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 8b1b3d7a2SStefano Zampini #undef __FUNCT__ 9b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 1008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 11b1b3d7a2SStefano Zampini { 12b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 1508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 1608122e43SStefano Zampini PetscScalar *work,lwork; 1708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 1808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 1908122e43SStefano Zampini PetscReal *eigs,thresh; 201b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 21f6f667cfSStefano Zampini PetscBool allocated_S_St; 2208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2308122e43SStefano Zampini PetscReal *rwork; 2408122e43SStefano Zampini #endif 25b1b3d7a2SStefano Zampini PetscErrorCode ierr; 26b1b3d7a2SStefano Zampini 27b1b3d7a2SStefano Zampini PetscFunctionBegin; 2808122e43SStefano Zampini if (!sub_schurs->use_mumps) { 2908122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3008122e43SStefano Zampini } 3108122e43SStefano Zampini 32fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 33fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 34fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 35fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 36fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 37fd14bc51SStefano Zampini } 38fd14bc51SStefano Zampini 39e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 40e496cd5dSStefano 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); 41e496cd5dSStefano Zampini } 42e496cd5dSStefano Zampini 43e496cd5dSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 44e496cd5dSStefano 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); 45e496cd5dSStefano Zampini } 46e496cd5dSStefano Zampini 4708122e43SStefano Zampini /* max size of subsets */ 4808122e43SStefano Zampini mss = 0; 4908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5008122e43SStefano Zampini PetscInt subset_size; 51862806e4SStefano Zampini 5208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5408122e43SStefano Zampini } 5508122e43SStefano Zampini 5608122e43SStefano Zampini /* min/max and threshold */ 5708122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 58f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 5908122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 60f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 61f6f667cfSStefano Zampini if (nmin) { 62f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 63f6f667cfSStefano Zampini } 6408122e43SStefano Zampini 6508122e43SStefano Zampini /* allocate lapack workspace */ 6608122e43SStefano Zampini cum = cum2 = 0; 6708122e43SStefano Zampini maxneigs = 0; 6808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6908122e43SStefano Zampini PetscInt n,subset_size; 70f6f667cfSStefano Zampini 7108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7208122e43SStefano Zampini n = PetscMin(subset_size,nmax); 739162d606SStefano Zampini cum += subset_size; 749162d606SStefano Zampini cum2 += subset_size*n; 7508122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7608122e43SStefano Zampini } 7708122e43SStefano Zampini if (mss) { 789ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 7908122e43SStefano Zampini PetscBLASInt B_itype = 1; 8008122e43SStefano Zampini PetscBLASInt B_N = mss; 814c6709b3SStefano Zampini PetscReal zero = 0.0; 824c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8308122e43SStefano Zampini 8408122e43SStefano Zampini B_lwork = -1; 8508122e43SStefano Zampini S = NULL; 8608122e43SStefano Zampini St = NULL; 87a58a30b4SStefano Zampini eigs = NULL; 88a58a30b4SStefano Zampini eigv = NULL; 89a58a30b4SStefano Zampini B_iwork = NULL; 90a58a30b4SStefano Zampini B_ifail = NULL; 918bec7fa6SStefano Zampini thresh = 1.0; 9208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9408122e43SStefano 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)); 9508122e43SStefano Zampini #else 9608122e43SStefano 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)); 9708122e43SStefano Zampini #endif 9808122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 9908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10008122e43SStefano Zampini } else { 10108122e43SStefano Zampini /* TODO */ 10208122e43SStefano Zampini } 10308122e43SStefano Zampini } else { 10408122e43SStefano Zampini lwork = 0; 10508122e43SStefano Zampini } 10608122e43SStefano Zampini 10708122e43SStefano Zampini nv = 0; 108d62866d3SStefano 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) */ 109d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 11008122e43SStefano Zampini } 1114c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 112f6f667cfSStefano Zampini if (allocated_S_St) { 113f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 114f6f667cfSStefano Zampini } 115f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11708122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 11808122e43SStefano Zampini #endif 1199162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1209162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1219162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12208122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1239162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12408122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12508122e43SStefano Zampini 12608122e43SStefano Zampini maxneigs = 0; 12708122e43SStefano Zampini cum = cum2 = cumarray = 0; 1289162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1299162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 130d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13108122e43SStefano Zampini const PetscInt *idxs; 13208122e43SStefano Zampini 133d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 13408122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 13508122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 13708122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1389162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1399162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 14008122e43SStefano Zampini } 14108122e43SStefano Zampini cum2 = cum; 142d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14308122e43SStefano Zampini } 14408122e43SStefano Zampini 14508122e43SStefano Zampini if (mss) { /* multilevel */ 14608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 14708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 14808122e43SStefano Zampini } 14908122e43SStefano Zampini 15008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15108122e43SStefano Zampini 15208122e43SStefano Zampini const PetscInt *idxs; 153f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 154862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 15508122e43SStefano Zampini PetscBLASInt B_N; 156aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 15708122e43SStefano Zampini 158862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 159f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 160f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1619ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 162aff50787SStefano Zampini PetscInt j,k; 163aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 164aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 165aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 16608122e43SStefano Zampini } 16708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 168aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 169aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 170aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 171aff50787SStefano Zampini } 17208122e43SStefano Zampini } 17308122e43SStefano Zampini } else { 17408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17608122e43SStefano Zampini } 1778bec7fa6SStefano Zampini } else { 178f6f667cfSStefano Zampini S = Sarray + cumarray; 179f6f667cfSStefano Zampini St = Starray + cumarray; 1808bec7fa6SStefano Zampini } 18108122e43SStefano Zampini 182f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 183aff50787SStefano Zampini /* see if we can save some work */ 184aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 185aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 186aff50787SStefano Zampini } 187aff50787SStefano Zampini 188aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 189aff50787SStefano Zampini B_neigs = 0; 190aff50787SStefano Zampini } else { 191aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 192f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 193f6f667cfSStefano Zampini 1949ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 19508122e43SStefano Zampini PetscBLASInt B_itype = 1; 196f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1974c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1989552c7c7SStefano Zampini PetscInt nmin_s; 19908122e43SStefano Zampini 200f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 201fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2028bec7fa6SStefano 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]]); 203fd14bc51SStefano Zampini } 20408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 20508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 206f6f667cfSStefano 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)); 20708122e43SStefano Zampini #else 208f6f667cfSStefano 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)); 20908122e43SStefano Zampini #endif 21008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 21108122e43SStefano Zampini if (B_ierr) { 21208122e43SStefano Zampini if (B_ierr < 0 ) { 21308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 21408122e43SStefano Zampini } else if (B_ierr <= B_N) { 21508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 21608122e43SStefano Zampini } else { 2179552c7c7SStefano 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); 21808122e43SStefano Zampini } 21908122e43SStefano Zampini } 22008122e43SStefano Zampini 22108122e43SStefano Zampini if (B_neigs > nmax) { 222fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 223fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 224fd14bc51SStefano Zampini } 225f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 22608122e43SStefano Zampini B_neigs = nmax; 22708122e43SStefano Zampini } 22808122e43SStefano Zampini 2299552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2309552c7c7SStefano Zampini if (B_neigs < nmin_s) { 23108122e43SStefano Zampini PetscBLASInt B_neigs2; 23208122e43SStefano Zampini 233f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 234f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 235fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 236fd14bc51SStefano 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); 237fd14bc51SStefano Zampini } 2389ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23908122e43SStefano Zampini PetscInt j; 24008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24108122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24208122e43SStefano Zampini } 24308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24408122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24508122e43SStefano Zampini } 24608122e43SStefano Zampini } else { 24708122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24808122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24908122e43SStefano Zampini } 25008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 252f6f667cfSStefano 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)); 25308122e43SStefano Zampini #else 254f6f667cfSStefano 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)); 25508122e43SStefano Zampini #endif 25608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 25708122e43SStefano Zampini B_neigs += B_neigs2; 25808122e43SStefano Zampini } 25908122e43SStefano Zampini if (B_ierr) { 26008122e43SStefano Zampini if (B_ierr < 0 ) { 26108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 26208122e43SStefano Zampini } else if (B_ierr <= B_N) { 26308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 26408122e43SStefano Zampini } else { 2659552c7c7SStefano 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); 26608122e43SStefano Zampini } 26708122e43SStefano Zampini } 268fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 269fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 27008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 27108122e43SStefano Zampini if (eigs[j] == 0.0) { 272fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 27308122e43SStefano Zampini } else { 274f6f667cfSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]); 275fd14bc51SStefano Zampini } 27608122e43SStefano Zampini } 27708122e43SStefano Zampini } 27808122e43SStefano Zampini } else { 27908122e43SStefano Zampini /* TODO */ 28008122e43SStefano Zampini } 281aff50787SStefano Zampini } 2828bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2838bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2849162d606SStefano Zampini if (B_neigs) { 2859162d606SStefano 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); 286fd14bc51SStefano Zampini 287fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2889552c7c7SStefano Zampini PetscInt ii; 2899552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 290fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2919552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2929162d606SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2939552c7c7SStefano Zampini } 2949552c7c7SStefano Zampini } 295fd14bc51SStefano Zampini } 29608122e43SStefano Zampini #if 0 2979162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29808122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29908122e43SStefano Zampini PetscScalar norm; 30008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3019162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3029162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 30308122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 30408122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30508122e43SStefano Zampini } else { 30608122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 30708122e43SStefano Zampini } 3089162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 309b1b3d7a2SStefano Zampini } 310b1b3d7a2SStefano Zampini #endif 3119162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3129162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3139162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3149162d606SStefano Zampini cum++; 31508122e43SStefano Zampini } 31608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 31708122e43SStefano Zampini /* shift for next computation */ 31808122e43SStefano Zampini cumarray += subset_size*subset_size; 31908122e43SStefano Zampini } 320fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 321fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 322fd14bc51SStefano Zampini } 32308122e43SStefano Zampini 32408122e43SStefano Zampini if (mss) { 32508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 32608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 327f6f667cfSStefano Zampini /* destroy matrices (junk) */ 328f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 329f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 33008122e43SStefano Zampini } 331f6f667cfSStefano Zampini if (allocated_S_St) { 332f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 333f6f667cfSStefano Zampini } 334f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 33508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 33608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33708122e43SStefano Zampini #endif 33808122e43SStefano Zampini if (pcbddc->dbg_flag) { 3391b968477SStefano Zampini PetscInt maxneigs_r; 34008122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3419b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 34208122e43SStefano Zampini } 34308122e43SStefano Zampini PetscFunctionReturn(0); 34408122e43SStefano Zampini } 345b1b3d7a2SStefano Zampini 346674ae819SStefano Zampini #undef __FUNCT__ 347c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 348c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 349c8587f34SStefano Zampini { 350c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3518629588bSStefano Zampini PetscScalar *coarse_submat_vals; 352c8587f34SStefano Zampini PetscErrorCode ierr; 353c8587f34SStefano Zampini 354c8587f34SStefano Zampini PetscFunctionBegin; 355f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3565e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 357c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 358c8587f34SStefano Zampini 359684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3600fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 361684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 362c8587f34SStefano Zampini 363c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 364b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 365c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 366c8587f34SStefano Zampini } 367c8587f34SStefano Zampini 3688629588bSStefano Zampini /* 3698629588bSStefano Zampini Setup local correction and local part of coarse basis. 3708629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3718629588bSStefano Zampini */ 37247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3738629588bSStefano Zampini 3748629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3758629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3768629588bSStefano Zampini 3778629588bSStefano Zampini /* free */ 3788629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 379c8587f34SStefano Zampini PetscFunctionReturn(0); 380c8587f34SStefano Zampini } 381c8587f34SStefano Zampini 382c8587f34SStefano Zampini #undef __FUNCT__ 383674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 384674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 385674ae819SStefano Zampini { 386674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 387674ae819SStefano Zampini PetscErrorCode ierr; 388674ae819SStefano Zampini 389674ae819SStefano Zampini PetscFunctionBegin; 390674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 391674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 392674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 393674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 394785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 395674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 396f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 397f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 398785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 40063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 401674ae819SStefano Zampini PetscFunctionReturn(0); 402674ae819SStefano Zampini } 403674ae819SStefano Zampini 404674ae819SStefano Zampini #undef __FUNCT__ 405674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 406674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 407674ae819SStefano Zampini { 408674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 409674ae819SStefano Zampini PetscErrorCode ierr; 410674ae819SStefano Zampini 411674ae819SStefano Zampini PetscFunctionBegin; 412b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 413674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 414674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 415674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 416b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 417674ae819SStefano Zampini PetscFunctionReturn(0); 418674ae819SStefano Zampini } 419674ae819SStefano Zampini 420674ae819SStefano Zampini #undef __FUNCT__ 421674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 422674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 423674ae819SStefano Zampini { 424674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 42506656605SStefano Zampini PetscScalar *array; 426674ae819SStefano Zampini PetscErrorCode ierr; 427674ae819SStefano Zampini 428674ae819SStefano Zampini PetscFunctionBegin; 429674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43058da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 43106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 43206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43358da7f69SStefano Zampini } 434674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 43615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 44106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 442674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 443674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4448ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 445674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 446674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 447674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 448f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 449f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 450f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 451f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 452727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4530e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 454f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 45570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4566e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 457674ae819SStefano Zampini PetscFunctionReturn(0); 458674ae819SStefano Zampini } 459674ae819SStefano Zampini 460674ae819SStefano Zampini #undef __FUNCT__ 461f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 462f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4636bfb1811SStefano Zampini { 4646bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4656bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4666bfb1811SStefano Zampini VecType impVecType; 467e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4686bfb1811SStefano Zampini PetscErrorCode ierr; 4696bfb1811SStefano Zampini 4706bfb1811SStefano Zampini PetscFunctionBegin; 471f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 472f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 473f4ddd8eeSStefano Zampini } 474e7b262bdSStefano Zampini /* get sizes */ 475b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 476b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4776bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 478e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 479e7b262bdSStefano Zampini /* R nodes */ 480e7b262bdSStefano Zampini old_size = -1; 481e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 482e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 483e7b262bdSStefano Zampini } 484e7b262bdSStefano Zampini if (n_R != old_size) { 485e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 486e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4876bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4886bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4896bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4906bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 491e7b262bdSStefano Zampini } 492e7b262bdSStefano Zampini /* local primal dofs */ 493e7b262bdSStefano Zampini old_size = -1; 494e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 495e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 496e7b262bdSStefano Zampini } 497e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 498e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 500e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5016bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 502e7b262bdSStefano Zampini } 503e7b262bdSStefano Zampini /* local explicit constraints */ 504e7b262bdSStefano Zampini old_size = -1; 505e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 506e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 507e7b262bdSStefano Zampini } 508e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 509e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 51083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 51183b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 51283b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 51383b7ccabSStefano Zampini } 5146bfb1811SStefano Zampini PetscFunctionReturn(0); 5156bfb1811SStefano Zampini } 5166bfb1811SStefano Zampini 5176bfb1811SStefano Zampini #undef __FUNCT__ 51847f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 52088ebb749SStefano Zampini { 52125084f0cSStefano Zampini PetscErrorCode ierr; 52225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 52388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 52488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 525d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 52625084f0cSStefano Zampini /* submatrices of local problem */ 52780677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52806656605SStefano Zampini /* submatrices of local coarse problem */ 52906656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 53025084f0cSStefano Zampini /* working matrices */ 53106656605SStefano Zampini Mat C_CR; 53225084f0cSStefano Zampini /* additional working stuff */ 53306656605SStefano Zampini PC pc_R; 53406656605SStefano Zampini Mat F; 53506656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 53606656605SStefano Zampini 53725084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53806656605SStefano Zampini PetscScalar *work; 53906656605SStefano Zampini PetscInt *idx_V_B; 54006656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 54106656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 542b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 54345a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 54488ebb749SStefano Zampini MatType impMatType; 54525084f0cSStefano Zampini /* some shortcuts to scalars */ 54606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54788ebb749SStefano Zampini 54888ebb749SStefano Zampini PetscFunctionBegin; 549b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 55088ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 55188ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 552b371cd4fSStefano Zampini n_B = pcis->n_B; 553b371cd4fSStefano Zampini n_D = pcis->n - n_B; 55488ebb749SStefano Zampini n_R = pcis->n - n_vertices; 55588ebb749SStefano Zampini 55688ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55788ebb749SStefano Zampini impMatType = MATSEQDENSE; 55888ebb749SStefano Zampini 55988ebb749SStefano Zampini /* vertices in boundary numbering */ 560785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5610e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 56288ebb749SStefano Zampini if (i != n_vertices) { 56322d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 56488ebb749SStefano Zampini } 56588ebb749SStefano Zampini 56606656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 56706656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 57006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 57106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 57206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 57306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 57406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 57506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 57606656605SStefano Zampini 57706656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57806656605SStefano Zampini /* allocate workspace */ 57906656605SStefano Zampini n = 0; 58006656605SStefano Zampini if (n_constraints) { 58106656605SStefano Zampini n += n_R*n_constraints; 58206656605SStefano Zampini } 58306656605SStefano Zampini if (n_vertices) { 58406656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 58580677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 58606656605SStefano Zampini } 5873301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 58806656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58906656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 59006656605SStefano Zampini } 59106656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 59206656605SStefano Zampini 59306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 59406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 59506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 59606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 59706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59806656605SStefano Zampini if (isLU || isILU || isCHOL) { 59906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 600d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 601d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 602d62866d3SStefano Zampini MatFactorType type; 603d62866d3SStefano Zampini 6046816873aSStefano Zampini F = reuse_mumps->F; 6056816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 606d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 60706656605SStefano Zampini } else { 60806656605SStefano Zampini F = NULL; 60906656605SStefano Zampini } 61006656605SStefano Zampini 61188ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 61288ebb749SStefano Zampini if (n_constraints) { 61306656605SStefano Zampini Mat M1,M2,M3; 61480677318SStefano Zampini Mat auxmat; 61506656605SStefano Zampini IS is_aux; 61680677318SStefano Zampini PetscScalar *array,*array2; 61706656605SStefano Zampini 618f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 61980677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 62088ebb749SStefano Zampini 62125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 62225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6238ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 62480677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 62588ebb749SStefano Zampini 62680677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 62780677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 62806656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62988ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 63006656605SStefano Zampini const PetscScalar *row_cmat_values; 63106656605SStefano Zampini const PetscInt *row_cmat_indices; 63206656605SStefano Zampini PetscInt size_of_constraint,j; 63388ebb749SStefano Zampini 63406656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63506656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 63606656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 63706656605SStefano Zampini } 63806656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63906656605SStefano Zampini } 64080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 64106656605SStefano Zampini if (F) { 64206656605SStefano Zampini Mat B; 64306656605SStefano Zampini 64406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 64580677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 64606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64706656605SStefano Zampini } else { 64880677318SStefano Zampini PetscScalar *marr; 64980677318SStefano Zampini 65080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 65206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 65380677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 65406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 65506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 65606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65706656605SStefano Zampini } 65880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65906656605SStefano Zampini } 66080677318SStefano Zampini if (!pcbddc->switch_static) { 66180677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 66480677318SStefano Zampini for (i=0;i<n_constraints;i++) { 66580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 66680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 66780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 67080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 67180677318SStefano Zampini } 67280677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67380677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 67480677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67580677318SStefano Zampini } else { 67680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 67780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 67825084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67980677318SStefano Zampini } 68080677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 68180677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 68280677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 68306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 68406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 68580677318SStefano Zampini if (isCHOL) { 68680677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 68780677318SStefano Zampini } else { 68825084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 68980677318SStefano Zampini } 69080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 69106656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 69225084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 69325084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 69425084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 69580677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 69680677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 69780677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 69806656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 69906656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 70088ebb749SStefano Zampini } 70188ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 70288ebb749SStefano Zampini if (n_vertices) { 70306656605SStefano Zampini IS is_aux; 7043a50541eSStefano Zampini 7056816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7066816873aSStefano Zampini IS tis; 7076816873aSStefano Zampini 7086816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7096816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7106816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7126816873aSStefano Zampini } else { 7133a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7146816873aSStefano Zampini } 7159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 717*04708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 71825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 71988ebb749SStefano Zampini } 72088ebb749SStefano Zampini 72188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 722f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 72306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 72406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 72506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 72606656605SStefano Zampini } 727f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 72806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 72906656605SStefano Zampini PetscScalar *marray; 73006656605SStefano Zampini 73106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 73206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 733f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 734f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 735f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 736f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 737f4ddd8eeSStefano Zampini } 738f4ddd8eeSStefano Zampini } 73906656605SStefano Zampini 740f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 74106656605SStefano Zampini PetscScalar *marray; 74288ebb749SStefano Zampini 74306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7448eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 74506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 74688ebb749SStefano Zampini } 7473301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 74806656605SStefano Zampini n *= 2; 749f4ddd8eeSStefano Zampini } 75006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 75106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 75206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 75506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75688ebb749SStefano Zampini } 7573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 75806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7591b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 76106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 76288ebb749SStefano Zampini } 76388ebb749SStefano Zampini } else { 764c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 765c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7661b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 767c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 768c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 769c0553b1fSStefano Zampini } 77088ebb749SStefano Zampini } 77106656605SStefano Zampini } 77206656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 77306656605SStefano Zampini /* vertices */ 77406656605SStefano Zampini if (n_vertices) { 77516f15bc4SStefano Zampini 776*04708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 777*04708bb6SStefano Zampini 77816f15bc4SStefano Zampini if (n_R) { 77906656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 78006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 78116f15bc4SStefano Zampini PetscScalar *x,*y; 782*04708bb6SStefano Zampini PetscBool isseqaij; 78306656605SStefano Zampini 78421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 78506656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 786*04708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7876816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 78806656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 78906656605SStefano Zampini } else { 79006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 79106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 79206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 79306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 79406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 79506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 79606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 79706656605SStefano Zampini } 79806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 79906656605SStefano Zampini } 80080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 80106656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 80206656605SStefano Zampini if (n_constraints) { 80306656605SStefano Zampini Mat B; 80480677318SStefano Zampini 805b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 80680677318SStefano Zampini for (i=0;i<n_vertices;i++) { 80780677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 80880677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 80980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 81280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 81380677318SStefano Zampini } 81480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 81580677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 81680677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 81706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 81880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 81906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 82006656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 82106656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 82206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82306656605SStefano Zampini } 824*04708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 825*04708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 826*04708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 827*04708bb6SStefano Zampini } 82806656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 82980677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 83006656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 83106656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 83206656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 83306656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 83406656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 83606656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 83706656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 83816f15bc4SStefano Zampini } else { 83916f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84016f15bc4SStefano Zampini } 84121eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 84206656605SStefano Zampini /* coarse basis functions */ 84306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 84416f15bc4SStefano Zampini PetscScalar *y; 84516f15bc4SStefano Zampini 84606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 84706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 84806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 85206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 85406656605SStefano Zampini 85506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 85606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 86106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86206656605SStefano Zampini } 86306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 86406656605SStefano Zampini } 865*04708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 866*04708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 86706656605SStefano Zampini } 86806656605SStefano Zampini 86906656605SStefano Zampini if (n_constraints) { 87006656605SStefano Zampini Mat B; 87106656605SStefano Zampini 87206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 87306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 87480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 87506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 87606656605SStefano Zampini if (n_vertices) { 87780677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 87880677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 87980677318SStefano Zampini } else { 88080677318SStefano Zampini Mat S_VCt; 88180677318SStefano Zampini 88280677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 88380677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 88480677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 88580677318SStefano Zampini } 88606656605SStefano Zampini } 88706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 88806656605SStefano Zampini /* coarse basis functions */ 88906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 89006656605SStefano Zampini PetscScalar *y; 89106656605SStefano Zampini 89206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 89306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 89506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 89706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 89806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 89906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 90006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 90206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90606656605SStefano Zampini } 90706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 90806656605SStefano Zampini } 90906656605SStefano Zampini } 91080677318SStefano Zampini if (n_constraints) { 91180677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 91280677318SStefano Zampini } 91306656605SStefano Zampini 91406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9153301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 91606656605SStefano Zampini 91706656605SStefano Zampini if (n_constraints) { 91816f15bc4SStefano Zampini Mat S_CCT,B_C; 91906656605SStefano Zampini 92080677318SStefano Zampini /* this is a lazy thing */ 92180677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 92206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 92306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 92406656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 92516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 92606656605SStefano Zampini if (n_vertices) { 92716f15bc4SStefano Zampini Mat B_V,S_VCT; 92806656605SStefano Zampini 92906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 93006656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 93106656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 93206656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 93316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 93406656605SStefano Zampini } 93506656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 93680677318SStefano Zampini } else { /* if there are no constraints, reset work */ 93780677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 93806656605SStefano Zampini } 93916f15bc4SStefano Zampini if (n_vertices && n_R) { 94006656605SStefano Zampini Mat A_VRT; 94180677318SStefano Zampini PetscScalar *marray; 94206656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 94306656605SStefano Zampini 94480677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 94580677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 94680677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 94706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 94880677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 94980677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 95016f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 95106656605SStefano Zampini } 95206656605SStefano Zampini 95306656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 95406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 95506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 95606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 95706656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 95806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96006656605SStefano Zampini } 96106656605SStefano Zampini } else { 96206656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96806656605SStefano Zampini } 96906656605SStefano Zampini } 97006656605SStefano Zampini /* coarse basis functions */ 97106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97206656605SStefano Zampini PetscScalar *y; 97306656605SStefano Zampini 97406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 97506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 97906656605SStefano Zampini if (i<n_vertices) { 98006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 98106656605SStefano Zampini } 98206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 98406656605SStefano Zampini 98506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 98606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 98706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 98806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 99106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99206656605SStefano Zampini } 99306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 99406656605SStefano Zampini } 99506656605SStefano Zampini } 996d62866d3SStefano Zampini /* free memory */ 99788ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 99806656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1002d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1003d62866d3SStefano Zampini if (n_vertices) { 1004d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1005d62866d3SStefano Zampini } 1006d62866d3SStefano Zampini if (n_constraints) { 1007d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1008d62866d3SStefano Zampini } 100988ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 101088ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 101188ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 101225084f0cSStefano Zampini if (pcbddc->dbg_flag) { 101388ebb749SStefano Zampini Mat coarse_sub_mat; 101425084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 101588ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 101688ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 101788ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10188bec7fa6SStefano Zampini Mat C_B,CPHI; 10198bec7fa6SStefano Zampini IS is_dummy; 10208bec7fa6SStefano Zampini Vec mones; 102188ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 102288ebb749SStefano Zampini PetscReal real_value; 102388ebb749SStefano Zampini 102488ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 102588ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 102688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 102788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 102888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 102988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1030c0553b1fSStefano Zampini if (unsymmetric_check) { 103188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 103388ebb749SStefano Zampini } 103488ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 103588ebb749SStefano Zampini 103625084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10373301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 103825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1039c0553b1fSStefano Zampini if (unsymmetric_check) { 104088ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104388ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105288ebb749SStefano Zampini } else { 105388ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106188ebb749SStefano Zampini } 106288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106588ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 106681d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10678bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10680fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 106906656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10708bec7fa6SStefano Zampini 10718bec7fa6SStefano Zampini /* check constraints */ 10728bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10738bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10748bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10758bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10768bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10778bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10788bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1079bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1080bdae7319SStefano Zampini if (unsymmetric_check) { 1081bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1082bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1083bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1084bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1085bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1086bdae7319SStefano Zampini } 10878bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10888bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10898bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10908bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10918bec7fa6SStefano Zampini 109225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 109388ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 109488ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 109588ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 109688ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 109788ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 109888ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 109988ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 110088ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 110188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 110288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1103c0553b1fSStefano Zampini if (unsymmetric_check) { 110488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 110688ebb749SStefano Zampini } 110788ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 110888ebb749SStefano Zampini } 11098629588bSStefano Zampini /* get back data */ 11108629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 111188ebb749SStefano Zampini PetscFunctionReturn(0); 111288ebb749SStefano Zampini } 111388ebb749SStefano Zampini 111488ebb749SStefano Zampini #undef __FUNCT__ 1115d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1116d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1117d65f70fdSStefano Zampini { 1118d65f70fdSStefano Zampini Mat *work_mat; 1119d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1120d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1121d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1122d65f70fdSStefano Zampini PetscErrorCode ierr; 1123d65f70fdSStefano Zampini 1124d65f70fdSStefano Zampini PetscFunctionBegin; 1125d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1126d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1127d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1128d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1129d65f70fdSStefano Zampini 1130d65f70fdSStefano Zampini if (!rsorted) { 1131d65f70fdSStefano Zampini const PetscInt *idxs; 1132d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1133d65f70fdSStefano Zampini 1134d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1135d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1136d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1137d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1138d65f70fdSStefano Zampini } 1139d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1140d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1142d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1143d65f70fdSStefano Zampini } 1144d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini } else { 1147d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1148d65f70fdSStefano Zampini isrow_s = isrow; 1149d65f70fdSStefano Zampini } 1150d65f70fdSStefano Zampini 1151d65f70fdSStefano Zampini if (!csorted) { 1152d65f70fdSStefano Zampini if (isrow == iscol) { 1153d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini iscol_s = isrow_s; 1155d65f70fdSStefano Zampini } else { 1156d65f70fdSStefano Zampini const PetscInt *idxs; 1157d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1158d65f70fdSStefano Zampini 1159d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1160d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1161d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1162d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1163d65f70fdSStefano Zampini } 1164d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1167d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1168d65f70fdSStefano Zampini } 1169d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini } 1172d65f70fdSStefano Zampini } else { 1173d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1174d65f70fdSStefano Zampini iscol_s = iscol; 1175d65f70fdSStefano Zampini } 1176d65f70fdSStefano Zampini 1177d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1178d65f70fdSStefano Zampini 1179d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1180d65f70fdSStefano Zampini Mat new_mat; 1181d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1182d65f70fdSStefano Zampini 1183d65f70fdSStefano Zampini if (!rsorted) { 1184d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1185d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1186d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1187d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1188d65f70fdSStefano Zampini } 1189d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1190d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini } else { 1192d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini 1196d65f70fdSStefano Zampini if (!csorted) { 1197d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1198d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1200d65f70fdSStefano Zampini } else { 1201d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1202d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1203d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1204d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1205d65f70fdSStefano Zampini } 1206d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1207d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini } 1209d65f70fdSStefano Zampini } else { 1210d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1211d65f70fdSStefano Zampini } 1212d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini 1214d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1215d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1216d65f70fdSStefano Zampini work_mat[0] = new_mat; 1217d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1219d65f70fdSStefano Zampini } 1220d65f70fdSStefano Zampini 1221d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1222d65f70fdSStefano Zampini *B = work_mat[0]; 1223d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1225d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1226d65f70fdSStefano Zampini PetscFunctionReturn(0); 1227d65f70fdSStefano Zampini } 1228d65f70fdSStefano Zampini 1229d65f70fdSStefano Zampini #undef __FUNCT__ 12305e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12315e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1232aa0d41d4SStefano Zampini { 1233aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12345e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1235d65f70fdSStefano Zampini Mat new_mat; 12365e8657edSStefano Zampini IS is_local,is_global; 1237d65f70fdSStefano Zampini PetscInt local_size; 1238d65f70fdSStefano Zampini PetscBool isseqaij; 1239aa0d41d4SStefano Zampini PetscErrorCode ierr; 1240aa0d41d4SStefano Zampini 1241aa0d41d4SStefano Zampini PetscFunctionBegin; 1242aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12435e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12445e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1245906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1246906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1247d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1248906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1249906d46d4SStefano Zampini 1250906d46d4SStefano Zampini /* check */ 1251906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1252906d46d4SStefano Zampini Vec x,x_change; 1253906d46d4SStefano Zampini PetscReal error; 1254906d46d4SStefano Zampini 12555e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1256906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12575e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12585e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12595e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1260d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12615e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12625e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1263906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1264906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1265906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1266906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1267906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1268906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1269906d46d4SStefano Zampini } 1270906d46d4SStefano Zampini 127122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12729b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 127322d5777bSStefano Zampini if (isseqaij) { 1274d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1275aa0d41d4SStefano Zampini } else { 1276aa0d41d4SStefano Zampini Mat work_mat; 1277aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1278d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1279aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1280aa0d41d4SStefano Zampini } 12813301b35fSStefano Zampini if (matis->A->symmetric_set) { 12823301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1283e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12843301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1285e496cd5dSStefano Zampini #endif 12863301b35fSStefano Zampini } 128745a1bb75SStefano Zampini /* 128845a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1289d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 129045a1bb75SStefano Zampini */ 1291d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1292aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1293aa0d41d4SStefano Zampini } 1294aa0d41d4SStefano Zampini 1295aa0d41d4SStefano Zampini #undef __FUNCT__ 1296a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 12978ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1298a64d13efSStefano Zampini { 1299a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1300a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1301d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 130253892102SStefano Zampini PetscInt *idx_R_local=NULL; 13033a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13043a50541eSStefano Zampini PetscInt vbs,bs; 13056816873aSStefano Zampini PetscBT bitmask=NULL; 1306a64d13efSStefano Zampini PetscErrorCode ierr; 1307a64d13efSStefano Zampini 1308a64d13efSStefano Zampini PetscFunctionBegin; 1309b23d619eSStefano Zampini /* 1310b23d619eSStefano Zampini No need to setup local scatters if 1311b23d619eSStefano Zampini - primal space is unchanged 1312b23d619eSStefano Zampini AND 1313b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1314b23d619eSStefano Zampini AND 1315b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1316b23d619eSStefano Zampini */ 1317b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1318f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1319f4ddd8eeSStefano Zampini } 1320f4ddd8eeSStefano Zampini /* destroy old objects */ 1321f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1322f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1323f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1324a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1325b371cd4fSStefano Zampini n_B = pcis->n_B; 1326b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1327b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13286816873aSStefano Zampini 13296816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13306816873aSStefano Zampini 133153892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13326816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 133353892102SStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13344641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13354641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13360e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13374641a718SStefano Zampini } 13383a50541eSStefano Zampini 1339a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13404641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13416816873aSStefano Zampini idx_R_local[n_R++] = i; 1342a64d13efSStefano Zampini } 1343a64d13efSStefano Zampini } 134453892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13456816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13466816873aSStefano Zampini 134753892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 134853892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13496816873aSStefano Zampini } 13503a50541eSStefano Zampini 13513a50541eSStefano Zampini /* Block code */ 13523a50541eSStefano Zampini vbs = 1; 13533a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13543a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13553a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13563a50541eSStefano Zampini PetscInt *vary; 1357d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1358785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13593a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1360d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1361d3df7717SStefano 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 */ 13620e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1363d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13643a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13653a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13663a50541eSStefano Zampini break; 13673a50541eSStefano Zampini } 13683a50541eSStefano Zampini } 1369d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1370d3df7717SStefano Zampini } else { 1371d3df7717SStefano Zampini /* Verify directly the R set */ 1372d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1373d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1374d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1375d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1376d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1377d3df7717SStefano Zampini break; 1378d3df7717SStefano Zampini } 1379d3df7717SStefano Zampini } 1380d3df7717SStefano Zampini } 1381d3df7717SStefano Zampini } 13823a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13833a50541eSStefano Zampini vbs = bs; 13843a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13853a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13863a50541eSStefano Zampini } 13873a50541eSStefano Zampini } 13883a50541eSStefano Zampini } 13893a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 139053892102SStefano Zampini if (sub_schurs->reuse_mumps) { 139153892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 139253892102SStefano Zampini 139353892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 139453892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 139553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 139653892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 139753892102SStefano Zampini } else { 13983a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 139953892102SStefano Zampini } 1400a64d13efSStefano Zampini 1401a64d13efSStefano Zampini /* print some info if requested */ 1402a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1403a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1404a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14050fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1406a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1407a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14083a50541eSStefano 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,pcbddc->local_primal_size);CHKERRQ(ierr); 1409a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1410a64d13efSStefano Zampini } 1411a64d13efSStefano Zampini 1412a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14136816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14146816873aSStefano Zampini IS is_aux1,is_aux2; 14156816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14166816873aSStefano Zampini 14173a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1418854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1419854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1420a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14214641a718SStefano Zampini for (i=0; i<n_D; i++) { 14224641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14234641a718SStefano Zampini } 1424a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1425a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14264641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14274641a718SStefano Zampini aux_array1[j++] = i; 1428a64d13efSStefano Zampini } 1429a64d13efSStefano Zampini } 1430a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1431a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1432a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14334641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14344641a718SStefano Zampini aux_array2[j++] = i; 1435a64d13efSStefano Zampini } 1436a64d13efSStefano Zampini } 1437a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1438a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1439a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1440a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1441a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1442a64d13efSStefano Zampini 14438eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1444785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1445a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14464641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14474641a718SStefano Zampini aux_array1[j++] = i; 1448a64d13efSStefano Zampini } 1449a64d13efSStefano Zampini } 1450a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1451a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1452a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1453a64d13efSStefano Zampini } 14544641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14553a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1456d62866d3SStefano Zampini } else { 145753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14586816873aSStefano Zampini IS tis; 14596816873aSStefano Zampini PetscInt schur_size; 14606816873aSStefano Zampini 146153892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14626816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 146353892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14646816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14656816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14666816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14676816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14686816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1469d62866d3SStefano Zampini } 1470d62866d3SStefano Zampini } 1471a64d13efSStefano Zampini PetscFunctionReturn(0); 1472a64d13efSStefano Zampini } 1473a64d13efSStefano Zampini 1474304d26faSStefano Zampini 1475304d26faSStefano Zampini #undef __FUNCT__ 1476304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1477684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1478304d26faSStefano Zampini { 1479304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1480304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1481304d26faSStefano Zampini PC pc_temp; 1482304d26faSStefano Zampini Mat A_RR; 1483f4ddd8eeSStefano Zampini MatReuse reuse; 1484304d26faSStefano Zampini PetscScalar m_one = -1.0; 1485304d26faSStefano Zampini PetscReal value; 1486*04708bb6SStefano Zampini PetscInt n_D,n_R; 14879577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1488304d26faSStefano Zampini PetscErrorCode ierr; 1489e604994aSStefano Zampini /* prefixes stuff */ 1490312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1491e604994aSStefano Zampini size_t len; 1492304d26faSStefano Zampini 1493304d26faSStefano Zampini PetscFunctionBegin; 1494304d26faSStefano Zampini 1495e604994aSStefano Zampini /* compute prefixes */ 1496e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1497e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1498e604994aSStefano Zampini if (!pcbddc->current_level) { 1499e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1500e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1501e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1502e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1503e604994aSStefano Zampini } else { 1504e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1505312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1506e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1507e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1508312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1509312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 151034d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 151134d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1512e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1513e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1514e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1515e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1516e604994aSStefano Zampini } 1517e604994aSStefano Zampini 1518304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1519684f6988SStefano Zampini if (dirichlet) { 1520d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15213301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15223301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1523964fefecSStefano Zampini } 1524ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1525964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1526304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1527304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1528304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1529304d26faSStefano Zampini /* default */ 1530304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1531e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15329577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1533304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15349577ea80SStefano Zampini if (issbaij) { 15359577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15369577ea80SStefano Zampini } else { 1537304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15389577ea80SStefano Zampini } 1539304d26faSStefano Zampini /* Allow user's customization */ 1540304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1541304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1542304d26faSStefano Zampini } 1543d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1544d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1545d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1546d62866d3SStefano Zampini 1547d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1548d5574798SStefano Zampini } 1549304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1550304d26faSStefano Zampini if (!n_D) { 1551304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1552304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1553304d26faSStefano Zampini } 1554304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1555304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1556304d26faSStefano Zampini /* set ksp_D into pcis data */ 1557304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1558304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1559304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1560684f6988SStefano Zampini } 1561304d26faSStefano Zampini 1562304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1563684f6988SStefano Zampini A_RR = 0; 1564684f6988SStefano Zampini if (neumann) { 1565d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1566*04708bb6SStefano Zampini PetscInt ibs,mbs; 1567*04708bb6SStefano Zampini PetscBool issbaij; 1568*04708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1569f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15708ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1571f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1572f4ddd8eeSStefano Zampini PetscInt nn_R; 157381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1574f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1575f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1576f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1577f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1578f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1579f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1580f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1581727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1582f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1583f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1584f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1585f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1586f4ddd8eeSStefano Zampini } 1587f4ddd8eeSStefano Zampini } 1588f4ddd8eeSStefano Zampini /* last check */ 1589d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1590f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1591f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1592f4ddd8eeSStefano Zampini } 1593f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1594f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1595f4ddd8eeSStefano Zampini } 15966816873aSStefano Zampini /* extract A_RR */ 15976816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 15986816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 1599*04708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1600*04708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 1601*04708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 1602*04708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1603*04708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16046816873aSStefano Zampini } else { 1605*04708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16066816873aSStefano Zampini } 1607*04708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 1608*04708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 1609*04708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1610*04708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1611*04708bb6SStefano Zampini } else { 1612*04708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 1613*04708bb6SStefano Zampini } 1614*04708bb6SStefano Zampini } 1615*04708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 1616*04708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16173301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16183301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16196816873aSStefano Zampini } 16206816873aSStefano Zampini } else { 16216816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16226816873aSStefano Zampini 16236816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16246816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16256816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16266816873aSStefano Zampini } 1627f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1628304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1629304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1630304d26faSStefano Zampini /* default */ 1631304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1632e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1633304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16349577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16359577ea80SStefano Zampini if (issbaij) { 16369577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16379577ea80SStefano Zampini } else { 1638304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16399577ea80SStefano Zampini } 1640304d26faSStefano Zampini /* Allow user's customization */ 1641304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1642304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1643304d26faSStefano Zampini } 1644d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1645304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1646304d26faSStefano Zampini if (!n_R) { 1647304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1648304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1649304d26faSStefano Zampini } 1650d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1651d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1652d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1653d62866d3SStefano Zampini 1654d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1655d62866d3SStefano Zampini } 1656304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1657304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1658684f6988SStefano Zampini } 16596816873aSStefano Zampini /* free Neumann problem's matrix */ 16606816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1661304d26faSStefano Zampini 1662304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16630fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1664684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1665684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1666684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1667684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1668684f6988SStefano Zampini } 1669684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16700fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16710fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16720fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16730fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16740fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1675304d26faSStefano Zampini /* need to be adapted? */ 1676b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1677b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1678b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1679304d26faSStefano Zampini /* print info */ 1680304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1681e604994aSStefano 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); 1682304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1683304d26faSStefano Zampini } 1684b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16858ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1686304d26faSStefano Zampini } 1687684f6988SStefano Zampini } 1688684f6988SStefano Zampini if (neumann) { /* Neumann */ 16896816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16900fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16910fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16920fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16930fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16940fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1695304d26faSStefano Zampini /* need to be adapted? */ 1696b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1697b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1698304d26faSStefano Zampini /* print info */ 1699304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1700e604994aSStefano 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); 1701304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1702304d26faSStefano Zampini } 1703b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 17048ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1705304d26faSStefano Zampini } 17060fccc4e9SStefano Zampini } 1707684f6988SStefano Zampini } 1708304d26faSStefano Zampini PetscFunctionReturn(0); 1709304d26faSStefano Zampini } 1710304d26faSStefano Zampini 1711304d26faSStefano Zampini #undef __FUNCT__ 1712ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 171380677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1714674ae819SStefano Zampini { 1715674ae819SStefano Zampini PetscErrorCode ierr; 1716674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1717be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1718674ae819SStefano Zampini 1719674ae819SStefano Zampini PetscFunctionBegin; 1720be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1722be83ff47SStefano Zampini } 172380677318SStefano Zampini if (!pcbddc->switch_static) { 172480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 172580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 172680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 172720c7b377SStefano Zampini } 1728be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173120c7b377SStefano Zampini } else { 1732be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1733be83ff47SStefano Zampini 173453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 173553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1736be83ff47SStefano Zampini } 1737be83ff47SStefano Zampini } else { 173880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 174380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 174480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 174580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174720c7b377SStefano Zampini } 1748674ae819SStefano Zampini } 1749be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 175080677318SStefano Zampini if (applytranspose) { 175180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 175280677318SStefano Zampini } else { 175380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 175480677318SStefano Zampini } 175553892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1756be83ff47SStefano Zampini } else { 1757be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1758be83ff47SStefano Zampini 1759be83ff47SStefano Zampini if (applytranspose) { 176053892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1761be83ff47SStefano Zampini } else { 176253892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1763be83ff47SStefano Zampini } 176453892102SStefano Zampini #endif 1765be83ff47SStefano Zampini } 176680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 176780677318SStefano Zampini if (!pcbddc->switch_static) { 1768be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1771be83ff47SStefano Zampini } else { 1772be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1773be83ff47SStefano Zampini 177453892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 177553892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1776be83ff47SStefano Zampini } 177780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 178080677318SStefano Zampini } 178180677318SStefano Zampini } else { 178280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 178980677318SStefano Zampini } 179080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179480677318SStefano Zampini } 1795674ae819SStefano Zampini PetscFunctionReturn(0); 1796674ae819SStefano Zampini } 1797674ae819SStefano Zampini 1798dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1799674ae819SStefano Zampini #undef __FUNCT__ 1800674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1801dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1802674ae819SStefano Zampini { 1803674ae819SStefano Zampini PetscErrorCode ierr; 1804674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1805674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1806674ae819SStefano Zampini const PetscScalar zero = 0.0; 1807674ae819SStefano Zampini 1808674ae819SStefano Zampini PetscFunctionBegin; 1809dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1810dc359a40SStefano Zampini if (applytranspose) { 1811674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18128eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1813dc359a40SStefano Zampini } else { 1814674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1815674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 181615aaf578SStefano Zampini } 181712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 181812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 181912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182112edc857SStefano Zampini 18229f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 182312edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 182412edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1825964fefecSStefano Zampini Vec rhs,sol; 1826964fefecSStefano Zampini 1827964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1828964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 182912edc857SStefano Zampini if (applytranspose) { 1830964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183112edc857SStefano Zampini } else { 1832964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183312edc857SStefano Zampini } 183412edc857SStefano Zampini } 1835674ae819SStefano Zampini 1836674ae819SStefano Zampini /* Local solution on R nodes */ 183780677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 183880677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18399f00e9b4SStefano Zampini } 1840674ae819SStefano Zampini 18419f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18429f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 184312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1844674ae819SStefano Zampini 1845674ae819SStefano Zampini /* Sum contributions from two levels */ 1846dc359a40SStefano Zampini if (applytranspose) { 1847dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1848dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1849dc359a40SStefano Zampini } else { 1850674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18518eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1852dc359a40SStefano Zampini } 1853674ae819SStefano Zampini PetscFunctionReturn(0); 1854674ae819SStefano Zampini } 1855674ae819SStefano Zampini 1856674ae819SStefano Zampini #undef __FUNCT__ 1857674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 185812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1859674ae819SStefano Zampini { 1860674ae819SStefano Zampini PetscErrorCode ierr; 1861674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 186258da7f69SStefano Zampini PetscScalar *array; 186312edc857SStefano Zampini Vec from,to; 1864674ae819SStefano Zampini 1865674ae819SStefano Zampini PetscFunctionBegin; 186612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 186712edc857SStefano Zampini from = pcbddc->coarse_vec; 186812edc857SStefano Zampini to = pcbddc->vec1_P; 186912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 187012edc857SStefano Zampini Vec tvec; 187158da7f69SStefano Zampini 187258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 187358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 187412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 187558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 187658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 187758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 187812edc857SStefano Zampini } 187912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 188012edc857SStefano Zampini from = pcbddc->vec1_P; 188112edc857SStefano Zampini to = pcbddc->coarse_vec; 188212edc857SStefano Zampini } 188312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1884674ae819SStefano Zampini PetscFunctionReturn(0); 1885674ae819SStefano Zampini } 1886674ae819SStefano Zampini 1887674ae819SStefano Zampini #undef __FUNCT__ 1888674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 188912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1890674ae819SStefano Zampini { 1891674ae819SStefano Zampini PetscErrorCode ierr; 1892674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 189358da7f69SStefano Zampini PetscScalar *array; 189412edc857SStefano Zampini Vec from,to; 1895674ae819SStefano Zampini 1896674ae819SStefano Zampini PetscFunctionBegin; 189712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 189812edc857SStefano Zampini from = pcbddc->coarse_vec; 189912edc857SStefano Zampini to = pcbddc->vec1_P; 190012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 190112edc857SStefano Zampini from = pcbddc->vec1_P; 190212edc857SStefano Zampini to = pcbddc->coarse_vec; 190312edc857SStefano Zampini } 190412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 190512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 190612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 190712edc857SStefano Zampini Vec tvec; 190858da7f69SStefano Zampini 190912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 191058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 191158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 191258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 191358da7f69SStefano Zampini } 191458da7f69SStefano Zampini } else { 191558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 191658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 191712edc857SStefano Zampini } 191812edc857SStefano Zampini } 1919674ae819SStefano Zampini PetscFunctionReturn(0); 1920674ae819SStefano Zampini } 1921674ae819SStefano Zampini 1922984c4197SStefano Zampini /* uncomment for testing purposes */ 1923984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1924674ae819SStefano Zampini #undef __FUNCT__ 1925674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1926674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1927674ae819SStefano Zampini { 1928674ae819SStefano Zampini PetscErrorCode ierr; 1929674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1930674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1931674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1932984c4197SStefano Zampini /* one and zero */ 1933984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1934984c4197SStefano Zampini /* space to store constraints and their local indices */ 19359162d606SStefano Zampini PetscScalar *constraints_data; 19369162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19379162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19389162d606SStefano Zampini PetscInt *constraints_n; 1939984c4197SStefano Zampini /* iterators */ 1940b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1941984c4197SStefano Zampini /* BLAS integers */ 1942e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1943e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1944c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1945727cdba6SStefano Zampini /* reuse */ 19460e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19470e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1948984c4197SStefano Zampini /* change of basis */ 1949b3d85658SStefano Zampini PetscBool qr_needed; 19509162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1951984c4197SStefano Zampini /* auxiliary stuff */ 195264efe560SStefano Zampini PetscInt *nnz,*is_indices; 19538a0068c3SStefano Zampini PetscInt ncc; 1954984c4197SStefano Zampini /* some quantities */ 195545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1956a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1957984c4197SStefano Zampini 1958674ae819SStefano Zampini PetscFunctionBegin; 19598e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19608e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19618e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1962088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1963088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19640e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19650e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19660e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19670e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19680e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1969088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1970cf5a6209SStefano Zampini 1971cf5a6209SStefano Zampini /* print some info */ 1972cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1973cf5a6209SStefano Zampini IS vertices; 1974cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1975cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1976cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1977cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1978cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1979cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1980cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1981fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1982fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1983cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1984cf5a6209SStefano Zampini } 1985cf5a6209SStefano Zampini 1986cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19879162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1988cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1989cf5a6209SStefano Zampini const Vec *nearnullvecs; 1990cf5a6209SStefano Zampini Vec *localnearnullsp; 1991cf5a6209SStefano Zampini PetscScalar *array; 1992cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1993cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1994cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1995b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1996cf5a6209SStefano Zampini PetscScalar *work; 1997cf5a6209SStefano Zampini PetscReal *singular_vals; 1998cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1999cf5a6209SStefano Zampini PetscReal *rwork; 2000cf5a6209SStefano Zampini #endif 2001cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2002cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2003cf5a6209SStefano Zampini #else 2004964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2005964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2006cf5a6209SStefano Zampini #endif 2007cf5a6209SStefano Zampini 2008674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2009d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2010d06fc5fdSStefano Zampini /* free unneeded index sets */ 2011d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2012d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2013674ae819SStefano Zampini } 2014d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2015d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2016d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2017d06fc5fdSStefano Zampini } 2018d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2019d06fc5fdSStefano Zampini n_ISForEdges = 0; 2020d06fc5fdSStefano Zampini } 2021d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2022d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2023d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2024d06fc5fdSStefano Zampini } 2025d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2026d06fc5fdSStefano Zampini n_ISForFaces = 0; 2027d06fc5fdSStefano Zampini } 202870022509SStefano Zampini 202970022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 203070022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 203170022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 203270022509SStefano Zampini if (pcbddc->NullSpace) { 203370022509SStefano Zampini PetscBool tbool[2],gbool[2]; 203470022509SStefano Zampini 203570022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2036b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2037d06fc5fdSStefano Zampini if (!ISForEdges) { 2038d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2039d06fc5fdSStefano Zampini } 2040b8ffe317SStefano Zampini } 2041d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2042d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2043d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2044d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2045d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 204698a51de6SStefano Zampini } 204770022509SStefano Zampini #endif 204808122e43SStefano Zampini 2049674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2050674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2051674ae819SStefano Zampini if (nearnullsp) { 2052674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2053f4ddd8eeSStefano Zampini /* remove any stored info */ 2054f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2055f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2056f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2057f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2058f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2059473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2060f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2061f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2062f4ddd8eeSStefano Zampini } 2063984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2064984c4197SStefano Zampini nnsp_size = 0; 2065674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2066674ae819SStefano Zampini } 2067984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2068984c4197SStefano Zampini max_constraints = nnsp_size; 2069984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2070984c4197SStefano Zampini 2071674ae819SStefano Zampini /* 2072674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20739162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20749162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20759162d606SStefano Zampini There can be multiple constraints per connected component 2076674ae819SStefano Zampini */ 2077674ae819SStefano Zampini n_vertices = 0; 2078674ae819SStefano Zampini if (ISForVertices) { 2079674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2080674ae819SStefano Zampini } 20819162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20829162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20839162d606SStefano Zampini 20849162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20859162d606SStefano Zampini total_counts *= max_constraints; 2086674ae819SStefano Zampini total_counts += n_vertices; 20874641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20889162d606SStefano Zampini 2089674ae819SStefano Zampini total_counts = 0; 2090674ae819SStefano Zampini max_size_of_constraint = 0; 2091674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20929162d606SStefano Zampini IS used_is; 2093674ae819SStefano Zampini if (i<n_ISForEdges) { 20949162d606SStefano Zampini used_is = ISForEdges[i]; 2095674ae819SStefano Zampini } else { 20969162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2097674ae819SStefano Zampini } 20989162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2099674ae819SStefano Zampini total_counts += j; 2100674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2101674ae819SStefano Zampini } 21029162d606SStefano 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); 21039162d606SStefano Zampini 2104984c4197SStefano Zampini /* get local part of global near null space vectors */ 2105785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2106984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2107984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2108984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2109984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2110984c4197SStefano Zampini } 2111674ae819SStefano Zampini 2112242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2113242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2114a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2115242a89d7SStefano Zampini 2116984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2117a773dcb8SStefano Zampini if (!skip_lapack) { 2118674ae819SStefano Zampini PetscScalar temp_work; 2119911cabfeSStefano Zampini 2120674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2121984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2122785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2123785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2124785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2125674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2126785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2127674ae819SStefano Zampini #endif 2128674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2129c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2130c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2131674ae819SStefano Zampini lwork = -1; 2132674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2133674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2134c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2135674ae819SStefano Zampini #else 2136c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2137674ae819SStefano Zampini #endif 2138674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2139984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2140674ae819SStefano Zampini #else /* on missing GESVD */ 2141674ae819SStefano Zampini /* SVD */ 2142674ae819SStefano Zampini PetscInt max_n,min_n; 2143674ae819SStefano Zampini max_n = max_size_of_constraint; 2144984c4197SStefano Zampini min_n = max_constraints; 2145984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2146674ae819SStefano Zampini min_n = max_size_of_constraint; 2147984c4197SStefano Zampini max_n = max_constraints; 2148674ae819SStefano Zampini } 2149785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2150674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2151785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2152674ae819SStefano Zampini #endif 2153674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2154674ae819SStefano Zampini lwork = -1; 2155e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2156e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2157b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2158674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2159674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21609162d606SStefano 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)); 2161674ae819SStefano Zampini #else 21629162d606SStefano 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)); 2163674ae819SStefano Zampini #endif 2164674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2165984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2166984c4197SStefano Zampini #endif /* on missing GESVD */ 2167674ae819SStefano Zampini /* Allocate optimal workspace */ 2168674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2169854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2170674ae819SStefano Zampini } 2171674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2172674ae819SStefano Zampini total_counts = 0; 21739162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21749162d606SStefano Zampini constraints_data_ptr[0] = 0; 2175674ae819SStefano Zampini /* vertices */ 21769162d606SStefano Zampini if (n_vertices) { 2177674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21789162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21799162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2180674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21819162d606SStefano Zampini constraints_n[total_counts] = 1; 21829162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21839162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21849162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2185674ae819SStefano Zampini total_counts++; 2186674ae819SStefano Zampini } 2187674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2188984c4197SStefano Zampini PetscBool used_vertex; 2189674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2190674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2191674ae819SStefano Zampini k = 0; 2192674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2193984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2194984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21959162d606SStefano Zampini constraints_n[total_counts] = 1; 21969162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21979162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21989162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21999162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2200674ae819SStefano Zampini total_counts++; 2201674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2202674ae819SStefano Zampini } 2203984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2204674ae819SStefano Zampini k++; 2205674ae819SStefano Zampini } 2206674ae819SStefano Zampini } 2207674ae819SStefano Zampini } 2208674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2209674ae819SStefano Zampini n_vertices = total_counts; 2210674ae819SStefano Zampini } 2211984c4197SStefano Zampini 2212674ae819SStefano Zampini /* edges and faces */ 22139162d606SStefano Zampini total_counts_cc = total_counts; 2214911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22159162d606SStefano Zampini IS used_is; 22169162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22179162d606SStefano Zampini 2218911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22199162d606SStefano Zampini used_is = ISForEdges[ncc]; 2220984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2221674ae819SStefano Zampini } else { 22229162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2223984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2224674ae819SStefano Zampini } 2225674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22269162d606SStefano Zampini 22279162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22289162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2229984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2230984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2231674ae819SStefano Zampini if (nnsp_has_cnst) { 22325b08dc53SStefano Zampini PetscScalar quad_value; 22339162d606SStefano Zampini 22349162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22359162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22369162d606SStefano Zampini 2237a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2238674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2239a773dcb8SStefano Zampini } else { 2240a773dcb8SStefano Zampini quad_value = 1.0; 2241a773dcb8SStefano Zampini } 2242674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22439162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2244674ae819SStefano Zampini } 22459162d606SStefano Zampini temp_constraints++; 2246674ae819SStefano Zampini total_counts++; 2247674ae819SStefano Zampini } 2248674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2249984c4197SStefano Zampini PetscReal real_value; 22509162d606SStefano Zampini PetscScalar *ptr_to_data; 22519162d606SStefano Zampini 2252984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22539162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2254674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22559162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2256674ae819SStefano Zampini } 2257984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2258984c4197SStefano Zampini /* check if array is null on the connected component */ 2259e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22609162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22615b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2262674ae819SStefano Zampini temp_constraints++; 2263674ae819SStefano Zampini total_counts++; 22649162d606SStefano Zampini if (!idxs_copied) { 22659162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22669162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2267674ae819SStefano Zampini } 2268674ae819SStefano Zampini } 22699162d606SStefano Zampini } 22709162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 227145a1bb75SStefano Zampini valid_constraints = temp_constraints; 2272eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2273a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22749162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22759162d606SStefano Zampini 22769162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2277a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22789162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2279a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22809162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2281a773dcb8SStefano Zampini } else { /* perform SVD */ 2282984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22839162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2284674ae819SStefano Zampini 2285674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2286984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2287984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2288984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2289984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2290984c4197SStefano Zampini from that computed using LAPACKgesvd 2291984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2292984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2293984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2294674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2295e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2296984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2297674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2298674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22999162d606SStefano 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)); 2300674ae819SStefano Zampini } 2301674ae819SStefano Zampini } 2302e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2303e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2304e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2305674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2306c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2307674ae819SStefano Zampini #else 2308c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2309674ae819SStefano Zampini #endif 2310674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2311984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2312984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2313674ae819SStefano Zampini j = 0; 2314984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2315674ae819SStefano Zampini total_counts = total_counts-j; 231645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2317e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2318c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2319c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2320c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2321c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2322c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2323c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2324674ae819SStefano Zampini if (j<temp_constraints) { 2325984c4197SStefano Zampini PetscInt ii; 2326984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2327674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23289162d606SStefano 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)); 2329674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2330984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2331674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23329162d606SStefano 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]; 2333674ae819SStefano Zampini } 2334674ae819SStefano Zampini } 2335674ae819SStefano Zampini } 2336674ae819SStefano Zampini #else /* on missing GESVD */ 2337e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2338e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2339b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2340674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2341674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23429162d606SStefano 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)); 2343674ae819SStefano Zampini #else 23449162d606SStefano 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)); 2345674ae819SStefano Zampini #endif 2346984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2347674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2348984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2349e310c8b4SStefano Zampini k = temp_constraints; 2350e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2351674ae819SStefano Zampini j = 0; 2352e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 235345a1bb75SStefano Zampini valid_constraints = k-j; 2354911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2355984c4197SStefano Zampini #endif /* on missing GESVD */ 2356674ae819SStefano Zampini } 2357a773dcb8SStefano Zampini } 23589162d606SStefano Zampini /* update pointers information */ 23599162d606SStefano Zampini if (valid_constraints) { 23609162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23619162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23629162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23639162d606SStefano Zampini /* set change_of_basis flag */ 236445a1bb75SStefano Zampini if (boolforchange) { 2365b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23669162d606SStefano Zampini } 2367b3d85658SStefano Zampini total_counts_cc++; 236845a1bb75SStefano Zampini } 2369674ae819SStefano Zampini } 2370984c4197SStefano Zampini /* free workspace */ 23718f1c130eSStefano Zampini if (!skip_lapack) { 2372984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2373984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2374984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2375984c4197SStefano Zampini #endif 2376984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2377984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2378984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2379984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2380984c4197SStefano Zampini #endif 2381984c4197SStefano Zampini } 2382984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2383984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2384984c4197SStefano Zampini } 2385984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2386cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2387cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2388cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2389cf5a6209SStefano Zampini } 2390cf5a6209SStefano Zampini if (n_ISForFaces) { 2391cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2392cf5a6209SStefano Zampini } 2393cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2394cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2395cf5a6209SStefano Zampini } 2396cf5a6209SStefano Zampini if (n_ISForEdges) { 2397cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2398cf5a6209SStefano Zampini } 2399cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 240008122e43SStefano Zampini } else { 240108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 240208122e43SStefano Zampini 240308122e43SStefano Zampini total_counts = 0; 240408122e43SStefano Zampini n_vertices = 0; 2405d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2406d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 240708122e43SStefano Zampini } 240808122e43SStefano Zampini max_constraints = 0; 24099162d606SStefano Zampini total_counts_cc = 0; 241008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 241108122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24129162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 241308122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 241408122e43SStefano Zampini } 24159162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24169162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24179162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24189162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 241974d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24209162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24219162d606SStefano Zampini total_counts_cc = 0; 24229162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24239162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24249162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 242508122e43SStefano Zampini } 242608122e43SStefano Zampini } 24279162d606SStefano Zampini #if 0 24289162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24299162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24309162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24319162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24329162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24339162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24349162d606SStefano Zampini } 24359162d606SStefano Zampini printf("\n"); 24369162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24379162d606SStefano Zampini } 24381b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24398bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24401b968477SStefano Zampini } 24411b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24428bec7fa6SStefano 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]); 24431b968477SStefano Zampini } 244408122e43SStefano Zampini #endif 244508122e43SStefano Zampini 24468bec7fa6SStefano Zampini max_size_of_constraint = 0; 24479162d606SStefano 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]); 24489162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 244908122e43SStefano Zampini /* Change of basis */ 2450b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 245108122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 245208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 245308122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2454b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 245508122e43SStefano Zampini } 245608122e43SStefano Zampini } 245708122e43SStefano Zampini } 245808122e43SStefano Zampini } 2459b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24600e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 246108122e43SStefano Zampini 24629162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24639162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24649162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24659162d606SStefano 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); 246608122e43SStefano Zampini } 2467984c4197SStefano Zampini 2468674ae819SStefano Zampini /* Create constraint matrix */ 2469674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 247016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2471984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2472984c4197SStefano Zampini 2473984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 247474d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 247574d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 247674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2477984c4197SStefano Zampini total_primal_vertices=0; 2478b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24799162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24809162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2481984c4197SStefano Zampini if (size_of_constraint == 1) { 24829162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2483b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 248464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24859162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24869162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24879162d606SStefano Zampini } 2488b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 248974d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 249074d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 249174d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 249274d5cdf7SStefano Zampini } 2493b3d85658SStefano Zampini } else { 2494b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2495984c4197SStefano Zampini } 2496984c4197SStefano Zampini } 2497b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2498b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 249970022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 250070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2501b3d85658SStefano Zampini 2502b3d85658SStefano Zampini ierr = PetscMalloc2(pcbddc->local_primal_size_cc,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc,&pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 25030e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25040e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 250570022509SStefano Zampini 2506984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 250774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2508785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2509984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 251074d5cdf7SStefano Zampini 2511984c4197SStefano Zampini j = total_primal_vertices; 251274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2513b3d85658SStefano Zampini cum = total_primal_vertices; 25149162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25154641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2516b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2517b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2518b3d85658SStefano Zampini cum++; 25199162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 252074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 252174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 252274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 252374d5cdf7SStefano Zampini } 25249162d606SStefano Zampini j += constraints_n[i]; 2525674ae819SStefano Zampini } 2526674ae819SStefano Zampini } 2527674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2528674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2529088faed8SStefano Zampini 2530674ae819SStefano Zampini /* set values in constraint matrix */ 2531984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25320e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2533674ae819SStefano Zampini } 2534984c4197SStefano Zampini total_counts = total_primal_vertices; 25359162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25364641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25379162d606SStefano Zampini PetscInt *cols; 25389162d606SStefano Zampini 25399162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25409162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25419162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25429162d606SStefano Zampini PetscInt row = total_counts+k; 25439162d606SStefano Zampini PetscScalar *vals; 25449162d606SStefano Zampini 25459162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25469162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25479162d606SStefano Zampini } 25489162d606SStefano Zampini total_counts += constraints_n[i]; 2549674ae819SStefano Zampini } 2550674ae819SStefano Zampini } 2551674ae819SStefano Zampini /* assembling */ 2552674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2553674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2554088faed8SStefano Zampini 2555984c4197SStefano Zampini /* 255645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2557984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2558984c4197SStefano Zampini */ 2559674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2560674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2561026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2562984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2563984c4197SStefano Zampini /* working stuff for GEQRF */ 256481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2565984c4197SStefano Zampini PetscBLASInt lqr_work; 2566984c4197SStefano Zampini /* working stuff for UNGQR */ 2567984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2568984c4197SStefano Zampini PetscBLASInt lgqr_work; 2569984c4197SStefano Zampini /* working stuff for TRTRS */ 2570984c4197SStefano Zampini PetscScalar *trs_rhs; 25713f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2572984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2573984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2574984c4197SStefano Zampini PetscScalar *start_vals; 2575984c4197SStefano Zampini /* working stuff for values insertion */ 25764641a718SStefano Zampini PetscBT is_primal; 257764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2578906d46d4SStefano Zampini /* matrix sizes */ 2579906d46d4SStefano Zampini PetscInt global_size,local_size; 2580906d46d4SStefano Zampini /* temporary change of basis */ 2581906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2582cf5a6209SStefano Zampini /* extra space for debugging */ 2583cf5a6209SStefano Zampini PetscScalar *dbg_work; 2584984c4197SStefano Zampini 2585906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2586906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 258716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2588bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2589906d46d4SStefano Zampini /* nonzeros for local mat */ 2590bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2591bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25929162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2593a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25949162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2595a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25969162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2597a717540cSStefano Zampini } else { 25989162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25999162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2600a717540cSStefano Zampini } 2601a717540cSStefano Zampini } 2602a717540cSStefano Zampini } 2603906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2604bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2605a717540cSStefano Zampini /* Set initial identity in the matrix */ 2606bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2607906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2608a717540cSStefano Zampini } 2609a717540cSStefano Zampini 2610a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2611a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2612a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2613a717540cSStefano Zampini } 2614a717540cSStefano Zampini 2615a717540cSStefano Zampini 2616a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2617a717540cSStefano Zampini /* 2618a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2619a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2620a717540cSStefano Zampini 2621a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2622a717540cSStefano Zampini 2623a6b551f4SStefano 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) 2624a6b551f4SStefano Zampini 2625a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2626a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2627a717540cSStefano Zampini | ... | 2628a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2629a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2630a717540cSStefano Zampini 2631a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2632a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2633a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2634a6b551f4SStefano Zampini 2635a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2636a717540cSStefano Zampini */ 2637a717540cSStefano Zampini if (qr_needed) { 2638984c4197SStefano Zampini /* space to store Q */ 2639854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2640984c4197SStefano Zampini /* first we issue queries for optimal work */ 26413f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2644984c4197SStefano Zampini lqr_work = -1; 26453f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2646984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2647984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2648785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2649984c4197SStefano Zampini lgqr_work = -1; 26503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26513f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26523f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26533f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26543f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26553f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2656984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2657984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2658785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2659984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2660785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2661984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2662785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2663a717540cSStefano Zampini /* allocating workspace for check */ 2664a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2665cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2666a717540cSStefano Zampini } 2667a717540cSStefano Zampini } 2668984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26694641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2670473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26710e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 267239e2fb2aSStefano Zampini if (i != total_primal_vertices) { 267339e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26744641a718SStefano Zampini } 267539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 267639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 267739e2fb2aSStefano Zampini } 267839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2679984c4197SStefano Zampini 2680a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26819162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26829162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26834641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2684984c4197SStefano Zampini /* get constraint info */ 26859162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2686984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2687984c4197SStefano Zampini 2688984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26899162d606SStefano 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); 2690674ae819SStefano Zampini } 2691984c4197SStefano Zampini 2692fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2693a717540cSStefano Zampini 2694a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2695a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26969162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2697a717540cSStefano Zampini } 2698984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26999162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2700984c4197SStefano Zampini 2701984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27023f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27033f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27043f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2705674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27063f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2707984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2708674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2709984c4197SStefano Zampini 2710984c4197SStefano Zampini /* explictly compute R^-T */ 2711984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2712984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27143f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2717984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27183f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2719984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2720984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2721984c4197SStefano Zampini 2722a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27233f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27253f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27263f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2727984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27283f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2729984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2730984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2731984c4197SStefano Zampini 2732984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2733984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2734984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27373f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27393f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27403f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2741984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27429162d606SStefano 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)); 2743984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27449162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2745984c4197SStefano Zampini 2746984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27479162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2748984c4197SStefano Zampini /* insert cols for primal dofs */ 2749984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2750984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27519162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2752906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2753984c4197SStefano Zampini } 2754984c4197SStefano Zampini /* insert cols for dual dofs */ 2755984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27569162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2757984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27589162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2759906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2760984c4197SStefano Zampini j++; 2761674ae819SStefano Zampini } 2762674ae819SStefano Zampini } 2763984c4197SStefano Zampini 2764984c4197SStefano Zampini /* check change of basis */ 2765984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2766984c4197SStefano Zampini PetscInt ii,jj; 2767984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2768c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2769c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2770c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2771c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2772c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2773c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2774984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2775cf5a6209SStefano 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)); 2776984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2777984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2778984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2779cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2780cf5a6209SStefano 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; 2781674ae819SStefano Zampini } 2782674ae819SStefano Zampini } 2783984c4197SStefano Zampini if (!valid_qr) { 278422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2785984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2786984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2787cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2788cf5a6209SStefano 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])); 2789674ae819SStefano Zampini } 2790cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2791cf5a6209SStefano 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])); 2792984c4197SStefano Zampini } 2793984c4197SStefano Zampini } 2794984c4197SStefano Zampini } 2795674ae819SStefano Zampini } else { 279622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2797674ae819SStefano Zampini } 2798674ae819SStefano Zampini } 2799a717540cSStefano Zampini } else { /* simple transformation block */ 2800a717540cSStefano Zampini PetscInt row,col; 2801a6b551f4SStefano Zampini PetscScalar val,norm; 2802a6b551f4SStefano Zampini 2803a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28049162d606SStefano 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)); 2805a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28069162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28079162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2808bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28099162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2810906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28119162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2812a717540cSStefano Zampini } else { 2813a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28149162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2815a717540cSStefano Zampini if (row != col) { 28169162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2817a717540cSStefano Zampini } else { 28189162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2819a717540cSStefano Zampini } 2820906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2821a717540cSStefano Zampini } 2822a717540cSStefano Zampini } 2823a717540cSStefano Zampini } 282498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 282522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2826a717540cSStefano Zampini } 2827674ae819SStefano Zampini } 2828984c4197SStefano Zampini } else { 2829984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28309162d606SStefano 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); 2831674ae819SStefano Zampini } 2832674ae819SStefano Zampini } 2833674ae819SStefano Zampini } 2834a717540cSStefano Zampini 2835a717540cSStefano Zampini /* free workspace */ 2836a717540cSStefano Zampini if (qr_needed) { 2837984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2838cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2839984c4197SStefano Zampini } 2840984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2841984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2842984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2843984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2844984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2845674ae819SStefano Zampini } 2846a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2847906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2848906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2849906d46d4SStefano Zampini 2850906d46d4SStefano Zampini /* assembling of global change of variable */ 2851bbb9e6c6SStefano Zampini { 2852bbb9e6c6SStefano Zampini Mat tmat; 285316f15bc4SStefano Zampini PetscInt bs; 285416f15bc4SStefano Zampini 2855906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2856906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2857bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2858bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2859bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2860bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 286116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 286216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2863906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2866bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2867bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2868bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2869bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2870bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2871bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2872bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2873906d46d4SStefano Zampini } 2874906d46d4SStefano Zampini /* check */ 2875906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2876906d46d4SStefano Zampini PetscReal error; 2877906d46d4SStefano Zampini Vec x,x_change; 2878906d46d4SStefano Zampini 2879906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2880906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2881906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2882906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2883bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2884bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2885bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2886bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2887bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2888906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2889906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2890906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2891906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2892bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2893906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2894906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2895906d46d4SStefano Zampini } 2896b96c3477SStefano Zampini 2897b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2898b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2899b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2900b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2901ac632422SStefano Zampini Mat S_new,tmat; 2902bbb9e6c6SStefano Zampini IS is_all_N; 2903bbb9e6c6SStefano Zampini 2904bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29056816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2906bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2907ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2908b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2909ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2910ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2911ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2912ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2913ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2914b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2915ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2916ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2917ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2918ac632422SStefano Zampini } 2919b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2920b96c3477SStefano Zampini } 2921b96c3477SStefano Zampini } 2922906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2923906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2924b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2925b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2926b9b85e73SStefano Zampini } 2927906d46d4SStefano Zampini 2928906d46d4SStefano Zampini /* set up change of basis context */ 2929906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2930906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2931906d46d4SStefano Zampini 2932906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2933906d46d4SStefano Zampini PetscInt global_size,local_size; 2934906d46d4SStefano Zampini 2935906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2936906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2937906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2938906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2939906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2940906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2941906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2942906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2943906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2944906d46d4SStefano Zampini } else { 2945906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2946906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2947906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2948906d46d4SStefano Zampini } 2949906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2950906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2951906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2952906d46d4SStefano Zampini } else { 2953906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2954906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2955906d46d4SStefano Zampini } 2956906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2957906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2958906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2959906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2960b9b85e73SStefano Zampini } 2961a717540cSStefano Zampini 2962727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2963727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2964727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2965aff50787SStefano 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); 29660e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29670e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2968aff50787SStefano 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); 2969c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2970727cdba6SStefano Zampini } 29710e6343abSStefano Zampini } 29720e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2973727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2974727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2975727cdba6SStefano Zampini 2976a717540cSStefano Zampini /* flush dbg viewer */ 2977b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2978b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2979b8ffe317SStefano Zampini } 2980a717540cSStefano Zampini 2981e310c8b4SStefano Zampini /* free workspace */ 2982a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29834641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 298408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29859162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29869162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 298708122e43SStefano Zampini } else { 29889162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29899162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29909162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 299108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 299208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29939162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29949162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 299508122e43SStefano Zampini } 2996674ae819SStefano Zampini PetscFunctionReturn(0); 2997674ae819SStefano Zampini } 2998674ae819SStefano Zampini 2999674ae819SStefano Zampini #undef __FUNCT__ 3000674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3001674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3002674ae819SStefano Zampini { 3003674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3004674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3005674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30067fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3007674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3008674ae819SStefano Zampini 3009674ae819SStefano Zampini PetscFunctionBegin; 30108e61c736SStefano Zampini /* Reset previously computed graph */ 30118e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3012674ae819SStefano Zampini /* Init local Graph struct */ 30137fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30147fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3015674ae819SStefano Zampini 3016575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3017575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3018575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3019575ad6abSStefano Zampini } 30209577ea80SStefano Zampini 3021674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30224d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30234d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30244d379d7bSStefano Zampini PetscInt nvtxs; 3025e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3026674ae819SStefano Zampini 30274d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30282fffb893SStefano Zampini 30292fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30302fffb893SStefano Zampini if (flg_row) { 30314d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3032b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30332fffb893SStefano Zampini } 30342fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30359b28b941SStefano 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 */ 30364d379d7bSStefano Zampini IS is_dummy; 30374d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30384d379d7bSStefano Zampini PetscInt j,sum; 30394d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30404d379d7bSStefano Zampini const PetscInt *idxs; 30414d379d7bSStefano Zampini PCBDDCGraph graph; 30424d379d7bSStefano Zampini PetscBT is_on_boundary; 30434d379d7bSStefano Zampini 30444d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30454d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30464d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30474d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30487fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30494d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3050e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3051e496cd5dSStefano Zampini if (flg_row) { 30524d379d7bSStefano Zampini graph->xadj = xadj; 30534d379d7bSStefano Zampini graph->adjncy = adjncy; 3054e496cd5dSStefano Zampini } 30554d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30564d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3057e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30584d379d7bSStefano Zampini 30594d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30609b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30614d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30624d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30634d379d7bSStefano Zampini } 30644d379d7bSStefano Zampini } 30654d379d7bSStefano Zampini 3066e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30674d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30684d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30694d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30704d379d7bSStefano Zampini } 30714d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30724d379d7bSStefano Zampini 3073e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30744d379d7bSStefano Zampini sum = 0; 30754d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30764d379d7bSStefano Zampini PetscInt sizecc = 0; 30774d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30784d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30794d379d7bSStefano Zampini sizecc++; 30804d379d7bSStefano Zampini } 30814d379d7bSStefano Zampini } 30824d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30834d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30844d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30854d379d7bSStefano Zampini } 30864d379d7bSStefano Zampini } 30874d379d7bSStefano Zampini sum += sizecc*sizecc; 30884d379d7bSStefano Zampini } 30894d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30904d379d7bSStefano Zampini sum = 0; 3091e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30924d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30934d379d7bSStefano Zampini cxadj[i] = sum; 30944d379d7bSStefano Zampini sum += temp; 30954d379d7bSStefano Zampini } 3096e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30974d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30984d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30994d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31004d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31014d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31024d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31034d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31044d379d7bSStefano Zampini sizecc++; 31054d379d7bSStefano Zampini } 31064d379d7bSStefano Zampini } 31074d379d7bSStefano Zampini } 31084d379d7bSStefano Zampini } 31094d379d7bSStefano Zampini } 31109b28b941SStefano Zampini if (sum) { 3111e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31124d379d7bSStefano Zampini } else { 31134d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31144d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31154d379d7bSStefano Zampini } 31164d379d7bSStefano Zampini graph->xadj = 0; 31174d379d7bSStefano Zampini graph->adjncy = 0; 31184d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31194d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31204d379d7bSStefano Zampini } 3121674ae819SStefano Zampini } 31229b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31239b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31249b28b941SStefano Zampini } 3125674ae819SStefano Zampini 312663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3127674ae819SStefano Zampini vertex_size = 1; 312863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 312963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 313095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3132a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3134674ae819SStefano Zampini } 313563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 313663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 313763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3138674ae819SStefano Zampini } 313963602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3140674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 314163602bcaSStefano Zampini } else { 314263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 314363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3144854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 314563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 314663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314763602bcaSStefano Zampini } 314863602bcaSStefano Zampini } 3149674ae819SStefano Zampini } 3150674ae819SStefano Zampini 3151674ae819SStefano Zampini /* Setup of Graph */ 3152785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3153a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3154785d1243SStefano Zampini } 3155785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3156a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3157785d1243SStefano Zampini } 3158302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3159674ae819SStefano Zampini 3160674ae819SStefano Zampini /* Graph's connected components analysis */ 3161674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3162674ae819SStefano Zampini 3163674ae819SStefano Zampini /* print some info to stdout */ 3164674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3165302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3166674ae819SStefano Zampini } 3167fb180af4SStefano Zampini 3168fb180af4SStefano Zampini /* mark topography has done */ 3169fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3170674ae819SStefano Zampini PetscFunctionReturn(0); 3171674ae819SStefano Zampini } 3172674ae819SStefano Zampini 3173dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3174674ae819SStefano Zampini #undef __FUNCT__ 3175674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3176dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3177674ae819SStefano Zampini { 3178dc456d91SStefano Zampini PetscSF sf; 3179dc456d91SStefano Zampini PetscLayout map; 3180dc456d91SStefano Zampini const PetscInt *idxs; 3181dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3182dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3183dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3184dc456d91SStefano Zampini PetscMPIInt commsize; 3185674ae819SStefano Zampini PetscBool first_found; 3186674ae819SStefano Zampini PetscErrorCode ierr; 3187674ae819SStefano Zampini 3188674ae819SStefano Zampini PetscFunctionBegin; 3189dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3190dc456d91SStefano Zampini if (subset_mult) { 3191dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3192dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3193dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3194674ae819SStefano Zampini } 3195dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3196dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3197dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3198dc456d91SStefano Zampini for (i=0;i<n;i++) { 3199dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3200dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3201674ae819SStefano Zampini } 3202dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3203dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3204dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3205dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3206dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3207dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3208dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3209dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3210dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3211dc456d91SStefano Zampini 3212dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3213dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3214dc456d91SStefano Zampini if (subset_mult) { 3215dc456d91SStefano Zampini const PetscInt* idxs_mult; 3216dc456d91SStefano Zampini 3217dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3218dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3219dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3220674ae819SStefano Zampini } else { 3221dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3222674ae819SStefano Zampini } 3223dc456d91SStefano Zampini /* local size of new subset */ 3224dc456d91SStefano Zampini n_n = 0; 3225dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3226dc456d91SStefano Zampini 3227dc456d91SStefano Zampini /* global indexes in layout */ 3228dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3229dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3230dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3231dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3232dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3233dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3234dc456d91SStefano Zampini 3235dc456d91SStefano Zampini /* reduce from leaves to roots */ 3236dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 3237dc456d91SStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPIU_MAX);CHKERRQ(ierr); 3238dc456d91SStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPIU_MAX);CHKERRQ(ierr); 3239dc456d91SStefano Zampini 3240dc456d91SStefano Zampini /* count indexes in local part of layout */ 3241674ae819SStefano Zampini nlocals = 0; 3242674ae819SStefano Zampini first_index = -1; 3243674ae819SStefano Zampini first_found = PETSC_FALSE; 3244dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3245dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3246674ae819SStefano Zampini first_found = PETSC_TRUE; 3247674ae819SStefano Zampini first_index = i; 3248674ae819SStefano Zampini } 3249dc456d91SStefano Zampini nlocals += root_data[i]; 3250674ae819SStefano Zampini } 3251dc456d91SStefano Zampini 3252dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 3253dc456d91SStefano Zampini start = 0; 3254dc456d91SStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPIU_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3255dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3256dc456d91SStefano Zampini *N_n = start + nlocals; 3257dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3258dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3259674ae819SStefano Zampini } 3260dc456d91SStefano Zampini /* adapt roof data with cumulative */ 3261674ae819SStefano Zampini if (first_found) { 3262dc456d91SStefano Zampini PetscInt old_index; 3263dc456d91SStefano Zampini 3264dc456d91SStefano Zampini root_data[first_index] += start; 3265674ae819SStefano Zampini old_index = first_index; 3266dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3267dc456d91SStefano Zampini if (root_data[i]) { 3268dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3269674ae819SStefano Zampini old_index = i; 3270674ae819SStefano Zampini } 3271674ae819SStefano Zampini } 3272674ae819SStefano Zampini } 3273dc456d91SStefano Zampini 3274dc456d91SStefano Zampini /* from roots to leaves */ 3275dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3276dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3277dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3278dc456d91SStefano Zampini 3279dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3280dc456d91SStefano Zampini if (subset_mult) { 3281dc456d91SStefano Zampini const PetscInt* idxs_mult; 3282dc456d91SStefano Zampini PetscInt cum; 3283dc456d91SStefano Zampini 3284dc456d91SStefano Zampini cum = 0; 3285dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3286dc456d91SStefano Zampini for (i=0;i<n;i++) { 3287dc456d91SStefano Zampini PetscInt j; 3288dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3289674ae819SStefano Zampini } 3290dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3291674ae819SStefano Zampini } else { 3292dc456d91SStefano Zampini for (i=0;i<n;i++) { 3293dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3294674ae819SStefano Zampini } 3295674ae819SStefano Zampini } 3296dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3297dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3298674ae819SStefano Zampini PetscFunctionReturn(0); 3299674ae819SStefano Zampini } 33009a7d3425SStefano Zampini 33019a7d3425SStefano Zampini #undef __FUNCT__ 33029a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33039a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33049a7d3425SStefano Zampini { 33059a7d3425SStefano Zampini PetscInt i,j; 33069a7d3425SStefano Zampini PetscScalar *alphas; 33079a7d3425SStefano Zampini PetscErrorCode ierr; 33089a7d3425SStefano Zampini 33099a7d3425SStefano Zampini PetscFunctionBegin; 33109a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3311785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33129a7d3425SStefano Zampini for (i=0;i<n;i++) { 33139a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33149a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33159a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33169a7d3425SStefano Zampini } 33179a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33189a7d3425SStefano Zampini PetscFunctionReturn(0); 33199a7d3425SStefano Zampini } 33209a7d3425SStefano Zampini 3321e7931f94SStefano Zampini #undef __FUNCT__ 332270cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 332328143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3324e7931f94SStefano Zampini { 3325e7931f94SStefano Zampini Mat subdomain_adj; 3326e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3327e7931f94SStefano Zampini MatPartitioning partitioner; 3328e7931f94SStefano Zampini Mat_IS *matis; 3329e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3330e7931f94SStefano Zampini PetscInt prank; 3331e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3332e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3333e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33343837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3335e7931f94SStefano Zampini PetscErrorCode ierr; 33362b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3337e7931f94SStefano Zampini PetscSubcomm subcomm; 3338a57a6d2fSStefano Zampini 3339e7931f94SStefano Zampini PetscFunctionBegin; 33402b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33412b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33422b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3343e7931f94SStefano Zampini 3344e7931f94SStefano Zampini /* Get info on mapping */ 3345e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3346e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3347e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3348e7931f94SStefano Zampini 3349e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3350785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3351e7931f94SStefano Zampini xadj[0] = 0; 3352e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3353785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3354785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3355e7931f94SStefano Zampini 33562b510759SStefano Zampini if (threshold) { 3357d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33582b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3359d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3360d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3361d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3362d023bfaeSStefano Zampini xadj_count++; 3363e7931f94SStefano Zampini } 3364e7931f94SStefano Zampini } 3365d023bfaeSStefano Zampini xadj[1] = xadj_count; 3366c8587f34SStefano Zampini } else { 3367e7931f94SStefano Zampini if (xadj[1]) { 3368e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3369e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3370c8587f34SStefano Zampini } 3371e7931f94SStefano Zampini } 3372e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3373e7931f94SStefano Zampini if (use_square) { 3374e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3375e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3376e7931f94SStefano Zampini } 3377e7931f94SStefano Zampini } 3378e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3379e7931f94SStefano Zampini 33803837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3381e7931f94SStefano Zampini 3382e7931f94SStefano Zampini /* 3383e7931f94SStefano Zampini Restrict work on active processes only. 3384e7931f94SStefano Zampini */ 3385e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3386e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3387e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33882b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3389d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3390e7931f94SStefano Zampini if (color) { 3391e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3392e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3393e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3394c8587f34SStefano Zampini } else { 339528143c3dSStefano Zampini PetscInt coarsening_ratio; 3396306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3397785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3398e7931f94SStefano Zampini prank = rank; 3399306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34008002ef2cSStefano Zampini /* 3401e7931f94SStefano Zampini for (i=0;i<size;i++) { 3402e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3403c8587f34SStefano Zampini } 34048002ef2cSStefano Zampini */ 3405e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3406e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3407c8587f34SStefano Zampini } 3408e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3409306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 341022b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3411e7931f94SStefano Zampini 3412e7931f94SStefano Zampini /* Partition */ 3413306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3414e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3415e7931f94SStefano Zampini if (use_vwgt) { 34163837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3417e7931f94SStefano Zampini v_wgt[0] = local_size; 3418e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3419c8587f34SStefano Zampini } 342028143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 342128143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 342228143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3423e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3424e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 342522b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3426e7931f94SStefano Zampini 3427e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 342828143c3dSStefano Zampini if (contiguous) { 342928143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 343028143c3dSStefano Zampini } else { 343184ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 343228143c3dSStefano Zampini } 3433e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3434e7931f94SStefano Zampini /* clean up */ 3435e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3436e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3437e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3438e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3439e7931f94SStefano Zampini } 3440e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3441e7931f94SStefano Zampini 3442e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3443e7931f94SStefano Zampini i = 1; 3444e7931f94SStefano Zampini if (color) i=0; 3445e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3446e7931f94SStefano Zampini 3447e7931f94SStefano Zampini /* get back IS */ 3448e7931f94SStefano Zampini *is_sends = ranks_send_to; 3449e7931f94SStefano Zampini PetscFunctionReturn(0); 3450e7931f94SStefano Zampini } 3451e7931f94SStefano Zampini 3452e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3453e7931f94SStefano Zampini 3454e7931f94SStefano Zampini #undef __FUNCT__ 3455e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 345653a05cb3SStefano 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[]) 3457e7931f94SStefano Zampini { 345870cf5478SStefano Zampini Mat local_mat; 3459e7931f94SStefano Zampini Mat_IS *matis; 3460e7931f94SStefano Zampini IS is_sends_internal; 34619d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 346228143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34639d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3464e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3465e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3466e7931f94SStefano Zampini const PetscInt* is_indices; 3467e7931f94SStefano Zampini MatType new_local_type; 3468e7931f94SStefano Zampini /* buffers */ 3469e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 347028143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34719d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3472e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3473e7931f94SStefano Zampini /* MPI */ 347428143c3dSStefano Zampini MPI_Comm comm,comm_n; 347528143c3dSStefano Zampini PetscSubcomm subcomm; 3476e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 347728143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 347828143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 347928143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 348028143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 348128143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3482e7931f94SStefano Zampini PetscErrorCode ierr; 3483e7931f94SStefano Zampini 3484e7931f94SStefano Zampini PetscFunctionBegin; 348528143c3dSStefano Zampini /* TODO: add missing checks */ 348628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 348728143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 348828143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 348928143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3490e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 349128143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3492e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3493e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3494e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3495e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3496e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 349728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 349870cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 349970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 350028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 350170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 350270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 350370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 350470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 350570cf5478SStefano Zampini } 3506e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3507e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3508e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3509e7931f94SStefano Zampini if (!is_sends) { 351028143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 351128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 351228143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3513c8587f34SStefano Zampini } else { 3514e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3515e7931f94SStefano Zampini is_sends_internal = is_sends; 3516c8587f34SStefano Zampini } 3517e7931f94SStefano Zampini 3518e7931f94SStefano Zampini /* get pointer of MATIS data */ 3519e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3520e7931f94SStefano Zampini 3521e7931f94SStefano Zampini /* get comm */ 3522a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3523e7931f94SStefano Zampini 3524e7931f94SStefano Zampini /* compute number of sends */ 3525e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3526e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3527e7931f94SStefano Zampini 3528e7931f94SStefano Zampini /* compute number of receives */ 3529e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3530785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3531e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3532e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3533e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3534e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3535e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3536e7931f94SStefano Zampini 353728143c3dSStefano Zampini /* restrict comm if requested */ 353828143c3dSStefano Zampini subcomm = 0; 353928143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 354028143c3dSStefano Zampini if (restrict_comm) { 3541779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3542779c1cceSStefano Zampini 354328143c3dSStefano Zampini color = 0; 354453a05cb3SStefano Zampini if (restrict_full) { 354553a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 354653a05cb3SStefano Zampini } else { 354753a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 354853a05cb3SStefano Zampini } 354928143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 355028143c3dSStefano Zampini subcommsize = commsize - subcommsize; 355128143c3dSStefano Zampini /* check if reuse has been requested */ 355228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 355328143c3dSStefano Zampini if (*mat_n) { 355428143c3dSStefano Zampini PetscMPIInt subcommsize2; 355528143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 355628143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 355728143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 355828143c3dSStefano Zampini } else { 355928143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 356028143c3dSStefano Zampini } 356128143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3562779c1cceSStefano Zampini PetscMPIInt rank; 3563779c1cceSStefano Zampini 3564779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 356528143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 356628143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 356728143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3568306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 356928143c3dSStefano Zampini } 357028143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 357128143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 357228143c3dSStefano Zampini } else { 357328143c3dSStefano Zampini comm_n = comm; 357428143c3dSStefano Zampini } 357528143c3dSStefano Zampini 3576e7931f94SStefano Zampini /* prepare send/receive buffers */ 3577785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3578e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3579785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3580e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 358128143c3dSStefano Zampini if (nis) { 3582854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 358328143c3dSStefano Zampini } 3584e7931f94SStefano Zampini 358528143c3dSStefano Zampini /* Get data from local matrices */ 3586e7931f94SStefano Zampini if (!isdense) { 3587a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3588e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3589e7931f94SStefano Zampini /* 3590e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3591e7931f94SStefano Zampini send_buffer_idxs should contain: 3592e7931f94SStefano Zampini - MatType_PRIVATE type 3593e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3594e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3595e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3596e7931f94SStefano Zampini */ 3597e7931f94SStefano Zampini } else { 3598e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3599e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3600854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3601e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3602e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3603e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3604e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3605e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3606e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3607e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3608e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3609e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3610c8587f34SStefano Zampini } 3611c8587f34SStefano Zampini } 3612e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 361328143c3dSStefano Zampini /* additional is (if any) */ 361428143c3dSStefano Zampini if (nis) { 361528143c3dSStefano Zampini PetscMPIInt psum; 361628143c3dSStefano Zampini PetscInt j; 361728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 361828143c3dSStefano Zampini PetscInt plen; 361928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 362028143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 362128143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 362228143c3dSStefano Zampini } 3623854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 362428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 362528143c3dSStefano Zampini PetscInt plen; 362628143c3dSStefano Zampini const PetscInt *is_array_idxs; 362728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 362828143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 362928143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 363028143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 363128143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 363228143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 363328143c3dSStefano Zampini } 363428143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 363528143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 363628143c3dSStefano Zampini } 363728143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 363828143c3dSStefano Zampini } 363928143c3dSStefano Zampini 3640e7931f94SStefano Zampini buf_size_idxs = 0; 3641e7931f94SStefano Zampini buf_size_vals = 0; 364228143c3dSStefano Zampini buf_size_idxs_is = 0; 3643e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3644e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3645e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 364628143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3647e7931f94SStefano Zampini } 3648785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3649785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 365095ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3651e7931f94SStefano Zampini 3652e7931f94SStefano Zampini /* get new tags for clean communications */ 3653e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3654e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 365528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3656e7931f94SStefano Zampini 3657e7931f94SStefano Zampini /* allocate for requests */ 3658785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3659785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 366095ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3661785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3662785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 366395ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3664e7931f94SStefano Zampini 3665e7931f94SStefano Zampini /* communications */ 3666e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3667e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 366828143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3669e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3670e7931f94SStefano Zampini source_dest = onodes[i]; 3671e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3672e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3673e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3674e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 367528143c3dSStefano Zampini if (nis) { 367628143c3dSStefano 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); 367728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 367828143c3dSStefano Zampini } 3679e7931f94SStefano Zampini } 3680e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3681e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3682e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3683e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 368428143c3dSStefano Zampini if (nis) { 368528143c3dSStefano 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); 368628143c3dSStefano Zampini } 3687e7931f94SStefano Zampini } 3688e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3689e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3690e7931f94SStefano Zampini 3691e7931f94SStefano Zampini /* assemble new l2g map */ 3692e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3693e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36949d30be91SStefano Zampini new_local_rows = 0; 3695e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36969d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3697e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3698e7931f94SStefano Zampini } 36999d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3700e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37019d30be91SStefano Zampini new_local_rows = 0; 3702e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37039d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37049d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3705e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3706e7931f94SStefano Zampini } 37079d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37089d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3709e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3710e7931f94SStefano Zampini 3711e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3712e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3713e7931f94SStefano 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) */ 3714e7931f94SStefano Zampini if (n_recvs) { 371528143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3716e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3717e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3718e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3719e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3720e7931f94SStefano Zampini break; 3721e7931f94SStefano Zampini } 3722e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3723e7931f94SStefano Zampini } 3724e7931f94SStefano Zampini switch (new_local_type_private) { 372528143c3dSStefano Zampini case MATDENSE_PRIVATE: 372628143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3727e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3728e7931f94SStefano Zampini bs = 1; 372928143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 373028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 373128143c3dSStefano Zampini bs = 1; 373228143c3dSStefano Zampini } 3733e7931f94SStefano Zampini break; 3734e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3735e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3736e7931f94SStefano Zampini bs = 1; 3737e7931f94SStefano Zampini break; 3738e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3739e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3740e7931f94SStefano Zampini break; 3741e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3742e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3743e7931f94SStefano Zampini break; 3744e7931f94SStefano Zampini default: 37459d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3746e7931f94SStefano Zampini break; 3747e7931f94SStefano Zampini } 374828143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 374928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 375028143c3dSStefano Zampini bs = 1; 3751e7931f94SStefano Zampini } 3752e7931f94SStefano Zampini 375370cf5478SStefano Zampini /* create MATIS object if needed */ 375470cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3755e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 375628143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 375770cf5478SStefano Zampini } else { 375870cf5478SStefano Zampini /* it also destroys the local matrices */ 375970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 376070cf5478SStefano Zampini } 376170cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3762e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37639d30be91SStefano Zampini 37649d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37659d30be91SStefano Zampini 37669d30be91SStefano Zampini /* Global to local map of received indices */ 37679d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37689d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37699d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37709d30be91SStefano Zampini 37719d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37729d30be91SStefano Zampini buf_size_idxs = 0; 37739d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37749d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37759d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37769d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37779d30be91SStefano Zampini } 37789d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37799d30be91SStefano Zampini 37809d30be91SStefano Zampini /* set preallocation */ 37819d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37829d30be91SStefano Zampini if (!newisdense) { 37839d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37849d30be91SStefano Zampini 37859d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37869d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37879d30be91SStefano Zampini if (n_recvs) { 37889d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37899d30be91SStefano Zampini } 37909d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37919d30be91SStefano Zampini PetscInt j; 37929d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37939d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37949d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37959d30be91SStefano Zampini } 37969d30be91SStefano Zampini } else { 37979d30be91SStefano Zampini /* TODO */ 37989d30be91SStefano Zampini } 37999d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38009d30be91SStefano Zampini } 38019d30be91SStefano Zampini if (new_local_nnz) { 38029d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38039d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38049d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38059d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38069d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38079d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38089d30be91SStefano Zampini } else { 38099d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38109d30be91SStefano Zampini } 38119d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38129d30be91SStefano Zampini } else { 38139d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38149d30be91SStefano Zampini } 3815e7931f94SStefano Zampini 3816e7931f94SStefano Zampini /* set values */ 3817e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38189d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3819e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3820e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3821e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38229d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3823e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3824e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3825e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 382628143c3dSStefano Zampini } else { 382728143c3dSStefano Zampini /* TODO */ 3828e7931f94SStefano Zampini } 3829e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3830e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3831e7931f94SStefano Zampini } 3832e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3833e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383470cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 383570cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38369d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38379d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3838e7931f94SStefano Zampini 3839dfd14d43SStefano Zampini #if 0 384028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3841e7931f94SStefano Zampini Vec lvec,rvec; 3842e7931f94SStefano Zampini PetscReal infty_error; 3843e7931f94SStefano Zampini 38442a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3845e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3846e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3847e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 384870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3849e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3850e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3851e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3852e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3853e7931f94SStefano Zampini } 385428143c3dSStefano Zampini #endif 3855e7931f94SStefano Zampini 385628143c3dSStefano Zampini /* assemble new additional is (if any) */ 385728143c3dSStefano Zampini if (nis) { 385828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 385928143c3dSStefano Zampini 386028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3861854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 386228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 386328143c3dSStefano Zampini psum = 0; 386428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 386528143c3dSStefano Zampini for (j=0;j<nis;j++) { 386628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 386728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 386828143c3dSStefano Zampini psum += plen; 386928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 387028143c3dSStefano Zampini } 387128143c3dSStefano Zampini } 3872854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3873854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 387428143c3dSStefano Zampini for (i=1;i<nis;i++) { 387528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 387628143c3dSStefano Zampini } 387728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 387828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 387928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 388028143c3dSStefano Zampini for (j=0;j<nis;j++) { 388128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 388228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 388328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 388428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 388528143c3dSStefano Zampini } 388628143c3dSStefano Zampini } 388728143c3dSStefano Zampini for (i=0;i<nis;i++) { 388828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 388928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 389028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 389128143c3dSStefano Zampini } 389228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 389328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 389428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 389528143c3dSStefano Zampini } 3896e7931f94SStefano Zampini /* free workspace */ 389728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3898e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3899e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3900e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3901e7931f94SStefano Zampini if (isdense) { 3902e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3903e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3904e7931f94SStefano Zampini } else { 3905e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3906e7931f94SStefano Zampini } 390728143c3dSStefano Zampini if (nis) { 390828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 390928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 391028143c3dSStefano Zampini } 3911e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3912e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 391328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3914e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3915e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 391628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3917e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3918e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3919e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3920e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3921e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 392228143c3dSStefano Zampini if (nis) { 392328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 392428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 392528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 392628143c3dSStefano Zampini } 392728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 392828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 392928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 393028143c3dSStefano Zampini for (i=0;i<nis;i++) { 393128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 393228143c3dSStefano Zampini } 393353a05cb3SStefano Zampini *mat_n = NULL; 393428143c3dSStefano Zampini } 3935e7931f94SStefano Zampini PetscFunctionReturn(0); 3936e7931f94SStefano Zampini } 3937a57a6d2fSStefano Zampini 393812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3939af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 394012edc857SStefano Zampini 3941c8587f34SStefano Zampini #undef __FUNCT__ 3942c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3943c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3944c8587f34SStefano Zampini { 3945c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3946c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 394720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39489881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 394920a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39506e683305SStefano Zampini IS coarse_is,*isarray; 39516e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39526e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3953f9eb5b7dSStefano Zampini PC pc_temp; 3954c8587f34SStefano Zampini PCType coarse_pc_type; 3955c8587f34SStefano Zampini KSPType coarse_ksp_type; 3956f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39574f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39586e683305SStefano Zampini Mat t_coarse_mat_is; 39596e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39606e683305SStefano Zampini PetscMPIInt all_procs; 396174e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 396268457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 396322bc73bbSStefano Zampini PetscScalar *array; 39649881197aSStefano Zampini PetscErrorCode ierr; 3965fdc09c96SStefano Zampini 3966c8587f34SStefano Zampini PetscFunctionBegin; 3967c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 396868457ee5SStefano 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 */ 3969fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39705a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3971fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3972f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3973f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3974f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3975fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3976fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3977727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3978fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3979fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3980fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3981f4ddd8eeSStefano Zampini } 3982fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3983fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3984f4ddd8eeSStefano Zampini } 398570cf5478SStefano Zampini /* reset any subassembling information */ 398670cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39876e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39886e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3989fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3990f4ddd8eeSStefano Zampini } 3991c8587f34SStefano Zampini 39926e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39932b510759SStefano Zampini im_active = !!(pcis->n); 39942b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39956e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39966e683305SStefano Zampini void_procs = all_procs-active_procs; 39976e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 399874e2c79eSStefano Zampini redist = PETSC_FALSE; 399922bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40006e683305SStefano Zampini csin_ml = PETSC_TRUE; 40016e683305SStefano Zampini ncoarse_ml = void_procs; 4002779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4003779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 400418a45a71SStefano Zampini csin_ds = PETSC_TRUE; 400518a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 400618a45a71SStefano Zampini redist = PETSC_TRUE; 400718a45a71SStefano Zampini } else { 40086e683305SStefano Zampini csin_ds = PETSC_TRUE; 4009779c1cceSStefano Zampini ncoarse_ds = active_procs; 4010779c1cceSStefano Zampini redist = PETSC_TRUE; 401118a45a71SStefano Zampini } 40126e683305SStefano Zampini } else { 40136e683305SStefano Zampini csin_ml = PETSC_FALSE; 40146e683305SStefano Zampini ncoarse_ml = all_procs; 40156e683305SStefano Zampini if (void_procs) { 40166e683305SStefano Zampini csin_ds = PETSC_TRUE; 40176e683305SStefano Zampini ncoarse_ds = void_procs; 40186e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40196e683305SStefano Zampini } else { 4020779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 402174e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 402274e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 402374e2c79eSStefano Zampini redist = PETSC_TRUE; 402474e2c79eSStefano Zampini } else { 40256e683305SStefano Zampini csin_ds = PETSC_FALSE; 40266e683305SStefano Zampini ncoarse_ds = all_procs; 40276e683305SStefano Zampini } 40286e683305SStefano Zampini } 402974e2c79eSStefano Zampini } 40306e683305SStefano Zampini 40316e683305SStefano Zampini /* 40326e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40336e683305SStefano Zampini - we have not exceeded the number of levels requested 40346e683305SStefano Zampini - we can actually subassemble the active processes 40356e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40366e683305SStefano Zampini */ 40376e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40386e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40396e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40406e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40416e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4042f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40432b510759SStefano Zampini } else { 4044f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4045c8587f34SStefano Zampini } 4046c8587f34SStefano Zampini } 40476e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40486e683305SStefano Zampini if (multilevel_allowed) { 40496e683305SStefano Zampini ncoarse = ncoarse_ml; 40506e683305SStefano Zampini csin = csin_ml; 405158da7f69SStefano Zampini redist = PETSC_FALSE; 40526e683305SStefano Zampini } else { 40536e683305SStefano Zampini ncoarse = ncoarse_ds; 40546e683305SStefano Zampini csin = csin_ds; 40556e683305SStefano Zampini } 4056e7931f94SStefano Zampini 4057abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4058abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4059abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4060abbbba34SStefano Zampini 4061abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 406222bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 406322bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 406422bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 406522bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4066b9b85e73SStefano Zampini #if 0 4067b9b85e73SStefano Zampini { 4068b9b85e73SStefano Zampini PetscViewer viewer; 4069b9b85e73SStefano Zampini char filename[256]; 4070b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4071b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4072b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4073b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4074b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4075b9b85e73SStefano Zampini } 4076b9b85e73SStefano Zampini #endif 40776e683305SStefano Zampini ierr = MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,&t_coarse_mat_is);CHKERRQ(ierr); 40786e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40796e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40806e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4081abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4082abbbba34SStefano Zampini 40836e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40846e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40856e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40866e683305SStefano Zampini const PetscInt *idxs; 40876e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40886e683305SStefano Zampini 40896e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40900be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40916e683305SStefano Zampini /* allocate space for temporary storage */ 4092854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4093854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40946e683305SStefano Zampini /* allocate for IS array */ 40956e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40966e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40976e683305SStefano Zampini nis = nisdofs + nisneu; 4098854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40996e683305SStefano Zampini /* dofs splitting */ 41006e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41016e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41026e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41036e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41046e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41056e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41066e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41076e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41086e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41096e683305SStefano Zampini } 41106e683305SStefano Zampini /* neumann boundaries */ 41116e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41126e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41136e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41146e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41156e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41166e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41176e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41186e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41196e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41206e683305SStefano Zampini } 41216e683305SStefano Zampini /* free memory */ 41226e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41236e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41246e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41256e683305SStefano Zampini } else { 41266e683305SStefano Zampini nis = 0; 41276e683305SStefano Zampini nisdofs = 0; 41286e683305SStefano Zampini nisneu = 0; 41296e683305SStefano Zampini isarray = NULL; 41306e683305SStefano Zampini } 41316e683305SStefano Zampini /* destroy no longer needed map */ 41326e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41336e683305SStefano Zampini 41346e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41356e683305SStefano Zampini coarse_mat_is = NULL; 41366e683305SStefano Zampini if (csin) { 41376e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 413874e2c79eSStefano Zampini if (redist) { 413974e2c79eSStefano Zampini PetscMPIInt rank; 4140779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 414174e2c79eSStefano Zampini 414274e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 414358da7f69SStefano Zampini spc = active_procs/ncoarse; 414458da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4145779c1cceSStefano Zampini if (im_active) { 4146779c1cceSStefano Zampini destsize = 1; 414774e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 414874e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 414974e2c79eSStefano Zampini } else { 415074e2c79eSStefano Zampini dest[0] = rank/(spc+1); 415174e2c79eSStefano Zampini } 415274e2c79eSStefano Zampini } else { 4153779c1cceSStefano Zampini destsize = 0; 41546e683305SStefano Zampini } 4155779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4156779c1cceSStefano Zampini } else if (csin_type_simple) { 41576e683305SStefano Zampini PetscMPIInt rank; 41586e683305SStefano Zampini PetscInt issize,isidx; 4159779c1cceSStefano Zampini 41606e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41616e683305SStefano Zampini if (im_active) { 41626e683305SStefano Zampini issize = 1; 41636e683305SStefano Zampini isidx = (PetscInt)rank; 41646e683305SStefano Zampini } else { 41656e683305SStefano Zampini issize = 0; 41666e683305SStefano Zampini isidx = -1; 41676e683305SStefano Zampini } 41686e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4169779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41706e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41716e683305SStefano Zampini } 4172779c1cceSStefano Zampini 4173779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4174779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4175779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4176779c1cceSStefano Zampini PetscInt *coarse_candidates; 4177779c1cceSStefano Zampini const PetscInt* tisindices; 4178779c1cceSStefano Zampini 4179779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4180779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4181779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4182779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4183779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4184779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4185779c1cceSStefano Zampini } 4186779c1cceSStefano Zampini } 4187779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4188779c1cceSStefano Zampini 4189779c1cceSStefano Zampini 41906e683305SStefano Zampini if (pcbddc->dbg_flag) { 41916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41936e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4195779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41976e683305SStefano Zampini } 41986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41996e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42006e683305SStefano Zampini } 42016e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42026e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42036e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4204854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42056e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42066e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42076e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42086e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42096e683305SStefano Zampini } 42106e683305SStefano Zampini if (pcbddc->dbg_flag) { 42116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42126e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42136e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42146e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42156e683305SStefano Zampini } 4216779c1cceSStefano Zampini } 42176e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 421853a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 421953a05cb3SStefano 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); 422053a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 422153a05cb3SStefano 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); 422253a05cb3SStefano Zampini } 42236e683305SStefano Zampini } else { 42246e683305SStefano Zampini if (pcbddc->dbg_flag) { 42256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42266e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42276e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42286e683305SStefano Zampini } 42296e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42306e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42316e683305SStefano Zampini } 42326e683305SStefano Zampini 42336e683305SStefano Zampini /* create local to global scatters for coarse problem */ 423468457ee5SStefano Zampini if (compute_vecs) { 42356e683305SStefano Zampini PetscInt lrows; 42366e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42376e683305SStefano Zampini if (coarse_mat_is) { 42386e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42396e683305SStefano Zampini } else { 42406e683305SStefano Zampini lrows = 0; 42416e683305SStefano Zampini } 42426e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42436e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42446e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42456e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42466e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42476e683305SStefano Zampini } 42486e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42496e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4250c8587f34SStefano Zampini 4251f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4252f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4253f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4254f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4255f9eb5b7dSStefano Zampini } else { 4256f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4257f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4258c8587f34SStefano Zampini } 4259c8587f34SStefano Zampini 42606e683305SStefano Zampini /* print some info if requested */ 42616e683305SStefano Zampini if (pcbddc->dbg_flag) { 42626e683305SStefano Zampini if (!multilevel_allowed) { 42636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42646e683305SStefano Zampini if (multilevel_requested) { 42656e683305SStefano 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); 42666e683305SStefano Zampini } else if (pcbddc->max_levels) { 42676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42686e683305SStefano Zampini } 42696e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42706e683305SStefano Zampini } 42716e683305SStefano Zampini } 42726e683305SStefano Zampini 4273f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42746e683305SStefano Zampini if (coarse_mat_is) { 42756e683305SStefano Zampini MatReuse coarse_mat_reuse; 42766a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42776e683305SStefano Zampini if (pcbddc->dbg_flag) { 42786e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42796e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42806e683305SStefano Zampini } 4281f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4282312be037SStefano Zampini char prefix[256],str_level[16]; 4283e604994aSStefano Zampini size_t len; 42846e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4285c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4286f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42875f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4288c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42896e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4290c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4291c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42923e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4293e604994aSStefano Zampini /* prefix */ 4294e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4295e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4296e604994aSStefano Zampini if (!pcbddc->current_level) { 4297e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4298e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4299c8587f34SStefano Zampini } else { 4300e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4301312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4302312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 430334d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4304312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4305e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4306e604994aSStefano Zampini } 4307e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43083e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43093e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43103e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43113e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4312f9eb5b7dSStefano Zampini /* allow user customization */ 4313f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43143e3c6dadSStefano Zampini } 43153e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43163e3c6dadSStefano Zampini if (nisdofs) { 43173e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43183e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43193e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43203e3c6dadSStefano Zampini } 43213e3c6dadSStefano Zampini } 43223e3c6dadSStefano Zampini if (nisneu) { 43233e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43243e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4325312be037SStefano Zampini } 4326f9eb5b7dSStefano Zampini 4327f9eb5b7dSStefano Zampini /* get some info after set from options */ 4328f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4329f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4330f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43314f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43326e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4333f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4334f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4335f9eb5b7dSStefano Zampini } 43364f3a063dSStefano Zampini if (isredundant) { 43374f3a063dSStefano Zampini KSP inner_ksp; 43384f3a063dSStefano Zampini PC inner_pc; 43394f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43404f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43414f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43424f3a063dSStefano Zampini } 4343f9eb5b7dSStefano Zampini 4344f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4345fa7f1dd8SStefano Zampini if (coarse_reuse) { 434681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4347fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43486e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4349fa7f1dd8SStefano Zampini } else { 43506e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4351fa7f1dd8SStefano Zampini } 4352c8587f34SStefano Zampini if (isbddc || isnn) { 435322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 435470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 435528143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 435622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43586e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43596e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43606e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 436122b6e8a2SStefano Zampini } 436270cf5478SStefano Zampini } 436353a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 436470cf5478SStefano Zampini } else { 436522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 436622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 436722bc73bbSStefano Zampini } 436822bc73bbSStefano Zampini } else { 43692e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4370c8587f34SStefano Zampini } 4371c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4372c8587f34SStefano Zampini 43733301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 43745a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 43753301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 43763301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 43773301b35fSStefano Zampini } 43783301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 43793301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 43803301b35fSStefano Zampini } 43813301b35fSStefano Zampini if (pc->pmat->spd_set) { 43823301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 43833301b35fSStefano Zampini } 43846e683305SStefano Zampini /* set operators */ 43855f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43866e683305SStefano Zampini if (pcbddc->dbg_flag) { 43876e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43886e683305SStefano Zampini } 43896e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43906e683305SStefano Zampini coarse_mat = 0; 43916e683305SStefano Zampini } 43926e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4393b9b85e73SStefano Zampini #if 0 4394b9b85e73SStefano Zampini { 4395b9b85e73SStefano Zampini PetscViewer viewer; 4396b9b85e73SStefano Zampini char filename[256]; 4397b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4398b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4399b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4400b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4401b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4402b9b85e73SStefano Zampini } 4403b9b85e73SStefano Zampini #endif 4404c8587f34SStefano Zampini 4405c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4406c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4407c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 440898a51de6SStefano Zampini } 440998a51de6SStefano Zampini 441098a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 441198a51de6SStefano Zampini Vec crhs,csol; 441298a51de6SStefano Zampini PetscBool ispreonly; 4413*04708bb6SStefano Zampini 441498a51de6SStefano Zampini if (CoarseNullSpace) { 4415c8587f34SStefano Zampini if (isbddc) { 4416c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4417c8587f34SStefano Zampini } else { 4418c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4419c8587f34SStefano Zampini } 4420c8587f34SStefano Zampini } 4421f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4422f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4423f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4424f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44256e683305SStefano Zampini /* hack */ 4426f347579bSStefano Zampini if (!csol) { 44272a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4428f9eb5b7dSStefano Zampini } 4429f347579bSStefano Zampini if (!crhs) { 44302a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4431f347579bSStefano Zampini } 4432cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4433cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44346e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4435c8587f34SStefano Zampini KSP check_ksp; 44362b510759SStefano Zampini KSPType check_ksp_type; 4437c8587f34SStefano Zampini PC check_pc; 44386e683305SStefano Zampini Vec check_vec,coarse_vec; 44396a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44402b510759SStefano Zampini PetscInt its; 44416e683305SStefano Zampini PetscBool compute_eigs; 44426e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44436e683305SStefano Zampini PetscInt neigs; 44448e185a42SStefano Zampini const char *prefix; 4445c8587f34SStefano Zampini 44462b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44476e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 444823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4449f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44502b510759SStefano Zampini if (ispreonly) { 44512b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44526e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44532b510759SStefano Zampini } else { 4454cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44556e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4456c8587f34SStefano Zampini } 4457c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44586e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44596e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44606e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4461a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4462a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4463a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4464a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4465c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4466c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4467c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4468c8587f34SStefano Zampini /* create random vec */ 44696e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44706e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4471c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4472c8587f34SStefano Zampini if (CoarseNullSpace) { 4473c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4474c8587f34SStefano Zampini } 44756e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4476c8587f34SStefano Zampini /* solve coarse problem */ 44776e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4478c8587f34SStefano Zampini if (CoarseNullSpace) { 44796e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4480c8587f34SStefano Zampini } 4481cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44826e683305SStefano Zampini if (compute_eigs) { 4483854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4484854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44856e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44866e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44876e683305SStefano Zampini lambda_min = eigs_r[0]; 44886e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44896e683305SStefano Zampini if (lambda_max>lambda_min) { 4490cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4491cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4492cbcc2c2aSStefano Zampini } 4493c8587f34SStefano Zampini } 4494c8587f34SStefano Zampini } 4495cbcc2c2aSStefano Zampini 4496c8587f34SStefano Zampini /* check coarse problem residual error */ 44976e683305SStefano Zampini if (pcbddc->dbg_flag) { 44986e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44996e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45006e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4501c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45026e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45036e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4504c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4505779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45066e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45076e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45106e683305SStefano Zampini if (compute_eigs) { 45116e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4512deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4513c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45146e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45156e683305SStefano 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); 45166e683305SStefano Zampini for (i=0;i<neigs;i++) { 45176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4518c8587f34SStefano Zampini } 45196e683305SStefano Zampini } 45206e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45216e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45226e683305SStefano Zampini } 4523c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45246e683305SStefano Zampini if (compute_eigs) { 45256e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45266e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4527c8587f34SStefano Zampini } 45286e683305SStefano Zampini } 45296e683305SStefano Zampini } 4530cbcc2c2aSStefano Zampini /* print additional info */ 4531cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45326e683305SStefano Zampini /* waits until all processes reaches this point */ 45336e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4534cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4535cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4536cbcc2c2aSStefano Zampini } 4537cbcc2c2aSStefano Zampini 45382b510759SStefano Zampini /* free memory */ 4539c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4540fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4541c8587f34SStefano Zampini PetscFunctionReturn(0); 4542c8587f34SStefano Zampini } 4543674ae819SStefano Zampini 4544f34684f1SStefano Zampini #undef __FUNCT__ 4545f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4546f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4547f34684f1SStefano Zampini { 4548f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4549f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4550f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4551dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4552dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 455373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4554dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4555f34684f1SStefano Zampini PetscErrorCode ierr; 4556f34684f1SStefano Zampini 4557f34684f1SStefano Zampini PetscFunctionBegin; 4558f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45590e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45600e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4561727cdba6SStefano Zampini } 4562dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 4563dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,subset_n,&subset);CHKERRQ(ierr); 4564dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4565dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4566dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4567dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4568dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4569dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 45700e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45710e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45720e6343abSStefano Zampini } 4573dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4574dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4575dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4576dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4577dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4578f34684f1SStefano Zampini 4579f34684f1SStefano Zampini /* check numbering */ 4580f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4581f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4582dc456d91SStefano Zampini PetscInt i; 4583b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4584f34684f1SStefano Zampini 4585f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4586f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4587f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45880fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4589f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4590f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4591727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4592f34684f1SStefano Zampini } 4593f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4594f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4595f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4596f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4597f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4598f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4599f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4600f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4601f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4602f34684f1SStefano Zampini if (array[i] == 1.0) { 4603b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4604f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4605f34684f1SStefano Zampini } 4606f34684f1SStefano Zampini } 4607b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4608f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4609f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4610f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4611f34684f1SStefano Zampini } 4612f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4613f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4614f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4615f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4616f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4617f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4618b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4619ca8b9ea9SStefano Zampini PetscInt *gidxs; 4620ca8b9ea9SStefano Zampini 4621ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4622ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4623f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4624f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4625f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4626f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46274bc2dc4bSStefano 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); 4628f34684f1SStefano Zampini } 4629f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4630ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4631f34684f1SStefano Zampini } 4632f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4633302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4634f34684f1SStefano Zampini } 46358bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4636f34684f1SStefano Zampini /* get back data */ 4637f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4638f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4639674ae819SStefano Zampini PetscFunctionReturn(0); 4640674ae819SStefano Zampini } 4641674ae819SStefano Zampini 4642e456f2a8SStefano Zampini #undef __FUNCT__ 4643e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4644a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4645e456f2a8SStefano Zampini { 4646e456f2a8SStefano Zampini IS localis_t; 4647a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4648e456f2a8SStefano Zampini PetscScalar *vals; 4649e456f2a8SStefano Zampini PetscErrorCode ierr; 4650e456f2a8SStefano Zampini 4651e456f2a8SStefano Zampini PetscFunctionBegin; 4652a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4653e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4654854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4655e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4656e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4657a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4658a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46591035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4660a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46611035eff8SStefano Zampini } 4662a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4663e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4664e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4665a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4666a7dc3881SStefano Zampini /* now compute set in local ordering */ 4667a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4668a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4669a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4670a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4671a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4672ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4673e456f2a8SStefano Zampini lsize++; 4674e456f2a8SStefano Zampini } 4675e456f2a8SStefano Zampini } 4676854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4677a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4678ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4679e456f2a8SStefano Zampini idxs[lsize++] = i; 4680e456f2a8SStefano Zampini } 4681e456f2a8SStefano Zampini } 4682a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4683a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4684e456f2a8SStefano Zampini *localis = localis_t; 4685e456f2a8SStefano Zampini PetscFunctionReturn(0); 4686e456f2a8SStefano Zampini } 4687906d46d4SStefano Zampini 4688906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4689906d46d4SStefano Zampini #undef __FUNCT__ 4690906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4691906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4692906d46d4SStefano Zampini { 4693906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4694906d46d4SStefano Zampini PetscErrorCode ierr; 4695906d46d4SStefano Zampini 4696906d46d4SStefano Zampini PetscFunctionBegin; 4697906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4698906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4699906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4700906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4701906d46d4SStefano Zampini PetscFunctionReturn(0); 4702906d46d4SStefano Zampini } 4703906d46d4SStefano Zampini 4704906d46d4SStefano Zampini #undef __FUNCT__ 4705906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4706906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4707906d46d4SStefano Zampini { 4708906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4709906d46d4SStefano Zampini PetscErrorCode ierr; 4710906d46d4SStefano Zampini 4711906d46d4SStefano Zampini PetscFunctionBegin; 4712906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4713906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4714906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4715906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4716906d46d4SStefano Zampini PetscFunctionReturn(0); 4717906d46d4SStefano Zampini } 4718b96c3477SStefano Zampini 4719b96c3477SStefano Zampini #undef __FUNCT__ 4720b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 472108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4722b96c3477SStefano Zampini { 4723a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4724b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4725b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4726a64f4aa4SStefano Zampini Mat S_j; 4727b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4728b96c3477SStefano Zampini PetscBool free_used_adj; 4729b96c3477SStefano Zampini PetscErrorCode ierr; 4730b96c3477SStefano Zampini 4731b96c3477SStefano Zampini PetscFunctionBegin; 4732b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4733b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 473408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4735b96c3477SStefano Zampini used_xadj = NULL; 4736b96c3477SStefano Zampini used_adjncy = NULL; 4737b96c3477SStefano Zampini } else { 473808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 473908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 474008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 474108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4742b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4743b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4744b96c3477SStefano Zampini } else { 47452fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4746b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4747b96c3477SStefano Zampini PetscInt nvtxs; 4748b96c3477SStefano Zampini 47492fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47502fffb893SStefano Zampini if (flg_row) { 4751b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4752b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4753b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4754b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47552fffb893SStefano Zampini } else { 47562fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47572fffb893SStefano Zampini used_xadj = NULL; 47582fffb893SStefano Zampini used_adjncy = NULL; 47592fffb893SStefano Zampini } 47602fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4761b96c3477SStefano Zampini } 4762b96c3477SStefano Zampini } 4763d5574798SStefano Zampini 4764d5574798SStefano Zampini /* setup sub_schurs data */ 4765a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4766a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4767a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4768a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4769862806e4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE);CHKERRQ(ierr); 4770a64f4aa4SStefano Zampini } else { 47716816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 4772*04708bb6SStefano Zampini PetscBool isseqaij; 47735feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 47745feab87aSStefano Zampini PetscInt n_vertices; 47755feab87aSStefano Zampini 47765feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 47772034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 47785feab87aSStefano Zampini } 4779*04708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 4780*04708bb6SStefano Zampini if (!isseqaij) { 4781*04708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4782*04708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 4783*04708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 4784*04708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4785*04708bb6SStefano Zampini } else { 4786*04708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 4787*04708bb6SStefano Zampini } 4788*04708bb6SStefano Zampini } 4789862806e4SStefano 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);CHKERRQ(ierr); 4790a64f4aa4SStefano Zampini } 4791a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4792b96c3477SStefano Zampini 4793b96c3477SStefano Zampini /* free adjacency */ 4794b96c3477SStefano Zampini if (free_used_adj) { 4795b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4796b96c3477SStefano Zampini } 4797b96c3477SStefano Zampini PetscFunctionReturn(0); 4798b96c3477SStefano Zampini } 4799b96c3477SStefano Zampini 4800b96c3477SStefano Zampini #undef __FUNCT__ 4801b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 480208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4803b96c3477SStefano Zampini { 4804b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4805b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4806b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4807b96c3477SStefano Zampini PCBDDCGraph graph; 4808b96c3477SStefano Zampini PetscErrorCode ierr; 4809b96c3477SStefano Zampini 4810b96c3477SStefano Zampini PetscFunctionBegin; 4811b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 481208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48133301b35fSStefano Zampini IS verticesIS,verticescomm; 48143301b35fSStefano Zampini PetscInt vsize,*idxs; 4815b96c3477SStefano Zampini 4816b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48173301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48183301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48193301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48203301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48213301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4822b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48237fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48243301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48253301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4826b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4827b96c3477SStefano Zampini /* 4828b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4829b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4830b96c3477SStefano Zampini } 4831b96c3477SStefano Zampini */ 4832b96c3477SStefano Zampini } else { 4833b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4834b96c3477SStefano Zampini } 4835b96c3477SStefano Zampini 4836b96c3477SStefano Zampini /* sub_schurs init */ 4837a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4838a64f4aa4SStefano Zampini 4839b96c3477SStefano Zampini /* free graph struct */ 484008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4841b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4842b96c3477SStefano Zampini } 4843b96c3477SStefano Zampini PetscFunctionReturn(0); 4844b96c3477SStefano Zampini } 4845