906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3041a717540cSStefano Zampini } 3042a717540cSStefano Zampini } 3043a717540cSStefano Zampini } 304498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 304522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3046a717540cSStefano Zampini } 3047674ae819SStefano Zampini } 3048984c4197SStefano Zampini } else { 3049984c4197SStefano Zampini if (pcbddc->dbg_flag) { 30509162d606SStefano 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); 3051674ae819SStefano Zampini } 3052674ae819SStefano Zampini } 3053674ae819SStefano Zampini } 3054a717540cSStefano Zampini 3055a717540cSStefano Zampini /* free workspace */ 3056a717540cSStefano Zampini if (qr_needed) { 3057984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3058cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 3059984c4197SStefano Zampini } 3060984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 3061984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 3062984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 3063984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 3064984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 3065674ae819SStefano Zampini } 3066a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 3067906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3068906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3069906d46d4SStefano Zampini 3070906d46d4SStefano Zampini /* assembling of global change of variable */ 3071bbb9e6c6SStefano Zampini { 3072bbb9e6c6SStefano Zampini Mat tmat; 307316f15bc4SStefano Zampini PetscInt bs; 307416f15bc4SStefano Zampini 3075906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3076906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3077bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 3078bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 3079bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3080bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 308116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 308216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 3083906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3084bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 3085bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3086bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3087bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 3088bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 3089e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3090e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3091bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 3092bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 3093906d46d4SStefano Zampini } 3094906d46d4SStefano Zampini /* check */ 3095906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 3096906d46d4SStefano Zampini PetscReal error; 3097906d46d4SStefano Zampini Vec x,x_change; 3098906d46d4SStefano Zampini 3099906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 3100906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 3101906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 3102906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 3103e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3104e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3105bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 3106e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3107e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3108906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 3109906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 3110906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 3111906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3112bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 3113906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 3114906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 3115906d46d4SStefano Zampini } 3116b96c3477SStefano Zampini 3117b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 3118b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 3119b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 3120b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 3121ac632422SStefano Zampini Mat S_new,tmat; 3122b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 3123b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 3124b087196eSStefano Zampini const PetscScalar *array; 3125b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 3126b087196eSStefano Zampini PetscInt i,n_V; 3127bbb9e6c6SStefano Zampini 3128bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 31296816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 3130b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 3131b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 3132b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 3133b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 3134bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 3135b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 3136ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3137b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 3138ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3139b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3140b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 3141b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3142b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3143b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 3144b087196eSStefano Zampini for (i=0;i<n_V;i++) { 3145b087196eSStefano Zampini PetscScalar val; 3146b087196eSStefano Zampini PetscInt idx; 3147b087196eSStefano Zampini 3148b087196eSStefano Zampini idx = idxs_V[i]; 3149b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 3150b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 3151b087196eSStefano Zampini } 3152b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3153b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3154ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 3155ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3156ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 3157ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 3158b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 3159ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 3160b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 3161ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 3162ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 3163ac632422SStefano Zampini } 3164b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 3165b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 3166b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 3167b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 3168b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 3169b96c3477SStefano Zampini } 3170b96c3477SStefano Zampini } 3171906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 3172906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 3173b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3174b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 3175b9b85e73SStefano Zampini } 3176906d46d4SStefano Zampini 3177906d46d4SStefano Zampini /* set up change of basis context */ 3178906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 3179906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 3180906d46d4SStefano Zampini 3181906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 3182906d46d4SStefano Zampini PetscInt global_size,local_size; 3183906d46d4SStefano Zampini 3184906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 3185906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 3186906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 3187906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 3188906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 3189906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 3190906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 3191906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 3192906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 3193906d46d4SStefano Zampini } else { 3194906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 3195906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 3196906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 3197906d46d4SStefano Zampini } 3198906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 3199906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 3200906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 3201906d46d4SStefano Zampini } else { 3202906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 3203906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 3204906d46d4SStefano Zampini } 3205906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 3206906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 3207906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3208906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3209b9b85e73SStefano Zampini } 3210a717540cSStefano Zampini 3211019a44ceSStefano Zampini /* add pressure dof to set of primal nodes for numbering purposes */ 3212c9ed8603SStefano Zampini if (pcbddc->benign_p0_lidx >= 0) { 3213c9ed8603SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx; 3214*d16cbb6bSStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx; 3215019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 3216019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 3217019a44ceSStefano Zampini pcbddc->local_primal_size++; 3218019a44ceSStefano Zampini } 3219019a44ceSStefano Zampini 3220019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 3221727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 3222727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 3223aff50787SStefano 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); 3224c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 32250e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 3226aff50787SStefano 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); 3227727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 3228727cdba6SStefano Zampini } 32290e6343abSStefano Zampini } 32300e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 3231727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 3232727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3233727cdba6SStefano Zampini 3234a717540cSStefano Zampini /* flush dbg viewer */ 3235b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 3236b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3237b8ffe317SStefano Zampini } 3238a717540cSStefano Zampini 3239e310c8b4SStefano Zampini /* free workspace */ 3240a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 32414641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 324208122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 32439162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 32449162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 324508122e43SStefano Zampini } else { 32469162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 32479162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 32489162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 324908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 325008122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 32519162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 32529162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 325308122e43SStefano Zampini } 3254674ae819SStefano Zampini PetscFunctionReturn(0); 3255674ae819SStefano Zampini } 3256674ae819SStefano Zampini 3257674ae819SStefano Zampini #undef __FUNCT__ 3258674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3259674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3260674ae819SStefano Zampini { 3261674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3262674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3263674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 32647fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3265674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3266674ae819SStefano Zampini 3267674ae819SStefano Zampini PetscFunctionBegin; 32688e61c736SStefano Zampini /* Reset previously computed graph */ 32698e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3270674ae819SStefano Zampini /* Init local Graph struct */ 32717fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 32723bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 3273674ae819SStefano Zampini 3274575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3275575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3276575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3277575ad6abSStefano Zampini } 32789577ea80SStefano Zampini 3279674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 32804d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 32814d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 32824d379d7bSStefano Zampini PetscInt nvtxs; 3283e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3284674ae819SStefano Zampini 32854d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 32862fffb893SStefano Zampini 32872fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32882fffb893SStefano Zampini if (flg_row) { 32894d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3290b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 32912fffb893SStefano Zampini } 32922fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 32939b28b941SStefano 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 */ 32944d379d7bSStefano Zampini IS is_dummy; 32954d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 32964d379d7bSStefano Zampini PetscInt j,sum; 32974d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 32984d379d7bSStefano Zampini const PetscInt *idxs; 32994d379d7bSStefano Zampini PCBDDCGraph graph; 33004d379d7bSStefano Zampini PetscBT is_on_boundary; 33014d379d7bSStefano Zampini 33024d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 33034d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 33044d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 33054d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 33067fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 33074d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3308e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3309e496cd5dSStefano Zampini if (flg_row) { 33104d379d7bSStefano Zampini graph->xadj = xadj; 33114d379d7bSStefano Zampini graph->adjncy = adjncy; 3312e496cd5dSStefano Zampini } 33134d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 33144d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3315e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 33164d379d7bSStefano Zampini 33174d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 33189b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 33194d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33204d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 33214d379d7bSStefano Zampini } 33224d379d7bSStefano Zampini } 33234d379d7bSStefano Zampini 3324e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 33254d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 33264d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 33274d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 33284d379d7bSStefano Zampini } 33294d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 33304d379d7bSStefano Zampini 3331e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 33324d379d7bSStefano Zampini sum = 0; 33334d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33344d379d7bSStefano Zampini PetscInt sizecc = 0; 33354d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33364d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33374d379d7bSStefano Zampini sizecc++; 33384d379d7bSStefano Zampini } 33394d379d7bSStefano Zampini } 33404d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33414d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33424d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 33434d379d7bSStefano Zampini } 33444d379d7bSStefano Zampini } 33454d379d7bSStefano Zampini sum += sizecc*sizecc; 33464d379d7bSStefano Zampini } 33474d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 33484d379d7bSStefano Zampini sum = 0; 3349e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 33504d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 33514d379d7bSStefano Zampini cxadj[i] = sum; 33524d379d7bSStefano Zampini sum += temp; 33534d379d7bSStefano Zampini } 3354e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 33554d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 33564d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 33574d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 33584d379d7bSStefano Zampini PetscInt k,sizecc = 0; 33594d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 33604d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 33614d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 33624d379d7bSStefano Zampini sizecc++; 33634d379d7bSStefano Zampini } 33644d379d7bSStefano Zampini } 33654d379d7bSStefano Zampini } 33664d379d7bSStefano Zampini } 33674d379d7bSStefano Zampini } 33689b28b941SStefano Zampini if (sum) { 3369e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 33704d379d7bSStefano Zampini } else { 33714d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 33724d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 33734d379d7bSStefano Zampini } 33744d379d7bSStefano Zampini graph->xadj = 0; 33754d379d7bSStefano Zampini graph->adjncy = 0; 33764d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 33774d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 33784d379d7bSStefano Zampini } 3379674ae819SStefano Zampini } 33809b28b941SStefano Zampini if (pcbddc->dbg_flag) { 33819b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3382674ae819SStefano Zampini } 3383674ae819SStefano Zampini 338463602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3385674ae819SStefano Zampini vertex_size = 1; 338663602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 338763602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 338895ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 338963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3390e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 339163602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3392674ae819SStefano Zampini } 339363602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 339463602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 339563602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3396674ae819SStefano Zampini } 339763602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3398674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 339963602bcaSStefano Zampini } else { 340063602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 340163602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3402854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 340363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 340463602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 340563602bcaSStefano Zampini } 340663602bcaSStefano Zampini } 3407674ae819SStefano Zampini } 3408674ae819SStefano Zampini 3409674ae819SStefano Zampini /* Setup of Graph */ 3410785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3411e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3412785d1243SStefano Zampini } 3413785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3414e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3415785d1243SStefano Zampini } 3416302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3417674ae819SStefano Zampini 3418674ae819SStefano Zampini /* Graph's connected components analysis */ 3419674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3420674ae819SStefano Zampini 3421674ae819SStefano Zampini /* print some info to stdout */ 3422674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3423302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3424674ae819SStefano Zampini } 3425fb180af4SStefano Zampini 3426fb180af4SStefano Zampini /* mark topography has done */ 3427fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3428674ae819SStefano Zampini PetscFunctionReturn(0); 3429674ae819SStefano Zampini } 3430674ae819SStefano Zampini 3431dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3432674ae819SStefano Zampini #undef __FUNCT__ 3433674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3434dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3435674ae819SStefano Zampini { 3436dc456d91SStefano Zampini PetscSF sf; 3437dc456d91SStefano Zampini PetscLayout map; 3438dc456d91SStefano Zampini const PetscInt *idxs; 3439dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3440dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3441dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3442dc456d91SStefano Zampini PetscMPIInt commsize; 3443674ae819SStefano Zampini PetscBool first_found; 3444674ae819SStefano Zampini PetscErrorCode ierr; 3445674ae819SStefano Zampini 3446674ae819SStefano Zampini PetscFunctionBegin; 3447dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3448dc456d91SStefano Zampini if (subset_mult) { 3449dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3450dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3451dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3452674ae819SStefano Zampini } 3453dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3454dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3455dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3456dc456d91SStefano Zampini for (i=0;i<n;i++) { 3457dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3458dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3459674ae819SStefano Zampini } 3460dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3461dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3462dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3463dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3464dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3465dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3466dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3467dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3468dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3469dc456d91SStefano Zampini 3470dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3471dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3472dc456d91SStefano Zampini if (subset_mult) { 3473dc456d91SStefano Zampini const PetscInt* idxs_mult; 3474dc456d91SStefano Zampini 3475dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3476dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3477dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3478674ae819SStefano Zampini } else { 3479dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3480674ae819SStefano Zampini } 3481dc456d91SStefano Zampini /* local size of new subset */ 3482dc456d91SStefano Zampini n_n = 0; 3483dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3484dc456d91SStefano Zampini 3485dc456d91SStefano Zampini /* global indexes in layout */ 3486dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3487dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3488dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3489dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3490dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3491dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3492dc456d91SStefano Zampini 3493dc456d91SStefano Zampini /* reduce from leaves to roots */ 3494dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 349564a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 349664a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 3497dc456d91SStefano Zampini 3498dc456d91SStefano Zampini /* count indexes in local part of layout */ 3499674ae819SStefano Zampini nlocals = 0; 3500674ae819SStefano Zampini first_index = -1; 3501674ae819SStefano Zampini first_found = PETSC_FALSE; 3502dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3503dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3504674ae819SStefano Zampini first_found = PETSC_TRUE; 3505674ae819SStefano Zampini first_index = i; 3506674ae819SStefano Zampini } 3507dc456d91SStefano Zampini nlocals += root_data[i]; 3508674ae819SStefano Zampini } 3509dc456d91SStefano Zampini 3510dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 35115fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 3512dc456d91SStefano Zampini start = 0; 351364a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 35145fa240b1SStefano Zampini #else 351564a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 35165fa240b1SStefano Zampini start = start-nlocals; 35175fa240b1SStefano Zampini #endif 35185fa240b1SStefano Zampini 3519dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3520dc456d91SStefano Zampini *N_n = start + nlocals; 3521dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3522dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3523674ae819SStefano Zampini } 35245fa240b1SStefano Zampini 35255fa240b1SStefano Zampini /* adapt root data with cumulative */ 3526674ae819SStefano Zampini if (first_found) { 3527dc456d91SStefano Zampini PetscInt old_index; 3528dc456d91SStefano Zampini 3529dc456d91SStefano Zampini root_data[first_index] += start; 3530674ae819SStefano Zampini old_index = first_index; 3531dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3532dc456d91SStefano Zampini if (root_data[i]) { 3533dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3534674ae819SStefano Zampini old_index = i; 3535674ae819SStefano Zampini } 3536674ae819SStefano Zampini } 3537674ae819SStefano Zampini } 3538dc456d91SStefano Zampini 3539dc456d91SStefano Zampini /* from roots to leaves */ 3540dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3541dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3542dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3543dc456d91SStefano Zampini 3544dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3545dc456d91SStefano Zampini if (subset_mult) { 3546dc456d91SStefano Zampini const PetscInt* idxs_mult; 3547dc456d91SStefano Zampini PetscInt cum; 3548dc456d91SStefano Zampini 3549dc456d91SStefano Zampini cum = 0; 3550dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3551dc456d91SStefano Zampini for (i=0;i<n;i++) { 3552dc456d91SStefano Zampini PetscInt j; 3553dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3554674ae819SStefano Zampini } 3555dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3556674ae819SStefano Zampini } else { 3557dc456d91SStefano Zampini for (i=0;i<n;i++) { 3558dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3559674ae819SStefano Zampini } 3560674ae819SStefano Zampini } 3561dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3562dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3563674ae819SStefano Zampini PetscFunctionReturn(0); 3564674ae819SStefano Zampini } 35659a7d3425SStefano Zampini 35669a7d3425SStefano Zampini #undef __FUNCT__ 35679a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 35689a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 35699a7d3425SStefano Zampini { 35709a7d3425SStefano Zampini PetscInt i,j; 35719a7d3425SStefano Zampini PetscScalar *alphas; 35729a7d3425SStefano Zampini PetscErrorCode ierr; 35739a7d3425SStefano Zampini 35749a7d3425SStefano Zampini PetscFunctionBegin; 35759a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3576785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 35779a7d3425SStefano Zampini for (i=0;i<n;i++) { 35789a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 35799a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 35809a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 35819a7d3425SStefano Zampini } 35829a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 35839a7d3425SStefano Zampini PetscFunctionReturn(0); 35849a7d3425SStefano Zampini } 35859a7d3425SStefano Zampini 3586e7931f94SStefano Zampini #undef __FUNCT__ 358770cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3588b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3589e7931f94SStefano Zampini { 359052e5ac9dSStefano Zampini IS ranks_send_to; 3591e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3592e7931f94SStefano Zampini PetscMPIInt size,rank,color; 359352e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 359452e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 35953837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 35962b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3597e7931f94SStefano Zampini PetscSubcomm subcomm; 359852e5ac9dSStefano Zampini PetscErrorCode ierr; 3599a57a6d2fSStefano Zampini 3600e7931f94SStefano Zampini PetscFunctionBegin; 36012b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 36022b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 36032b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3604e7931f94SStefano Zampini 3605e7931f94SStefano Zampini /* Get info on mapping */ 36063bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 36073bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3608e7931f94SStefano Zampini 3609e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3610785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3611e7931f94SStefano Zampini xadj[0] = 0; 3612e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3613785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3614785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3615e7931f94SStefano Zampini 36162b510759SStefano Zampini if (threshold) { 3617d023bfaeSStefano Zampini PetscInt xadj_count = 0; 36182b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3619d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3620d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3621d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3622d023bfaeSStefano Zampini xadj_count++; 3623e7931f94SStefano Zampini } 3624e7931f94SStefano Zampini } 3625d023bfaeSStefano Zampini xadj[1] = xadj_count; 3626c8587f34SStefano Zampini } else { 3627e7931f94SStefano Zampini if (xadj[1]) { 3628e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3629e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3630c8587f34SStefano Zampini } 3631e7931f94SStefano Zampini } 36323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3633e7931f94SStefano Zampini if (use_square) { 3634e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3635e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3636e7931f94SStefano Zampini } 3637e7931f94SStefano Zampini } 3638e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3639e7931f94SStefano Zampini 36403837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3641e7931f94SStefano Zampini 3642e7931f94SStefano Zampini /* 3643e7931f94SStefano Zampini Restrict work on active processes only. 3644e7931f94SStefano Zampini */ 3645e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3646e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3647e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 36482b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3649d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3650e7931f94SStefano Zampini if (color) { 3651e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3652e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3653e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3654c8587f34SStefano Zampini } else { 365552e5ac9dSStefano Zampini Mat subdomain_adj; 365652e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 365752e5ac9dSStefano Zampini MatPartitioning partitioner; 365852e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 365952e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 3660b0c7d250SStefano Zampini PetscBool aggregate; 3661b0c7d250SStefano Zampini 3662306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3663785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3664e7931f94SStefano Zampini prank = rank; 3665306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 36668002ef2cSStefano Zampini /* 3667e7931f94SStefano Zampini for (i=0;i<size;i++) { 3668e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3669c8587f34SStefano Zampini } 36708002ef2cSStefano Zampini */ 3671e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3672e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3673c8587f34SStefano Zampini } 3674e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3675b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3676b0c7d250SStefano Zampini if (aggregate) { 3677b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3678b0c7d250SStefano Zampini PetscMPIInt nrank; 3679b0c7d250SStefano Zampini PetscScalar *vals; 3680b0c7d250SStefano Zampini 3681b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3682b0c7d250SStefano Zampini lrows = 0; 3683b0c7d250SStefano Zampini if (nrank<redprocs) { 3684b0c7d250SStefano Zampini lrows = size/redprocs; 3685b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3686b0c7d250SStefano Zampini } 36875fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 3688b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3689b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3690b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3691b0c7d250SStefano Zampini row = nrank; 3692b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3693b0c7d250SStefano Zampini cols = adjncy; 3694b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3695b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3696b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3697b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3698b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 369952e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 370052e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 370152e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3702b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3703b0c7d250SStefano Zampini } else { 3704306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3705b0c7d250SStefano Zampini } 370622b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3707e7931f94SStefano Zampini 3708e7931f94SStefano Zampini /* Partition */ 3709306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3710e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3711e7931f94SStefano Zampini if (use_vwgt) { 37123837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3713e7931f94SStefano Zampini v_wgt[0] = local_size; 3714e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3715c8587f34SStefano Zampini } 371628143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 371728143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3718e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3719e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 372022b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3721e7931f94SStefano Zampini 372252e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 372352e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 372452e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 372552e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3726b0c7d250SStefano Zampini if (!redprocs) { 3727b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 372828143c3dSStefano Zampini } else { 3729b0c7d250SStefano Zampini PetscInt idxs[1]; 3730b0c7d250SStefano Zampini PetscMPIInt tag; 3731b0c7d250SStefano Zampini MPI_Request *reqs; 3732b0c7d250SStefano Zampini 3733b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3734b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3735b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3736b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 373728143c3dSStefano Zampini } 3738b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3739b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3740b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3741b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 3742e7931f94SStefano Zampini } 374352e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3744e7931f94SStefano Zampini /* clean up */ 3745e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 374652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 3747e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3748e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3749e7931f94SStefano Zampini } 3750e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3751e7931f94SStefano Zampini 3752e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3753e7931f94SStefano Zampini i = 1; 3754e7931f94SStefano Zampini if (color) i=0; 3755e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3756e7931f94SStefano Zampini /* get back IS */ 3757e7931f94SStefano Zampini *is_sends = ranks_send_to; 3758e7931f94SStefano Zampini PetscFunctionReturn(0); 3759e7931f94SStefano Zampini } 3760e7931f94SStefano Zampini 3761e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3762e7931f94SStefano Zampini 3763e7931f94SStefano Zampini #undef __FUNCT__ 3764e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 376553a05cb3SStefano 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[]) 3766e7931f94SStefano Zampini { 376770cf5478SStefano Zampini Mat local_mat; 3768e7931f94SStefano Zampini IS is_sends_internal; 37699d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 377028143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 37719d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3772e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3773e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3774e7931f94SStefano Zampini const PetscInt* is_indices; 3775e7931f94SStefano Zampini MatType new_local_type; 3776e7931f94SStefano Zampini /* buffers */ 3777e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 377828143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 37799d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3780e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3781e7931f94SStefano Zampini /* MPI */ 378228143c3dSStefano Zampini MPI_Comm comm,comm_n; 378328143c3dSStefano Zampini PetscSubcomm subcomm; 3784e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 378528143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 378628143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 378728143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 378828143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 378928143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3790e7931f94SStefano Zampini PetscErrorCode ierr; 3791e7931f94SStefano Zampini 3792e7931f94SStefano Zampini PetscFunctionBegin; 379328143c3dSStefano Zampini /* TODO: add missing checks */ 379428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 379528143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 379628143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 379728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3798e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 379928143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3800e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3801e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3802e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3803e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3804e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 380528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 380670cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 380770cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 380828143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 380970cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 381070cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 381170cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 381270cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 381370cf5478SStefano Zampini } 3814e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3815e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3816e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3817e7931f94SStefano Zampini if (!is_sends) { 381828143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3819b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3820c8587f34SStefano Zampini } else { 3821e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3822e7931f94SStefano Zampini is_sends_internal = is_sends; 3823c8587f34SStefano Zampini } 3824e7931f94SStefano Zampini 3825e7931f94SStefano Zampini /* get comm */ 3826a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3827e7931f94SStefano Zampini 3828e7931f94SStefano Zampini /* compute number of sends */ 3829e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3830e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3831e7931f94SStefano Zampini 3832e7931f94SStefano Zampini /* compute number of receives */ 3833e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3834785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3835e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3836e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3837e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3838e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3839e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3840e7931f94SStefano Zampini 384128143c3dSStefano Zampini /* restrict comm if requested */ 384228143c3dSStefano Zampini subcomm = 0; 384328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 384428143c3dSStefano Zampini if (restrict_comm) { 3845779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3846779c1cceSStefano Zampini 384728143c3dSStefano Zampini color = 0; 384853a05cb3SStefano Zampini if (restrict_full) { 384953a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 385053a05cb3SStefano Zampini } else { 385153a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 385253a05cb3SStefano Zampini } 385328143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 385428143c3dSStefano Zampini subcommsize = commsize - subcommsize; 385528143c3dSStefano Zampini /* check if reuse has been requested */ 385628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 385728143c3dSStefano Zampini if (*mat_n) { 385828143c3dSStefano Zampini PetscMPIInt subcommsize2; 385928143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 386028143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 386128143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 386228143c3dSStefano Zampini } else { 386328143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 386428143c3dSStefano Zampini } 386528143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3866779c1cceSStefano Zampini PetscMPIInt rank; 3867779c1cceSStefano Zampini 3868779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 386928143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 387028143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 387128143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3872306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 387328143c3dSStefano Zampini } 387428143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 387528143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 387628143c3dSStefano Zampini } else { 387728143c3dSStefano Zampini comm_n = comm; 387828143c3dSStefano Zampini } 387928143c3dSStefano Zampini 3880e7931f94SStefano Zampini /* prepare send/receive buffers */ 3881785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3882e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3883785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3884e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 388528143c3dSStefano Zampini if (nis) { 3886854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 388728143c3dSStefano Zampini } 3888e7931f94SStefano Zampini 388928143c3dSStefano Zampini /* Get data from local matrices */ 3890e7931f94SStefano Zampini if (!isdense) { 3891a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3892e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3893e7931f94SStefano Zampini /* 3894e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3895e7931f94SStefano Zampini send_buffer_idxs should contain: 3896e7931f94SStefano Zampini - MatType_PRIVATE type 3897e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3898e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3899e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3900e7931f94SStefano Zampini */ 3901e7931f94SStefano Zampini } else { 3902e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 39033bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 3904854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3905e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3906e7931f94SStefano Zampini send_buffer_idxs[1] = i; 39073bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 39093bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3910e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3911e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3912e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3913e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3914c8587f34SStefano Zampini } 3915c8587f34SStefano Zampini } 3916e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 391728143c3dSStefano Zampini /* additional is (if any) */ 391828143c3dSStefano Zampini if (nis) { 391928143c3dSStefano Zampini PetscMPIInt psum; 392028143c3dSStefano Zampini PetscInt j; 392128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 392228143c3dSStefano Zampini PetscInt plen; 392328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 392428143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 392528143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 392628143c3dSStefano Zampini } 3927854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 392828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 392928143c3dSStefano Zampini PetscInt plen; 393028143c3dSStefano Zampini const PetscInt *is_array_idxs; 393128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 393228143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 393328143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 393428143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 393528143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 393628143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 393728143c3dSStefano Zampini } 393828143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 393928143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 394028143c3dSStefano Zampini } 394128143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 394228143c3dSStefano Zampini } 394328143c3dSStefano Zampini 3944e7931f94SStefano Zampini buf_size_idxs = 0; 3945e7931f94SStefano Zampini buf_size_vals = 0; 394628143c3dSStefano Zampini buf_size_idxs_is = 0; 3947e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3948e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3949e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 395028143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3951e7931f94SStefano Zampini } 3952785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3953785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 395495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3955e7931f94SStefano Zampini 3956e7931f94SStefano Zampini /* get new tags for clean communications */ 3957e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3958e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 395928143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3960e7931f94SStefano Zampini 3961e7931f94SStefano Zampini /* allocate for requests */ 3962785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3963785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 396495ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3965785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3966785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 396795ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3968e7931f94SStefano Zampini 3969e7931f94SStefano Zampini /* communications */ 3970e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3971e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 397228143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3973e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3974e7931f94SStefano Zampini source_dest = onodes[i]; 3975e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3976e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3977e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3978e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 397928143c3dSStefano Zampini if (nis) { 398028143c3dSStefano 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); 398128143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 398228143c3dSStefano Zampini } 3983e7931f94SStefano Zampini } 3984e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3985e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3986e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3987e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 398828143c3dSStefano Zampini if (nis) { 398928143c3dSStefano 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); 399028143c3dSStefano Zampini } 3991e7931f94SStefano Zampini } 3992e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3993e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3994e7931f94SStefano Zampini 3995e7931f94SStefano Zampini /* assemble new l2g map */ 3996e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3997e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 39989d30be91SStefano Zampini new_local_rows = 0; 3999e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 40009d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4001e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4002e7931f94SStefano Zampini } 40039d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4004e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 40059d30be91SStefano Zampini new_local_rows = 0; 4006e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 40079d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 40089d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4009e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4010e7931f94SStefano Zampini } 40119d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 40129d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4013e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4014e7931f94SStefano Zampini 4015e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4016e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4017e7931f94SStefano 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) */ 4018e7931f94SStefano Zampini if (n_recvs) { 401928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4020e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4021e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4022e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4023e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4024e7931f94SStefano Zampini break; 4025e7931f94SStefano Zampini } 4026e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4027e7931f94SStefano Zampini } 4028e7931f94SStefano Zampini switch (new_local_type_private) { 402928143c3dSStefano Zampini case MATDENSE_PRIVATE: 403028143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4031e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4032e7931f94SStefano Zampini bs = 1; 403328143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 403428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 403528143c3dSStefano Zampini bs = 1; 403628143c3dSStefano Zampini } 4037e7931f94SStefano Zampini break; 4038e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4039e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4040e7931f94SStefano Zampini bs = 1; 4041e7931f94SStefano Zampini break; 4042e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4043e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4044e7931f94SStefano Zampini break; 4045e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4046e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4047e7931f94SStefano Zampini break; 4048e7931f94SStefano Zampini default: 40499d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4050e7931f94SStefano Zampini break; 4051e7931f94SStefano Zampini } 405228143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 405328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 405428143c3dSStefano Zampini bs = 1; 4055e7931f94SStefano Zampini } 4056e7931f94SStefano Zampini 405770cf5478SStefano Zampini /* create MATIS object if needed */ 405870cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4059e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4060e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 406170cf5478SStefano Zampini } else { 406270cf5478SStefano Zampini /* it also destroys the local matrices */ 406370cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 406470cf5478SStefano Zampini } 406570cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4066e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 40679d30be91SStefano Zampini 40689d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 40699d30be91SStefano Zampini 40709d30be91SStefano Zampini /* Global to local map of received indices */ 40719d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 40729d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 40739d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 40749d30be91SStefano Zampini 40759d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 40769d30be91SStefano Zampini buf_size_idxs = 0; 40779d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40789d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 40799d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 40809d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 40819d30be91SStefano Zampini } 40829d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 40839d30be91SStefano Zampini 40849d30be91SStefano Zampini /* set preallocation */ 40859d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 40869d30be91SStefano Zampini if (!newisdense) { 40879d30be91SStefano Zampini PetscInt *new_local_nnz=0; 40889d30be91SStefano Zampini 40899d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 40909d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 40919d30be91SStefano Zampini if (n_recvs) { 40929d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 40939d30be91SStefano Zampini } 40949d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 40959d30be91SStefano Zampini PetscInt j; 40969d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 40979d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 40989d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 40999d30be91SStefano Zampini } 41009d30be91SStefano Zampini } else { 41019d30be91SStefano Zampini /* TODO */ 41029d30be91SStefano Zampini } 41039d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 41049d30be91SStefano Zampini } 41059d30be91SStefano Zampini if (new_local_nnz) { 41069d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 41079d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 41089d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 41099d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 41109d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 41119d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 41129d30be91SStefano Zampini } else { 41139d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 41149d30be91SStefano Zampini } 41159d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 41169d30be91SStefano Zampini } else { 41179d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 41189d30be91SStefano Zampini } 4119e7931f94SStefano Zampini 4120e7931f94SStefano Zampini /* set values */ 4121e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 41229d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 4123e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4124e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 4125e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 41269d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 4127e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4128e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 4129e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 413028143c3dSStefano Zampini } else { 413128143c3dSStefano Zampini /* TODO */ 4132e7931f94SStefano Zampini } 4133e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4134e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 4135e7931f94SStefano Zampini } 4136e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4137e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 413870cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 413970cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41409d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 41419d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 4142e7931f94SStefano Zampini 4143dfd14d43SStefano Zampini #if 0 414428143c3dSStefano Zampini if (!restrict_comm) { /* check */ 4145e7931f94SStefano Zampini Vec lvec,rvec; 4146e7931f94SStefano Zampini PetscReal infty_error; 4147e7931f94SStefano Zampini 41482a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 4149e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 4150e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 4151e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 415270cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 4153e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 4154e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 4155e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 4156e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 4157e7931f94SStefano Zampini } 415828143c3dSStefano Zampini #endif 4159e7931f94SStefano Zampini 416028143c3dSStefano Zampini /* assemble new additional is (if any) */ 416128143c3dSStefano Zampini if (nis) { 416228143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 416328143c3dSStefano Zampini 416428143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4165854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 416628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 416728143c3dSStefano Zampini psum = 0; 416828143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 416928143c3dSStefano Zampini for (j=0;j<nis;j++) { 417028143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 417128143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 417228143c3dSStefano Zampini psum += plen; 417328143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 417428143c3dSStefano Zampini } 417528143c3dSStefano Zampini } 4176854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 4177854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 417828143c3dSStefano Zampini for (i=1;i<nis;i++) { 417928143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 418028143c3dSStefano Zampini } 418128143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 418228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 418328143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 418428143c3dSStefano Zampini for (j=0;j<nis;j++) { 418528143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 418628143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 418728143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 418828143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 418928143c3dSStefano Zampini } 419028143c3dSStefano Zampini } 419128143c3dSStefano Zampini for (i=0;i<nis;i++) { 419228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 419328143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 419428143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 419528143c3dSStefano Zampini } 419628143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 419728143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 419828143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 419928143c3dSStefano Zampini } 4200e7931f94SStefano Zampini /* free workspace */ 420128143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 4202e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4203e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 4204e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4205e7931f94SStefano Zampini if (isdense) { 4206e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4207e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 4208e7931f94SStefano Zampini } else { 4209e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 4210e7931f94SStefano Zampini } 421128143c3dSStefano Zampini if (nis) { 421228143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 421328143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 421428143c3dSStefano Zampini } 4215e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 4216e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 421728143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 4218e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 4219e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 422028143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 4221e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 4222e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 4223e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 4224e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 4225e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 422628143c3dSStefano Zampini if (nis) { 422728143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 422828143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 422928143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 423028143c3dSStefano Zampini } 423128143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 423228143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 423328143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 423428143c3dSStefano Zampini for (i=0;i<nis;i++) { 423528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 423628143c3dSStefano Zampini } 423753a05cb3SStefano Zampini *mat_n = NULL; 423828143c3dSStefano Zampini } 4239e7931f94SStefano Zampini PetscFunctionReturn(0); 4240e7931f94SStefano Zampini } 4241a57a6d2fSStefano Zampini 424212edc857SStefano Zampini /* temporary hack into ksp private data structure */ 4243af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 424412edc857SStefano Zampini 4245c8587f34SStefano Zampini #undef __FUNCT__ 4246c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 4247c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 4248c8587f34SStefano Zampini { 4249c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4250c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 425120a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 42529881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 425320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 42546e683305SStefano Zampini IS coarse_is,*isarray; 42556e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 42566e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 4257f9eb5b7dSStefano Zampini PC pc_temp; 4258c8587f34SStefano Zampini PCType coarse_pc_type; 4259c8587f34SStefano Zampini KSPType coarse_ksp_type; 4260f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 42614f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 42626e683305SStefano Zampini Mat t_coarse_mat_is; 42636e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 42646e683305SStefano Zampini PetscMPIInt all_procs; 426574e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 426668457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 426722bc73bbSStefano Zampini PetscScalar *array; 42689881197aSStefano Zampini PetscErrorCode ierr; 4269fdc09c96SStefano Zampini 4270c8587f34SStefano Zampini PetscFunctionBegin; 4271c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 427268457ee5SStefano 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 */ 4273fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 42745a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4275fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4276f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4277f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4278f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4279fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 428051bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 428151bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 4282dc4bcba2SStefano Zampini PC pc; 4283dc4bcba2SStefano Zampini PetscBool isbddc; 4284dc4bcba2SStefano Zampini 4285dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 4286dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 4287dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 4288dc4bcba2SStefano Zampini if (isbddc) { 4289dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 4290dc4bcba2SStefano Zampini } 4291727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4292fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4293fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4294fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4295f4ddd8eeSStefano Zampini } 4296fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4297fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4298f4ddd8eeSStefano Zampini } 429970cf5478SStefano Zampini /* reset any subassembling information */ 430070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 43016e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 43026e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4303fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4304f4ddd8eeSStefano Zampini } 4305c8587f34SStefano Zampini 43066e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 43072b510759SStefano Zampini im_active = !!(pcis->n); 43082b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43096e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 43106e683305SStefano Zampini void_procs = all_procs-active_procs; 43116e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 431274e2c79eSStefano Zampini redist = PETSC_FALSE; 431322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 43146e683305SStefano Zampini csin_ml = PETSC_TRUE; 43156e683305SStefano Zampini ncoarse_ml = void_procs; 4316779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4317779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 43186e683305SStefano Zampini csin_ds = PETSC_TRUE; 431918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 432018a45a71SStefano Zampini redist = PETSC_TRUE; 432118a45a71SStefano Zampini } else { 43226e683305SStefano Zampini csin_ds = PETSC_TRUE; 4323779c1cceSStefano Zampini ncoarse_ds = active_procs; 4324779c1cceSStefano Zampini redist = PETSC_TRUE; 432518a45a71SStefano Zampini } 43266e683305SStefano Zampini } else { 43276e683305SStefano Zampini csin_ml = PETSC_FALSE; 43286e683305SStefano Zampini ncoarse_ml = all_procs; 43296e683305SStefano Zampini if (void_procs) { 43306e683305SStefano Zampini csin_ds = PETSC_TRUE; 43316e683305SStefano Zampini ncoarse_ds = void_procs; 43326e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 43336e683305SStefano Zampini } else { 4334779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 433574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 433674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 433774e2c79eSStefano Zampini redist = PETSC_TRUE; 433874e2c79eSStefano Zampini } else { 43396e683305SStefano Zampini csin_ds = PETSC_FALSE; 43406e683305SStefano Zampini ncoarse_ds = all_procs; 43416e683305SStefano Zampini } 43426e683305SStefano Zampini } 434374e2c79eSStefano Zampini } 43446e683305SStefano Zampini 43456e683305SStefano Zampini /* 43466e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 43476e683305SStefano Zampini - we have not exceeded the number of levels requested 43486e683305SStefano Zampini - we can actually subassemble the active processes 43496e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 43506e683305SStefano Zampini */ 43516e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 43526e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 43536e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 43546e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 43556e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4356f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 43572b510759SStefano Zampini } else { 4358f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4359c8587f34SStefano Zampini } 4360c8587f34SStefano Zampini } 43616e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 43626e683305SStefano Zampini if (multilevel_allowed) { 43636e683305SStefano Zampini ncoarse = ncoarse_ml; 43646e683305SStefano Zampini csin = csin_ml; 436558da7f69SStefano Zampini redist = PETSC_FALSE; 43666e683305SStefano Zampini } else { 43676e683305SStefano Zampini ncoarse = ncoarse_ds; 43686e683305SStefano Zampini csin = csin_ds; 43696e683305SStefano Zampini } 4370e7931f94SStefano Zampini 4371abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4372abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4373abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4374abbbba34SStefano Zampini 4375abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 437622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 437722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 437822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 437922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4380b1ecc7b1SStefano Zampini #if 0 4381b9b85e73SStefano Zampini { 4382b9b85e73SStefano Zampini PetscViewer viewer; 4383b9b85e73SStefano Zampini char filename[256]; 4384b1ecc7b1SStefano Zampini sprintf(filename,"local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 4385b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4386b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4387b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4388b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4389b9b85e73SStefano Zampini } 4390b9b85e73SStefano Zampini #endif 4391e176bc59SStefano 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); 43926e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 43936e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43946e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4395abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4396abbbba34SStefano Zampini 43976e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 43986e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 43996e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 44006e683305SStefano Zampini const PetscInt *idxs; 44016e683305SStefano Zampini ISLocalToGlobalMapping tmap; 44026e683305SStefano Zampini 44036e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 44040be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 44056e683305SStefano Zampini /* allocate space for temporary storage */ 4406854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4407854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 44086e683305SStefano Zampini /* allocate for IS array */ 44096e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 44106e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 44116e683305SStefano Zampini nis = nisdofs + nisneu; 4412854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 44136e683305SStefano Zampini /* dofs splitting */ 44146e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 44156e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 44166e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 44176e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 44186e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 44196e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 44206e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 44216e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 44226e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 44236e683305SStefano Zampini } 44246e683305SStefano Zampini /* neumann boundaries */ 44256e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 44266e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 44276e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 44286e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 44296e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 44306e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 44316e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 44326e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 44336e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 44346e683305SStefano Zampini } 44356e683305SStefano Zampini /* free memory */ 44366e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 44376e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 44386e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 44396e683305SStefano Zampini } else { 44406e683305SStefano Zampini nis = 0; 44416e683305SStefano Zampini nisdofs = 0; 44426e683305SStefano Zampini nisneu = 0; 44436e683305SStefano Zampini isarray = NULL; 44446e683305SStefano Zampini } 44456e683305SStefano Zampini /* destroy no longer needed map */ 44466e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 44476e683305SStefano Zampini 44486e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 44496e683305SStefano Zampini coarse_mat_is = NULL; 44506e683305SStefano Zampini if (csin) { 44516e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 445274e2c79eSStefano Zampini if (redist) { 445374e2c79eSStefano Zampini PetscMPIInt rank; 4454779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 445574e2c79eSStefano Zampini 445674e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 445758da7f69SStefano Zampini spc = active_procs/ncoarse; 445858da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4459779c1cceSStefano Zampini if (im_active) { 4460779c1cceSStefano Zampini destsize = 1; 446174e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 446274e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 446374e2c79eSStefano Zampini } else { 446474e2c79eSStefano Zampini dest[0] = rank/(spc+1); 446574e2c79eSStefano Zampini } 446674e2c79eSStefano Zampini } else { 4467779c1cceSStefano Zampini destsize = 0; 44686e683305SStefano Zampini } 4469779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4470779c1cceSStefano Zampini } else if (csin_type_simple) { 44716e683305SStefano Zampini PetscMPIInt rank; 44726e683305SStefano Zampini PetscInt issize,isidx; 4473779c1cceSStefano Zampini 44746e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 44756e683305SStefano Zampini if (im_active) { 44766e683305SStefano Zampini issize = 1; 44776e683305SStefano Zampini isidx = (PetscInt)rank; 44786e683305SStefano Zampini } else { 44796e683305SStefano Zampini issize = 0; 44806e683305SStefano Zampini isidx = -1; 44816e683305SStefano Zampini } 44826e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4483779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4484b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 44856e683305SStefano Zampini } 4486779c1cceSStefano Zampini 4487779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4488779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4489779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4490779c1cceSStefano Zampini PetscInt *coarse_candidates; 4491779c1cceSStefano Zampini const PetscInt* tisindices; 4492779c1cceSStefano Zampini 4493779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4494779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4495779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4496779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4497779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4498779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4499779c1cceSStefano Zampini } 4500779c1cceSStefano Zampini } 4501779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4502779c1cceSStefano Zampini 4503779c1cceSStefano Zampini 45046e683305SStefano Zampini if (pcbddc->dbg_flag) { 45056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45066e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 45076e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 45086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4509779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 45106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 45116e683305SStefano Zampini } 45126e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 45136e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45146e683305SStefano Zampini } 45156e683305SStefano Zampini /* shift the pattern on coarse candidates */ 45166e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 45176e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4518854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 45196e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 45206e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 45216e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 45226e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 45236e683305SStefano Zampini } 45246e683305SStefano Zampini if (pcbddc->dbg_flag) { 45256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 45276e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 45286e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45296e683305SStefano Zampini } 4530779c1cceSStefano Zampini } 45316e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 453253a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 453353a05cb3SStefano 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); 453453a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 453553a05cb3SStefano 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); 453653a05cb3SStefano Zampini } 45376e683305SStefano Zampini } else { 45386e683305SStefano Zampini if (pcbddc->dbg_flag) { 45396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 45416e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45426e683305SStefano Zampini } 45436e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 45446e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 45456e683305SStefano Zampini } 45466e683305SStefano Zampini 45476e683305SStefano Zampini /* create local to global scatters for coarse problem */ 454868457ee5SStefano Zampini if (compute_vecs) { 45496e683305SStefano Zampini PetscInt lrows; 45506e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 45516e683305SStefano Zampini if (coarse_mat_is) { 45526e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 45536e683305SStefano Zampini } else { 45546e683305SStefano Zampini lrows = 0; 45556e683305SStefano Zampini } 45566e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 45576e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 45586e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 45596e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45606e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 45616e683305SStefano Zampini } 45626e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 45636e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4564c8587f34SStefano Zampini 4565f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4566f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4567f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4568f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4569f9eb5b7dSStefano Zampini } else { 4570f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4571f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4572c8587f34SStefano Zampini } 4573c8587f34SStefano Zampini 45746e683305SStefano Zampini /* print some info if requested */ 45756e683305SStefano Zampini if (pcbddc->dbg_flag) { 45766e683305SStefano Zampini if (!multilevel_allowed) { 45776e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 45786e683305SStefano Zampini if (multilevel_requested) { 45796e683305SStefano 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); 45806e683305SStefano Zampini } else if (pcbddc->max_levels) { 45816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 45826e683305SStefano Zampini } 45836e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 45846e683305SStefano Zampini } 45856e683305SStefano Zampini } 45866e683305SStefano Zampini 4587f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 45886e683305SStefano Zampini if (coarse_mat_is) { 45896e683305SStefano Zampini MatReuse coarse_mat_reuse; 45906a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 45916e683305SStefano Zampini if (pcbddc->dbg_flag) { 45926e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 45936e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 45946e683305SStefano Zampini } 4595f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4596312be037SStefano Zampini char prefix[256],str_level[16]; 4597e604994aSStefano Zampini size_t len; 45986e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4599422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 4600c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4601f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 46025f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4603c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 46046e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4605c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4606c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4607e604994aSStefano Zampini /* prefix */ 4608e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4609e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4610e604994aSStefano Zampini if (!pcbddc->current_level) { 4611e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4612e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4613c8587f34SStefano Zampini } else { 4614e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4615312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4616312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 461734d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4618312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4619e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4620e604994aSStefano Zampini } 4621e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 46223e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 46233e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 46243e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 46253e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4626f9eb5b7dSStefano Zampini /* allow user customization */ 4627f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 46283e3c6dadSStefano Zampini } 46293e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 463051bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 46313e3c6dadSStefano Zampini if (nisdofs) { 46323e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 46333e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 46343e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 46353e3c6dadSStefano Zampini } 46363e3c6dadSStefano Zampini } 46373e3c6dadSStefano Zampini if (nisneu) { 46383e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 46393e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4640312be037SStefano Zampini } 4641f9eb5b7dSStefano Zampini 4642f9eb5b7dSStefano Zampini /* get some info after set from options */ 4643f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4644f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 46454f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 46466e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4647f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4648f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4649f9eb5b7dSStefano Zampini } 465039f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 46514f3a063dSStefano Zampini if (isredundant) { 46524f3a063dSStefano Zampini KSP inner_ksp; 46534f3a063dSStefano Zampini PC inner_pc; 46544f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 46554f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 46564f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 46574f3a063dSStefano Zampini } 4658f9eb5b7dSStefano Zampini 4659f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4660fa7f1dd8SStefano Zampini if (coarse_reuse) { 466181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4662fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 46636e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4664fa7f1dd8SStefano Zampini } else { 46656e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4666fa7f1dd8SStefano Zampini } 4667c8587f34SStefano Zampini if (isbddc || isnn) { 466822bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 466970cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4670b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 467122b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 46726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 46736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 46746e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 46756e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 467622b6e8a2SStefano Zampini } 467770cf5478SStefano Zampini } 467853a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 467970cf5478SStefano Zampini } else { 468022bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 468122bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 468222bc73bbSStefano Zampini } 468322bc73bbSStefano Zampini } else { 46842e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4685c8587f34SStefano Zampini } 4686c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4687c8587f34SStefano Zampini 46883301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 46895a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 46903301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 46913301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 46923301b35fSStefano Zampini } 46933301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 46943301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 46953301b35fSStefano Zampini } 46963301b35fSStefano Zampini if (pc->pmat->spd_set) { 46973301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 46983301b35fSStefano Zampini } 46996e683305SStefano Zampini /* set operators */ 47005f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 47016e683305SStefano Zampini if (pcbddc->dbg_flag) { 47026e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 47036e683305SStefano Zampini } 47046e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 47056e683305SStefano Zampini coarse_mat = 0; 47066e683305SStefano Zampini } 47076e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4708b1ecc7b1SStefano Zampini #if 0 4709b9b85e73SStefano Zampini { 4710b9b85e73SStefano Zampini PetscViewer viewer; 4711b9b85e73SStefano Zampini char filename[256]; 4712b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 4713b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 4714b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4715b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4716b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4717b9b85e73SStefano Zampini } 4718b9b85e73SStefano Zampini #endif 4719c8587f34SStefano Zampini 4720c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4721298c0119SStefano Zampini #if 0 4722c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4723c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 472498a51de6SStefano Zampini } 4725298c0119SStefano Zampini #endif 472698a51de6SStefano Zampini 472798a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 472898a51de6SStefano Zampini Vec crhs,csol; 472998a51de6SStefano Zampini PetscBool ispreonly; 473004708bb6SStefano Zampini 473198a51de6SStefano Zampini if (CoarseNullSpace) { 4732c8587f34SStefano Zampini if (isbddc) { 4733c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4734c8587f34SStefano Zampini } else { 47355fa7ec2dSBarry Smith ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 4736c8587f34SStefano Zampini } 4737c8587f34SStefano Zampini } 4738f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4739f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4740f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4741f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 47426e683305SStefano Zampini /* hack */ 4743f347579bSStefano Zampini if (!csol) { 47442a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4745f9eb5b7dSStefano Zampini } 4746f347579bSStefano Zampini if (!crhs) { 47472a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4748f347579bSStefano Zampini } 4749cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4750cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 47516e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4752c8587f34SStefano Zampini KSP check_ksp; 47532b510759SStefano Zampini KSPType check_ksp_type; 4754c8587f34SStefano Zampini PC check_pc; 47556e683305SStefano Zampini Vec check_vec,coarse_vec; 47566a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 47572b510759SStefano Zampini PetscInt its; 47586e683305SStefano Zampini PetscBool compute_eigs; 47596e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 47606e683305SStefano Zampini PetscInt neigs; 47618e185a42SStefano Zampini const char *prefix; 4762c8587f34SStefano Zampini 47632b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 47646e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 4765422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 476623ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4767f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 47682b510759SStefano Zampini if (ispreonly) { 47692b510759SStefano Zampini check_ksp_type = KSPPREONLY; 47706e683305SStefano Zampini compute_eigs = PETSC_FALSE; 47712b510759SStefano Zampini } else { 4772cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 47736e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4774c8587f34SStefano Zampini } 4775c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 47766e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 47776e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 47786e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4779a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4780a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4781a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4782a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4783c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4784c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4785c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4786c8587f34SStefano Zampini /* create random vec */ 47876e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 47886e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4789c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4790c8587f34SStefano Zampini if (CoarseNullSpace) { 4791c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4792c8587f34SStefano Zampini } 47936e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4794c8587f34SStefano Zampini /* solve coarse problem */ 47956e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4796c8587f34SStefano Zampini if (CoarseNullSpace) { 47976e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4798c8587f34SStefano Zampini } 4799cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 48006e683305SStefano Zampini if (compute_eigs) { 4801854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4802854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 48036e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 48046e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 48056e683305SStefano Zampini lambda_min = eigs_r[0]; 48066e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 48076e683305SStefano Zampini if (lambda_max>lambda_min) { 4808cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4809cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4810cbcc2c2aSStefano Zampini } 4811c8587f34SStefano Zampini } 4812c8587f34SStefano Zampini } 4813cbcc2c2aSStefano Zampini 4814c8587f34SStefano Zampini /* check coarse problem residual error */ 48156e683305SStefano Zampini if (pcbddc->dbg_flag) { 48166e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 48176e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48186e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4819c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 48206e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 48216e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4822c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4823779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 48246e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 48256e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 48266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 48276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 48286e683305SStefano Zampini if (compute_eigs) { 48296e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4830deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4831c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 48326e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 48336e683305SStefano 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); 48346e683305SStefano Zampini for (i=0;i<neigs;i++) { 48356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4836c8587f34SStefano Zampini } 48376e683305SStefano Zampini } 48386e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 48396e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 48406e683305SStefano Zampini } 4841c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 48426e683305SStefano Zampini if (compute_eigs) { 48436e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 48446e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4845c8587f34SStefano Zampini } 48466e683305SStefano Zampini } 48476e683305SStefano Zampini } 4848cbcc2c2aSStefano Zampini /* print additional info */ 4849cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 48506e683305SStefano Zampini /* waits until all processes reaches this point */ 48516e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4852cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4853cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4854cbcc2c2aSStefano Zampini } 4855cbcc2c2aSStefano Zampini 48562b510759SStefano Zampini /* free memory */ 4857c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4858fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4859c8587f34SStefano Zampini PetscFunctionReturn(0); 4860c8587f34SStefano Zampini } 4861674ae819SStefano Zampini 4862f34684f1SStefano Zampini #undef __FUNCT__ 4863f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4864f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4865f34684f1SStefano Zampini { 4866f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4867f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4868f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4869dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4870dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 487173be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4872dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4873f34684f1SStefano Zampini PetscErrorCode ierr; 4874f34684f1SStefano Zampini 4875f34684f1SStefano Zampini PetscFunctionBegin; 4876f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 48770e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 48780e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4879727cdba6SStefano Zampini } 4880dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 48813bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 4882dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4883dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4884dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4885dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4886dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4887dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 48880e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 48890e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 48900e6343abSStefano Zampini } 4891dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4892dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4893dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4894dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4895dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4896f34684f1SStefano Zampini 4897f34684f1SStefano Zampini /* check numbering */ 4898f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4899019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 4900dc456d91SStefano Zampini PetscInt i; 4901b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4902f34684f1SStefano Zampini 4903f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4904f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4905f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 49060fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4907019a44ceSStefano Zampini /* counter */ 4908019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4909019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4910019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4911019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4912019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4913019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4914019a44ceSStefano Zampini 4915f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4916f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4917727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4918f34684f1SStefano Zampini } 4919f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4920f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4921f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4922e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4923e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4924e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4925e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4926f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4927019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4928f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4929019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 4930019a44ceSStefano Zampini PetscInt owned = (PetscInt)(array[i]); 4931019a44ceSStefano Zampini PetscInt neigh = (PetscInt)(array2[i]); 4932b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4933019a44ceSStefano 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); 4934f34684f1SStefano Zampini } 4935f34684f1SStefano Zampini } 4936019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 4937b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4938f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4939f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4940f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4941f34684f1SStefano Zampini } 4942f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4943f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4944e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4945e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4946f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4947f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4948b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4949ca8b9ea9SStefano Zampini PetscInt *gidxs; 4950ca8b9ea9SStefano Zampini 4951ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 49523bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4953f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4954f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4955f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4956f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 49574bc2dc4bSStefano 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); 4958f34684f1SStefano Zampini } 4959f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4960ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4961f34684f1SStefano Zampini } 4962f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4963302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4964f34684f1SStefano Zampini } 49658bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4966f34684f1SStefano Zampini /* get back data */ 4967f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4968f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4969674ae819SStefano Zampini PetscFunctionReturn(0); 4970674ae819SStefano Zampini } 4971674ae819SStefano Zampini 4972e456f2a8SStefano Zampini #undef __FUNCT__ 4973e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4974a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4975e456f2a8SStefano Zampini { 4976e456f2a8SStefano Zampini IS localis_t; 4977a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4978e456f2a8SStefano Zampini PetscScalar *vals; 4979e456f2a8SStefano Zampini PetscErrorCode ierr; 4980e456f2a8SStefano Zampini 4981e456f2a8SStefano Zampini PetscFunctionBegin; 4982a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4983e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4984854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4985e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4986e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4987a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4988a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 49891035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4990a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 49911035eff8SStefano Zampini } 4992a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4993e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4994e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4995a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4996a7dc3881SStefano Zampini /* now compute set in local ordering */ 4997a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4998a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4999a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5000a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5001a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5002ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5003e456f2a8SStefano Zampini lsize++; 5004e456f2a8SStefano Zampini } 5005e456f2a8SStefano Zampini } 5006854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5007a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5008ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5009e456f2a8SStefano Zampini idxs[lsize++] = i; 5010e456f2a8SStefano Zampini } 5011e456f2a8SStefano Zampini } 5012a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5013a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5014e456f2a8SStefano Zampini *localis = localis_t; 5015e456f2a8SStefano Zampini PetscFunctionReturn(0); 5016e456f2a8SStefano Zampini } 5017906d46d4SStefano Zampini 5018906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5019906d46d4SStefano Zampini #undef __FUNCT__ 5020906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5021906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5022906d46d4SStefano Zampini { 5023906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5024906d46d4SStefano Zampini PetscErrorCode ierr; 5025906d46d4SStefano Zampini 5026906d46d4SStefano Zampini PetscFunctionBegin; 5027906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5028906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5029906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5030906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5031906d46d4SStefano Zampini PetscFunctionReturn(0); 5032906d46d4SStefano Zampini } 5033906d46d4SStefano Zampini 5034906d46d4SStefano Zampini #undef __FUNCT__ 5035906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 5036906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 5037906d46d4SStefano Zampini { 5038906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5039906d46d4SStefano Zampini PetscErrorCode ierr; 5040906d46d4SStefano Zampini 5041906d46d4SStefano Zampini PetscFunctionBegin; 5042906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5043906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5044906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5045906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5046906d46d4SStefano Zampini PetscFunctionReturn(0); 5047906d46d4SStefano Zampini } 5048b96c3477SStefano Zampini 5049b96c3477SStefano Zampini #undef __FUNCT__ 5050b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 505108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 5052b96c3477SStefano Zampini { 5053a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5054b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5055b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5056a64f4aa4SStefano Zampini Mat S_j; 5057b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 5058b96c3477SStefano Zampini PetscBool free_used_adj; 5059b96c3477SStefano Zampini PetscErrorCode ierr; 5060b96c3477SStefano Zampini 5061b96c3477SStefano Zampini PetscFunctionBegin; 5062b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 5063b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 506408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 5065b96c3477SStefano Zampini used_xadj = NULL; 5066b96c3477SStefano Zampini used_adjncy = NULL; 5067b96c3477SStefano Zampini } else { 506808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 506908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 507008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 507108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 5072b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 5073b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 5074b96c3477SStefano Zampini } else { 50752fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 5076b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 5077b96c3477SStefano Zampini PetscInt nvtxs; 5078b96c3477SStefano Zampini 50792fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 50802fffb893SStefano Zampini if (flg_row) { 5081b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 5082b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 5083b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 5084b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 50852fffb893SStefano Zampini } else { 50862fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 50872fffb893SStefano Zampini used_xadj = NULL; 50882fffb893SStefano Zampini used_adjncy = NULL; 50892fffb893SStefano Zampini } 50902fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 5091b96c3477SStefano Zampini } 5092b96c3477SStefano Zampini } 5093d5574798SStefano Zampini 5094d5574798SStefano Zampini /* setup sub_schurs data */ 5095a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 5096a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 5097a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 5098a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 509906a4e24aSStefano 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); 5100a64f4aa4SStefano Zampini } else { 51016816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 510204708bb6SStefano Zampini PetscBool isseqaij; 51035feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 51045feab87aSStefano Zampini PetscInt n_vertices; 51055feab87aSStefano Zampini 51065feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 51072034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 51085feab87aSStefano Zampini } 510904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 511004708bb6SStefano Zampini if (!isseqaij) { 511104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 511204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 511304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 511404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 511504708bb6SStefano Zampini } else { 511604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 511704708bb6SStefano Zampini } 511804708bb6SStefano Zampini } 511906a4e24aSStefano 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); 5120a64f4aa4SStefano Zampini } 5121a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 5122b96c3477SStefano Zampini 5123b96c3477SStefano Zampini /* free adjacency */ 5124b96c3477SStefano Zampini if (free_used_adj) { 5125b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 5126b96c3477SStefano Zampini } 5127b96c3477SStefano Zampini PetscFunctionReturn(0); 5128b96c3477SStefano Zampini } 5129b96c3477SStefano Zampini 5130b96c3477SStefano Zampini #undef __FUNCT__ 5131b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 513208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 5133b96c3477SStefano Zampini { 5134b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 5135b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 5136b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 5137b96c3477SStefano Zampini PCBDDCGraph graph; 5138b96c3477SStefano Zampini PetscErrorCode ierr; 5139b96c3477SStefano Zampini 5140b96c3477SStefano Zampini PetscFunctionBegin; 5141b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 514208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 51433301b35fSStefano Zampini IS verticesIS,verticescomm; 51443301b35fSStefano Zampini PetscInt vsize,*idxs; 5145b96c3477SStefano Zampini 5146b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 51473301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 51483301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51493301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 51503301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 51513301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 5152b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 51537fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 51543301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 51553301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 5156b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 5157b96c3477SStefano Zampini /* 5158b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 5159b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 5160b96c3477SStefano Zampini } 5161b96c3477SStefano Zampini */ 5162b96c3477SStefano Zampini } else { 5163b96c3477SStefano Zampini graph = pcbddc->mat_graph; 5164b96c3477SStefano Zampini } 5165b96c3477SStefano Zampini 5166b96c3477SStefano Zampini /* sub_schurs init */ 5167a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 5168a64f4aa4SStefano Zampini 5169b96c3477SStefano Zampini /* free graph struct */ 517008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 5171b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 5172b96c3477SStefano Zampini } 5173b96c3477SStefano Zampini PetscFunctionReturn(0); 5174b96c3477SStefano Zampini } 5175