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; 51*862806e4SStefano 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; 154*862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 15508122e43SStefano Zampini PetscBLASInt B_N; 156aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 15708122e43SStefano Zampini 158*862806e4SStefano 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 Mat newmat; 70406656605SStefano Zampini IS is_aux; 7053a50541eSStefano Zampini PetscInt ibs,mbs; 7063a50541eSStefano Zampini PetscBool issbaij; 7073a50541eSStefano Zampini 7086816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7096816873aSStefano Zampini IS tis; 7106816873aSStefano Zampini 7116816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7126816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7136816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7146816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7156816873aSStefano Zampini } else { 7163a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7176816873aSStefano Zampini } 7189577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 7199577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 7203a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 7219577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7229577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7239577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7249577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 72580677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7269577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7279577ea80SStefano Zampini } else { 7283a50541eSStefano Zampini /* this is safe */ 7299577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 7303a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7313a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 732b3d85658SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 733b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7343a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7353a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 73606656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7373a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7383a50541eSStefano Zampini } else { 739b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7409577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7419577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7429577ea80SStefano Zampini } 74380677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 74480677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7453a50541eSStefano Zampini } 74625084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 74788ebb749SStefano Zampini } 74888ebb749SStefano Zampini 74988ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 750f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 75106656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 75206656605SStefano Zampini if (pcbddc->coarse_phi_D) { 75306656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 75406656605SStefano Zampini } 755f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 75606656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 75706656605SStefano Zampini PetscScalar *marray; 75806656605SStefano Zampini 75906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 76006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 761f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 762f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 763f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 764f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 765f4ddd8eeSStefano Zampini } 766f4ddd8eeSStefano Zampini } 76706656605SStefano Zampini 768f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 76906656605SStefano Zampini PetscScalar *marray; 77088ebb749SStefano Zampini 77106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7728eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 77488ebb749SStefano Zampini } 7753301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 77606656605SStefano Zampini n *= 2; 777f4ddd8eeSStefano Zampini } 77806656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 77906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 78006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 78106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 78206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 78306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 78488ebb749SStefano Zampini } 7853301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 78606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7871b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 78806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 78906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 79088ebb749SStefano Zampini } 79188ebb749SStefano Zampini } else { 792c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 793c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7941b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 795c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 796c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 797c0553b1fSStefano Zampini } 79888ebb749SStefano Zampini } 79906656605SStefano Zampini } 80006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 80106656605SStefano Zampini /* vertices */ 80206656605SStefano Zampini if (n_vertices) { 80316f15bc4SStefano Zampini 80416f15bc4SStefano Zampini if (n_R) { 80506656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 80606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 80716f15bc4SStefano Zampini PetscScalar *x,*y; 80806656605SStefano Zampini 80906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 81021eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 81106656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 8126816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 81306656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 81406656605SStefano Zampini } else { 81506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 81606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 81706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 81806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 81906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 82006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 82106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 82206656605SStefano Zampini } 82306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 82406656605SStefano Zampini } 82580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 82606656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 82706656605SStefano Zampini if (n_constraints) { 82806656605SStefano Zampini Mat B; 82980677318SStefano Zampini 830b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 83180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 83280677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 83380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 83480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 83680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 83780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 83880677318SStefano Zampini } 83980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 84080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 84180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 84380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 84606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 84706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 84806656605SStefano Zampini } 84906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 85080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 85206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 85406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 85506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 85606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 85706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85806656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 85916f15bc4SStefano Zampini } else { 86016f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 86116f15bc4SStefano Zampini } 86221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 86306656605SStefano Zampini /* coarse basis functions */ 86406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 86516f15bc4SStefano Zampini PetscScalar *y; 86616f15bc4SStefano Zampini 86706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 86806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 87006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 87306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 87406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 87506656605SStefano Zampini 87606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 87706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 87906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 88106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 88206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 88306656605SStefano Zampini } 88406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 88506656605SStefano Zampini } 88606656605SStefano Zampini } 88706656605SStefano Zampini 88806656605SStefano Zampini if (n_constraints) { 88906656605SStefano Zampini Mat B; 89006656605SStefano Zampini 89106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 89206656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 89380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 89406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 89506656605SStefano Zampini if (n_vertices) { 89680677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 89780677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 89880677318SStefano Zampini } else { 89980677318SStefano Zampini Mat S_VCt; 90080677318SStefano Zampini 90180677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 90280677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 90380677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 90480677318SStefano Zampini } 90506656605SStefano Zampini } 90606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 90706656605SStefano Zampini /* coarse basis functions */ 90806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 90906656605SStefano Zampini PetscScalar *y; 91006656605SStefano Zampini 91106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 91206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 91406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 91906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 92006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 92106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 92206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 92306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 92406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 92506656605SStefano Zampini } 92606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92706656605SStefano Zampini } 92806656605SStefano Zampini } 92980677318SStefano Zampini if (n_constraints) { 93080677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 93180677318SStefano Zampini } 93206656605SStefano Zampini 93306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9343301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 93506656605SStefano Zampini 93606656605SStefano Zampini if (n_constraints) { 93716f15bc4SStefano Zampini Mat S_CCT,B_C; 93806656605SStefano Zampini 93980677318SStefano Zampini /* this is a lazy thing */ 94080677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 94106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 94416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 94506656605SStefano Zampini if (n_vertices) { 94616f15bc4SStefano Zampini Mat B_V,S_VCT; 94706656605SStefano Zampini 94806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 94906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 95006656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 95106656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 95216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 95306656605SStefano Zampini } 95406656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 95580677318SStefano Zampini } else { /* if there are no constraints, reset work */ 95680677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 95706656605SStefano Zampini } 95816f15bc4SStefano Zampini if (n_vertices && n_R) { 95906656605SStefano Zampini Mat A_VRT; 96080677318SStefano Zampini PetscScalar *marray; 96106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 96206656605SStefano Zampini 96380677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 96480677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 96580677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 96606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 96780677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 96880677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 96916f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 97006656605SStefano Zampini } 97106656605SStefano Zampini 97206656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 97306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97906656605SStefano Zampini } 98006656605SStefano Zampini } else { 98106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 98506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 98606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 98706656605SStefano Zampini } 98806656605SStefano Zampini } 98906656605SStefano Zampini /* coarse basis functions */ 99006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 99106656605SStefano Zampini PetscScalar *y; 99206656605SStefano Zampini 99306656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 99406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 99506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 99606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99806656605SStefano Zampini if (i<n_vertices) { 99906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 100006656605SStefano Zampini } 100106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 100306656605SStefano Zampini 100406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 100506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 100606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 100706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 101106656605SStefano Zampini } 101206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 101306656605SStefano Zampini } 101406656605SStefano Zampini } 1015d62866d3SStefano Zampini /* free memory */ 101688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 101706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 101806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 101906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 102006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1021d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1022d62866d3SStefano Zampini if (n_vertices) { 1023d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1024d62866d3SStefano Zampini } 1025d62866d3SStefano Zampini if (n_constraints) { 1026d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1027d62866d3SStefano Zampini } 102888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 102988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 103088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 103125084f0cSStefano Zampini if (pcbddc->dbg_flag) { 103288ebb749SStefano Zampini Mat coarse_sub_mat; 103325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 103488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 103588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 103688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10378bec7fa6SStefano Zampini Mat C_B,CPHI; 10388bec7fa6SStefano Zampini IS is_dummy; 10398bec7fa6SStefano Zampini Vec mones; 104088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 104188ebb749SStefano Zampini PetscReal real_value; 104288ebb749SStefano Zampini 104388ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 104588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 104688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1049c0553b1fSStefano Zampini if (unsymmetric_check) { 105088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 105288ebb749SStefano Zampini } 105388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 105488ebb749SStefano Zampini 105525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10563301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 105725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1058c0553b1fSStefano Zampini if (unsymmetric_check) { 105988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107188ebb749SStefano Zampini } else { 107288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 107388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 107488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 108088ebb749SStefano Zampini } 108188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 108288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 108388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 108488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 108581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10868bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10870fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 108806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10898bec7fa6SStefano Zampini 10908bec7fa6SStefano Zampini /* check constraints */ 10918bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10928bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10938bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10948bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10958bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10968bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10978bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1098bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1099bdae7319SStefano Zampini if (unsymmetric_check) { 1100bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1101bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1102bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1103bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1104bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1105bdae7319SStefano Zampini } 11068bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 11078bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 11088bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 11098bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 11108bec7fa6SStefano Zampini 111125084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 111288ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 111388ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 111488ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 111588ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 111688ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 111788ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 111888ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 111988ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 112088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 112188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1122c0553b1fSStefano Zampini if (unsymmetric_check) { 112388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 112488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 112588ebb749SStefano Zampini } 112688ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 112788ebb749SStefano Zampini } 11288629588bSStefano Zampini /* get back data */ 11298629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 113088ebb749SStefano Zampini PetscFunctionReturn(0); 113188ebb749SStefano Zampini } 113288ebb749SStefano Zampini 113388ebb749SStefano Zampini #undef __FUNCT__ 1134d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1135d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1136d65f70fdSStefano Zampini { 1137d65f70fdSStefano Zampini Mat *work_mat; 1138d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1139d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1140d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1141d65f70fdSStefano Zampini PetscErrorCode ierr; 1142d65f70fdSStefano Zampini 1143d65f70fdSStefano Zampini PetscFunctionBegin; 1144d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1148d65f70fdSStefano Zampini 1149d65f70fdSStefano Zampini if (!rsorted) { 1150d65f70fdSStefano Zampini const PetscInt *idxs; 1151d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1152d65f70fdSStefano Zampini 1153d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1155d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1156d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1157d65f70fdSStefano Zampini } 1158d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1160d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1161d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1162d65f70fdSStefano Zampini } 1163d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1164d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini } else { 1166d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1167d65f70fdSStefano Zampini isrow_s = isrow; 1168d65f70fdSStefano Zampini } 1169d65f70fdSStefano Zampini 1170d65f70fdSStefano Zampini if (!csorted) { 1171d65f70fdSStefano Zampini if (isrow == iscol) { 1172d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1173d65f70fdSStefano Zampini iscol_s = isrow_s; 1174d65f70fdSStefano Zampini } else { 1175d65f70fdSStefano Zampini const PetscInt *idxs; 1176d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1177d65f70fdSStefano Zampini 1178d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1180d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1181d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1182d65f70fdSStefano Zampini } 1183d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1184d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1185d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1186d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1187d65f70fdSStefano Zampini } 1188d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1189d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1190d65f70fdSStefano Zampini } 1191d65f70fdSStefano Zampini } else { 1192d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1193d65f70fdSStefano Zampini iscol_s = iscol; 1194d65f70fdSStefano Zampini } 1195d65f70fdSStefano Zampini 1196d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1197d65f70fdSStefano Zampini 1198d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1199d65f70fdSStefano Zampini Mat new_mat; 1200d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1201d65f70fdSStefano Zampini 1202d65f70fdSStefano Zampini if (!rsorted) { 1203d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1204d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1205d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1206d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1207d65f70fdSStefano Zampini } 1208d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1210d65f70fdSStefano Zampini } else { 1211d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini } 1213d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1214d65f70fdSStefano Zampini 1215d65f70fdSStefano Zampini if (!csorted) { 1216d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1217d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1219d65f70fdSStefano Zampini } else { 1220d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1221d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1222d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1223d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1224d65f70fdSStefano Zampini } 1225d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1226d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1227d65f70fdSStefano Zampini } 1228d65f70fdSStefano Zampini } else { 1229d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1230d65f70fdSStefano Zampini } 1231d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1232d65f70fdSStefano Zampini 1233d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1234d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1235d65f70fdSStefano Zampini work_mat[0] = new_mat; 1236d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1237d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1238d65f70fdSStefano Zampini } 1239d65f70fdSStefano Zampini 1240d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1241d65f70fdSStefano Zampini *B = work_mat[0]; 1242d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1243d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1244d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1245d65f70fdSStefano Zampini PetscFunctionReturn(0); 1246d65f70fdSStefano Zampini } 1247d65f70fdSStefano Zampini 1248d65f70fdSStefano Zampini #undef __FUNCT__ 12495e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12505e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1251aa0d41d4SStefano Zampini { 1252aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12535e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1254d65f70fdSStefano Zampini Mat new_mat; 12555e8657edSStefano Zampini IS is_local,is_global; 1256d65f70fdSStefano Zampini PetscInt local_size; 1257d65f70fdSStefano Zampini PetscBool isseqaij; 1258aa0d41d4SStefano Zampini PetscErrorCode ierr; 1259aa0d41d4SStefano Zampini 1260aa0d41d4SStefano Zampini PetscFunctionBegin; 1261aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12625e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12635e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1264906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1265906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1266d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1267906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1268906d46d4SStefano Zampini 1269906d46d4SStefano Zampini /* check */ 1270906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1271906d46d4SStefano Zampini Vec x,x_change; 1272906d46d4SStefano Zampini PetscReal error; 1273906d46d4SStefano Zampini 12745e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1275906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12765e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12775e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12785e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1279d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12805e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12815e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1282906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1283906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1284906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1285906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1286906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1287906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1288906d46d4SStefano Zampini } 1289906d46d4SStefano Zampini 129022d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12919b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 129222d5777bSStefano Zampini if (isseqaij) { 1293d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1294aa0d41d4SStefano Zampini } else { 1295aa0d41d4SStefano Zampini Mat work_mat; 1296aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1297d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1298aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1299aa0d41d4SStefano Zampini } 13003301b35fSStefano Zampini if (matis->A->symmetric_set) { 13013301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1302e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 13033301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1304e496cd5dSStefano Zampini #endif 13053301b35fSStefano Zampini } 130645a1bb75SStefano Zampini /* 130745a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1308d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 130945a1bb75SStefano Zampini */ 1310d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1311aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1312aa0d41d4SStefano Zampini } 1313aa0d41d4SStefano Zampini 1314aa0d41d4SStefano Zampini #undef __FUNCT__ 1315a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13168ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1317a64d13efSStefano Zampini { 1318a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1319a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1320d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 132153892102SStefano Zampini PetscInt *idx_R_local=NULL; 13223a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13233a50541eSStefano Zampini PetscInt vbs,bs; 13246816873aSStefano Zampini PetscBT bitmask=NULL; 1325a64d13efSStefano Zampini PetscErrorCode ierr; 1326a64d13efSStefano Zampini 1327a64d13efSStefano Zampini PetscFunctionBegin; 1328b23d619eSStefano Zampini /* 1329b23d619eSStefano Zampini No need to setup local scatters if 1330b23d619eSStefano Zampini - primal space is unchanged 1331b23d619eSStefano Zampini AND 1332b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1333b23d619eSStefano Zampini AND 1334b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1335b23d619eSStefano Zampini */ 1336b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1337f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1338f4ddd8eeSStefano Zampini } 1339f4ddd8eeSStefano Zampini /* destroy old objects */ 1340f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1341f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1342f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1343a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1344b371cd4fSStefano Zampini n_B = pcis->n_B; 1345b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1346b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13476816873aSStefano Zampini 13486816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13496816873aSStefano Zampini 135053892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13516816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 135253892102SStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13534641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13544641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13550e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13564641a718SStefano Zampini } 13573a50541eSStefano Zampini 1358a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13594641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13606816873aSStefano Zampini idx_R_local[n_R++] = i; 1361a64d13efSStefano Zampini } 1362a64d13efSStefano Zampini } 136353892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13646816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13656816873aSStefano Zampini 136653892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 136753892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13686816873aSStefano Zampini } 13693a50541eSStefano Zampini 13703a50541eSStefano Zampini /* Block code */ 13713a50541eSStefano Zampini vbs = 1; 13723a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13733a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13743a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13753a50541eSStefano Zampini PetscInt *vary; 13763a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1377785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13783a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13790e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1380964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13813a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13823a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13833a50541eSStefano Zampini break; 13843a50541eSStefano Zampini } 13853a50541eSStefano Zampini } 13863a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13873a50541eSStefano Zampini vbs = bs; 13883a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13893a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13903a50541eSStefano Zampini } 13913a50541eSStefano Zampini } 13923a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13933a50541eSStefano Zampini } 13943a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 139553892102SStefano Zampini if (sub_schurs->reuse_mumps) { 139653892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 139753892102SStefano Zampini 139853892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 139953892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 140053892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140153892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 140253892102SStefano Zampini } else { 14033a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 140453892102SStefano Zampini } 1405a64d13efSStefano Zampini 1406a64d13efSStefano Zampini /* print some info if requested */ 1407a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1408a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1409a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14100fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1411a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1412a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14133a50541eSStefano 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); 1414a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1415a64d13efSStefano Zampini } 1416a64d13efSStefano Zampini 1417a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14186816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14196816873aSStefano Zampini IS is_aux1,is_aux2; 14206816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14216816873aSStefano Zampini 14223a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1423854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1424854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1425a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14264641a718SStefano Zampini for (i=0; i<n_D; i++) { 14274641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14284641a718SStefano Zampini } 1429a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1430a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14314641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14324641a718SStefano Zampini aux_array1[j++] = i; 1433a64d13efSStefano Zampini } 1434a64d13efSStefano Zampini } 1435a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1436a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1437a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14384641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14394641a718SStefano Zampini aux_array2[j++] = i; 1440a64d13efSStefano Zampini } 1441a64d13efSStefano Zampini } 1442a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1443a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1444a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1445a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1446a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1447a64d13efSStefano Zampini 14488eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1449785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1450a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14514641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14524641a718SStefano Zampini aux_array1[j++] = i; 1453a64d13efSStefano Zampini } 1454a64d13efSStefano Zampini } 1455a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1456a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1457a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1458a64d13efSStefano Zampini } 14594641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14603a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1461d62866d3SStefano Zampini } else { 146253892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14636816873aSStefano Zampini IS tis; 14646816873aSStefano Zampini PetscInt schur_size; 14656816873aSStefano Zampini 146653892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14676816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 146853892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14696816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14706816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14716816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14726816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14736816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1474d62866d3SStefano Zampini } 1475d62866d3SStefano Zampini } 1476a64d13efSStefano Zampini PetscFunctionReturn(0); 1477a64d13efSStefano Zampini } 1478a64d13efSStefano Zampini 1479304d26faSStefano Zampini 1480304d26faSStefano Zampini #undef __FUNCT__ 1481304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1482684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1483304d26faSStefano Zampini { 1484304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1485304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1486304d26faSStefano Zampini PC pc_temp; 1487304d26faSStefano Zampini Mat A_RR; 1488f4ddd8eeSStefano Zampini MatReuse reuse; 1489304d26faSStefano Zampini PetscScalar m_one = -1.0; 1490304d26faSStefano Zampini PetscReal value; 1491af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14929577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1493304d26faSStefano Zampini PetscErrorCode ierr; 1494e604994aSStefano Zampini /* prefixes stuff */ 1495312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1496e604994aSStefano Zampini size_t len; 1497304d26faSStefano Zampini 1498304d26faSStefano Zampini PetscFunctionBegin; 1499304d26faSStefano Zampini 1500e604994aSStefano Zampini /* compute prefixes */ 1501e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1502e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1503e604994aSStefano Zampini if (!pcbddc->current_level) { 1504e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1505e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1506e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1507e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1508e604994aSStefano Zampini } else { 1509e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1510312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1511e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1512e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1513312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1514312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 151534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 151634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1517e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1518e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1519e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1520e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1521e604994aSStefano Zampini } 1522e604994aSStefano Zampini 1523304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1524684f6988SStefano Zampini if (dirichlet) { 1525d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15263301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15273301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1528964fefecSStefano Zampini } 1529ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1530964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1531304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1532304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1533304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1534304d26faSStefano Zampini /* default */ 1535304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1536e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15379577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1538304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15399577ea80SStefano Zampini if (issbaij) { 15409577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15419577ea80SStefano Zampini } else { 1542304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15439577ea80SStefano Zampini } 1544304d26faSStefano Zampini /* Allow user's customization */ 1545304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1546304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1547304d26faSStefano Zampini } 1548d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1549d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1550d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1551d62866d3SStefano Zampini 1552d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1553d5574798SStefano Zampini } 1554304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1555304d26faSStefano Zampini if (!n_D) { 1556304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1557304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1558304d26faSStefano Zampini } 1559304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1560304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1561304d26faSStefano Zampini /* set ksp_D into pcis data */ 1562304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1563304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1564304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1565684f6988SStefano Zampini } 1566304d26faSStefano Zampini 1567304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1568684f6988SStefano Zampini A_RR = 0; 1569684f6988SStefano Zampini if (neumann) { 1570d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1571f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15728ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1573f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1574f4ddd8eeSStefano Zampini PetscInt nn_R; 157581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1576f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1577f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1578f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1579f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1580f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1581f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1582f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1583727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1584f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1585f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1586f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1587f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1588f4ddd8eeSStefano Zampini } 1589f4ddd8eeSStefano Zampini } 1590f4ddd8eeSStefano Zampini /* last check */ 1591d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1592f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1593f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1594f4ddd8eeSStefano Zampini } 1595f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1596f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1597f4ddd8eeSStefano Zampini } 15986816873aSStefano Zampini /* extract A_RR */ 15996816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16006816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 16016816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 16026816873aSStefano Zampini if (ibs != mbs) { 16036816873aSStefano Zampini Mat newmat; 16046816873aSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 16056816873aSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16066816873aSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 16076816873aSStefano Zampini } else { 16086816873aSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16096816873aSStefano Zampini } 16103301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16113301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16126816873aSStefano Zampini } 16136816873aSStefano Zampini } else { 16146816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16156816873aSStefano Zampini 16166816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16176816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16186816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16196816873aSStefano Zampini } 1620f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1621304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1622304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1623304d26faSStefano Zampini /* default */ 1624304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1625e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1626304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16279577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16289577ea80SStefano Zampini if (issbaij) { 16299577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16309577ea80SStefano Zampini } else { 1631304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16329577ea80SStefano Zampini } 1633304d26faSStefano Zampini /* Allow user's customization */ 1634304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1635304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1636304d26faSStefano Zampini } 1637d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1638304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1639304d26faSStefano Zampini if (!n_R) { 1640304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1641304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1642304d26faSStefano Zampini } 1643d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1644d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1645d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1646d62866d3SStefano Zampini 1647d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1648d62866d3SStefano Zampini } 1649304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1650304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1651684f6988SStefano Zampini } 16526816873aSStefano Zampini /* free Neumann problem's matrix */ 16536816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1654304d26faSStefano Zampini 1655304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16560fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1657684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1658684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1659684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1660684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1661684f6988SStefano Zampini } 1662684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16630fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16640fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16650fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16660fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16670fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1668304d26faSStefano Zampini /* need to be adapted? */ 1669b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1670b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1671b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1672304d26faSStefano Zampini /* print info */ 1673304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1674e604994aSStefano 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); 1675304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1676304d26faSStefano Zampini } 1677b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16788ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1679304d26faSStefano Zampini } 1680684f6988SStefano Zampini } 1681684f6988SStefano Zampini if (neumann) { /* Neumann */ 16826816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16830fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16840fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16850fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16860fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16870fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1688304d26faSStefano Zampini /* need to be adapted? */ 1689b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1690b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1691304d26faSStefano Zampini /* print info */ 1692304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1693e604994aSStefano 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); 1694304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1695304d26faSStefano Zampini } 1696b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16978ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1698304d26faSStefano Zampini } 16990fccc4e9SStefano Zampini } 1700684f6988SStefano Zampini } 1701304d26faSStefano Zampini PetscFunctionReturn(0); 1702304d26faSStefano Zampini } 1703304d26faSStefano Zampini 1704304d26faSStefano Zampini #undef __FUNCT__ 1705ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 170680677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1707674ae819SStefano Zampini { 1708674ae819SStefano Zampini PetscErrorCode ierr; 1709674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1710be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1711674ae819SStefano Zampini 1712674ae819SStefano Zampini PetscFunctionBegin; 1713be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 171480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1715be83ff47SStefano Zampini } 171680677318SStefano Zampini if (!pcbddc->switch_static) { 171780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 171880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 171980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 172020c7b377SStefano Zampini } 1721be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172420c7b377SStefano Zampini } else { 1725be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1726be83ff47SStefano Zampini 172753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1729be83ff47SStefano Zampini } 1730be83ff47SStefano Zampini } else { 173180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173580677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173680677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 173780677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 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); 174020c7b377SStefano Zampini } 1741674ae819SStefano Zampini } 1742be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174380677318SStefano Zampini if (applytranspose) { 174480677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174580677318SStefano Zampini } else { 174680677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 174780677318SStefano Zampini } 174853892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1749be83ff47SStefano Zampini } else { 1750be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1751be83ff47SStefano Zampini 1752be83ff47SStefano Zampini if (applytranspose) { 175353892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1754be83ff47SStefano Zampini } else { 175553892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1756be83ff47SStefano Zampini } 175753892102SStefano Zampini #endif 1758be83ff47SStefano Zampini } 175980677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 176080677318SStefano Zampini if (!pcbddc->switch_static) { 1761be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 176280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1764be83ff47SStefano Zampini } else { 1765be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1766be83ff47SStefano Zampini 176753892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 176853892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1769be83ff47SStefano Zampini } 177080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 177180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 177280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 177380677318SStefano Zampini } 177480677318SStefano Zampini } else { 177580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 178280677318SStefano Zampini } 178380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178780677318SStefano Zampini } 1788674ae819SStefano Zampini PetscFunctionReturn(0); 1789674ae819SStefano Zampini } 1790674ae819SStefano Zampini 1791dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1792674ae819SStefano Zampini #undef __FUNCT__ 1793674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1794dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1795674ae819SStefano Zampini { 1796674ae819SStefano Zampini PetscErrorCode ierr; 1797674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1798674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1799674ae819SStefano Zampini const PetscScalar zero = 0.0; 1800674ae819SStefano Zampini 1801674ae819SStefano Zampini PetscFunctionBegin; 1802dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1803dc359a40SStefano Zampini if (applytranspose) { 1804674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18058eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1806dc359a40SStefano Zampini } else { 1807674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1808674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 180915aaf578SStefano Zampini } 181012edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 181112edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 181212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181412edc857SStefano Zampini 18159f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 181612edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 181712edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1818964fefecSStefano Zampini Vec rhs,sol; 1819964fefecSStefano Zampini 1820964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1821964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 182212edc857SStefano Zampini if (applytranspose) { 1823964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182412edc857SStefano Zampini } else { 1825964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 182612edc857SStefano Zampini } 182712edc857SStefano Zampini } 1828674ae819SStefano Zampini 1829674ae819SStefano Zampini /* Local solution on R nodes */ 183080677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 183180677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18329f00e9b4SStefano Zampini } 1833674ae819SStefano Zampini 18349f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18359f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 183612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1837674ae819SStefano Zampini 1838674ae819SStefano Zampini /* Sum contributions from two levels */ 1839dc359a40SStefano Zampini if (applytranspose) { 1840dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1841dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1842dc359a40SStefano Zampini } else { 1843674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18448eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1845dc359a40SStefano Zampini } 1846674ae819SStefano Zampini PetscFunctionReturn(0); 1847674ae819SStefano Zampini } 1848674ae819SStefano Zampini 1849674ae819SStefano Zampini #undef __FUNCT__ 1850674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 185112edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1852674ae819SStefano Zampini { 1853674ae819SStefano Zampini PetscErrorCode ierr; 1854674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 185558da7f69SStefano Zampini PetscScalar *array; 185612edc857SStefano Zampini Vec from,to; 1857674ae819SStefano Zampini 1858674ae819SStefano Zampini PetscFunctionBegin; 185912edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 186012edc857SStefano Zampini from = pcbddc->coarse_vec; 186112edc857SStefano Zampini to = pcbddc->vec1_P; 186212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 186312edc857SStefano Zampini Vec tvec; 186458da7f69SStefano Zampini 186558da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186658da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 186712edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 186858da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 186958da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 187058da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 187112edc857SStefano Zampini } 187212edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 187312edc857SStefano Zampini from = pcbddc->vec1_P; 187412edc857SStefano Zampini to = pcbddc->coarse_vec; 187512edc857SStefano Zampini } 187612edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1877674ae819SStefano Zampini PetscFunctionReturn(0); 1878674ae819SStefano Zampini } 1879674ae819SStefano Zampini 1880674ae819SStefano Zampini #undef __FUNCT__ 1881674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 188212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1883674ae819SStefano Zampini { 1884674ae819SStefano Zampini PetscErrorCode ierr; 1885674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 188658da7f69SStefano Zampini PetscScalar *array; 188712edc857SStefano Zampini Vec from,to; 1888674ae819SStefano Zampini 1889674ae819SStefano Zampini PetscFunctionBegin; 189012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 189112edc857SStefano Zampini from = pcbddc->coarse_vec; 189212edc857SStefano Zampini to = pcbddc->vec1_P; 189312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 189412edc857SStefano Zampini from = pcbddc->vec1_P; 189512edc857SStefano Zampini to = pcbddc->coarse_vec; 189612edc857SStefano Zampini } 189712edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 189812edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 189912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 190012edc857SStefano Zampini Vec tvec; 190158da7f69SStefano Zampini 190212edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 190358da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 190458da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 190558da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 190658da7f69SStefano Zampini } 190758da7f69SStefano Zampini } else { 190858da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 190958da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 191012edc857SStefano Zampini } 191112edc857SStefano Zampini } 1912674ae819SStefano Zampini PetscFunctionReturn(0); 1913674ae819SStefano Zampini } 1914674ae819SStefano Zampini 1915984c4197SStefano Zampini /* uncomment for testing purposes */ 1916984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1917674ae819SStefano Zampini #undef __FUNCT__ 1918674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1919674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1920674ae819SStefano Zampini { 1921674ae819SStefano Zampini PetscErrorCode ierr; 1922674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1923674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1924674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1925984c4197SStefano Zampini /* one and zero */ 1926984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1927984c4197SStefano Zampini /* space to store constraints and their local indices */ 19289162d606SStefano Zampini PetscScalar *constraints_data; 19299162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19309162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19319162d606SStefano Zampini PetscInt *constraints_n; 1932984c4197SStefano Zampini /* iterators */ 1933b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1934984c4197SStefano Zampini /* BLAS integers */ 1935e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1936e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1937c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1938727cdba6SStefano Zampini /* reuse */ 19390e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19400e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1941984c4197SStefano Zampini /* change of basis */ 1942b3d85658SStefano Zampini PetscBool qr_needed; 19439162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1944984c4197SStefano Zampini /* auxiliary stuff */ 194564efe560SStefano Zampini PetscInt *nnz,*is_indices; 19468a0068c3SStefano Zampini PetscInt ncc; 1947984c4197SStefano Zampini /* some quantities */ 194845a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1949a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1950984c4197SStefano Zampini 1951674ae819SStefano Zampini PetscFunctionBegin; 19528e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19538e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19548e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1955088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1956088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19570e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19580e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19590e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19600e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19610e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1962088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1963cf5a6209SStefano Zampini 1964cf5a6209SStefano Zampini /* print some info */ 1965cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1966cf5a6209SStefano Zampini IS vertices; 1967cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1968cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1969cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1970cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1971cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1972cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1973cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1974fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1975fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1976cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1977cf5a6209SStefano Zampini } 1978cf5a6209SStefano Zampini 1979cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19809162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1981cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1982cf5a6209SStefano Zampini const Vec *nearnullvecs; 1983cf5a6209SStefano Zampini Vec *localnearnullsp; 1984cf5a6209SStefano Zampini PetscScalar *array; 1985cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1986cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1987cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1988b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1989cf5a6209SStefano Zampini PetscScalar *work; 1990cf5a6209SStefano Zampini PetscReal *singular_vals; 1991cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1992cf5a6209SStefano Zampini PetscReal *rwork; 1993cf5a6209SStefano Zampini #endif 1994cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1995cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1996cf5a6209SStefano Zampini #else 1997964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1998964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1999cf5a6209SStefano Zampini #endif 2000cf5a6209SStefano Zampini 2001674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2002d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2003d06fc5fdSStefano Zampini /* free unneeded index sets */ 2004d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2005d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2006674ae819SStefano Zampini } 2007d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2008d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2009d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2010d06fc5fdSStefano Zampini } 2011d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2012d06fc5fdSStefano Zampini n_ISForEdges = 0; 2013d06fc5fdSStefano Zampini } 2014d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2015d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2016d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2017d06fc5fdSStefano Zampini } 2018d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2019d06fc5fdSStefano Zampini n_ISForFaces = 0; 2020d06fc5fdSStefano Zampini } 202170022509SStefano Zampini 202270022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 202370022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 202470022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 202570022509SStefano Zampini if (pcbddc->NullSpace) { 202670022509SStefano Zampini PetscBool tbool[2],gbool[2]; 202770022509SStefano Zampini 202870022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2029b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2030d06fc5fdSStefano Zampini if (!ISForEdges) { 2031d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2032d06fc5fdSStefano Zampini } 2033b8ffe317SStefano Zampini } 2034d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2035d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2036d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2037d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2038d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 203998a51de6SStefano Zampini } 204070022509SStefano Zampini #endif 204108122e43SStefano Zampini 2042674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2043674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2044674ae819SStefano Zampini if (nearnullsp) { 2045674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2046f4ddd8eeSStefano Zampini /* remove any stored info */ 2047f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2048f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2049f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2050f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2051f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2052473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2053f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2054f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2055f4ddd8eeSStefano Zampini } 2056984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2057984c4197SStefano Zampini nnsp_size = 0; 2058674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2059674ae819SStefano Zampini } 2060984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2061984c4197SStefano Zampini max_constraints = nnsp_size; 2062984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2063984c4197SStefano Zampini 2064674ae819SStefano Zampini /* 2065674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20669162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20679162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20689162d606SStefano Zampini There can be multiple constraints per connected component 2069674ae819SStefano Zampini */ 2070674ae819SStefano Zampini n_vertices = 0; 2071674ae819SStefano Zampini if (ISForVertices) { 2072674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2073674ae819SStefano Zampini } 20749162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20759162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20769162d606SStefano Zampini 20779162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20789162d606SStefano Zampini total_counts *= max_constraints; 2079674ae819SStefano Zampini total_counts += n_vertices; 20804641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20819162d606SStefano Zampini 2082674ae819SStefano Zampini total_counts = 0; 2083674ae819SStefano Zampini max_size_of_constraint = 0; 2084674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20859162d606SStefano Zampini IS used_is; 2086674ae819SStefano Zampini if (i<n_ISForEdges) { 20879162d606SStefano Zampini used_is = ISForEdges[i]; 2088674ae819SStefano Zampini } else { 20899162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2090674ae819SStefano Zampini } 20919162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2092674ae819SStefano Zampini total_counts += j; 2093674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2094674ae819SStefano Zampini } 20959162d606SStefano 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); 20969162d606SStefano Zampini 2097984c4197SStefano Zampini /* get local part of global near null space vectors */ 2098785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2099984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2100984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2101984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2102984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2103984c4197SStefano Zampini } 2104674ae819SStefano Zampini 2105242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2106242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2107a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2108242a89d7SStefano Zampini 2109984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2110a773dcb8SStefano Zampini if (!skip_lapack) { 2111674ae819SStefano Zampini PetscScalar temp_work; 2112911cabfeSStefano Zampini 2113674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2114984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2115785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2116785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2117785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2118674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2119785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2120674ae819SStefano Zampini #endif 2121674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2122c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2123c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2124674ae819SStefano Zampini lwork = -1; 2125674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2126674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2127c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2128674ae819SStefano Zampini #else 2129c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2130674ae819SStefano Zampini #endif 2131674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2132984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2133674ae819SStefano Zampini #else /* on missing GESVD */ 2134674ae819SStefano Zampini /* SVD */ 2135674ae819SStefano Zampini PetscInt max_n,min_n; 2136674ae819SStefano Zampini max_n = max_size_of_constraint; 2137984c4197SStefano Zampini min_n = max_constraints; 2138984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2139674ae819SStefano Zampini min_n = max_size_of_constraint; 2140984c4197SStefano Zampini max_n = max_constraints; 2141674ae819SStefano Zampini } 2142785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2143674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2144785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2145674ae819SStefano Zampini #endif 2146674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2147674ae819SStefano Zampini lwork = -1; 2148e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2149e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2150b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2151674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2152674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21539162d606SStefano 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)); 2154674ae819SStefano Zampini #else 21559162d606SStefano 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)); 2156674ae819SStefano Zampini #endif 2157674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2158984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2159984c4197SStefano Zampini #endif /* on missing GESVD */ 2160674ae819SStefano Zampini /* Allocate optimal workspace */ 2161674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2162854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2163674ae819SStefano Zampini } 2164674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2165674ae819SStefano Zampini total_counts = 0; 21669162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21679162d606SStefano Zampini constraints_data_ptr[0] = 0; 2168674ae819SStefano Zampini /* vertices */ 21699162d606SStefano Zampini if (n_vertices) { 2170674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21719162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21729162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2173674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21749162d606SStefano Zampini constraints_n[total_counts] = 1; 21759162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21769162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21779162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2178674ae819SStefano Zampini total_counts++; 2179674ae819SStefano Zampini } 2180674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2181984c4197SStefano Zampini PetscBool used_vertex; 2182674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2183674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2184674ae819SStefano Zampini k = 0; 2185674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2186984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2187984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21889162d606SStefano Zampini constraints_n[total_counts] = 1; 21899162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21909162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21919162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21929162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2193674ae819SStefano Zampini total_counts++; 2194674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2195674ae819SStefano Zampini } 2196984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2197674ae819SStefano Zampini k++; 2198674ae819SStefano Zampini } 2199674ae819SStefano Zampini } 2200674ae819SStefano Zampini } 2201674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2202674ae819SStefano Zampini n_vertices = total_counts; 2203674ae819SStefano Zampini } 2204984c4197SStefano Zampini 2205674ae819SStefano Zampini /* edges and faces */ 22069162d606SStefano Zampini total_counts_cc = total_counts; 2207911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22089162d606SStefano Zampini IS used_is; 22099162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22109162d606SStefano Zampini 2211911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22129162d606SStefano Zampini used_is = ISForEdges[ncc]; 2213984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2214674ae819SStefano Zampini } else { 22159162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2216984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2217674ae819SStefano Zampini } 2218674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22199162d606SStefano Zampini 22209162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22219162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2222984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2223984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2224674ae819SStefano Zampini if (nnsp_has_cnst) { 22255b08dc53SStefano Zampini PetscScalar quad_value; 22269162d606SStefano Zampini 22279162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22289162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22299162d606SStefano Zampini 2230a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2231674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2232a773dcb8SStefano Zampini } else { 2233a773dcb8SStefano Zampini quad_value = 1.0; 2234a773dcb8SStefano Zampini } 2235674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22369162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2237674ae819SStefano Zampini } 22389162d606SStefano Zampini temp_constraints++; 2239674ae819SStefano Zampini total_counts++; 2240674ae819SStefano Zampini } 2241674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2242984c4197SStefano Zampini PetscReal real_value; 22439162d606SStefano Zampini PetscScalar *ptr_to_data; 22449162d606SStefano Zampini 2245984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22469162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2247674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22489162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2249674ae819SStefano Zampini } 2250984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2251984c4197SStefano Zampini /* check if array is null on the connected component */ 2252e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22539162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22545b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2255674ae819SStefano Zampini temp_constraints++; 2256674ae819SStefano Zampini total_counts++; 22579162d606SStefano Zampini if (!idxs_copied) { 22589162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22599162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2260674ae819SStefano Zampini } 2261674ae819SStefano Zampini } 22629162d606SStefano Zampini } 22639162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 226445a1bb75SStefano Zampini valid_constraints = temp_constraints; 2265eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2266a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22679162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22689162d606SStefano Zampini 22699162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2270a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22719162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2272a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22739162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2274a773dcb8SStefano Zampini } else { /* perform SVD */ 2275984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22769162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2277674ae819SStefano Zampini 2278674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2279984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2280984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2281984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2282984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2283984c4197SStefano Zampini from that computed using LAPACKgesvd 2284984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2285984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2286984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2287674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2288e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2289984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2290674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2291674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22929162d606SStefano 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)); 2293674ae819SStefano Zampini } 2294674ae819SStefano Zampini } 2295e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2296e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2297e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2298674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2299c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2300674ae819SStefano Zampini #else 2301c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2302674ae819SStefano Zampini #endif 2303674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2304984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2305984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2306674ae819SStefano Zampini j = 0; 2307984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2308674ae819SStefano Zampini total_counts = total_counts-j; 230945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2310e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2311c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2312c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2313c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2314c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2315c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2316c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2317674ae819SStefano Zampini if (j<temp_constraints) { 2318984c4197SStefano Zampini PetscInt ii; 2319984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2320674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23219162d606SStefano 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)); 2322674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2323984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2324674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23259162d606SStefano 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]; 2326674ae819SStefano Zampini } 2327674ae819SStefano Zampini } 2328674ae819SStefano Zampini } 2329674ae819SStefano Zampini #else /* on missing GESVD */ 2330e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2331e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2332b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2333674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2334674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23359162d606SStefano 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)); 2336674ae819SStefano Zampini #else 23379162d606SStefano 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)); 2338674ae819SStefano Zampini #endif 2339984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2340674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2341984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2342e310c8b4SStefano Zampini k = temp_constraints; 2343e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2344674ae819SStefano Zampini j = 0; 2345e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 234645a1bb75SStefano Zampini valid_constraints = k-j; 2347911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2348984c4197SStefano Zampini #endif /* on missing GESVD */ 2349674ae819SStefano Zampini } 2350a773dcb8SStefano Zampini } 23519162d606SStefano Zampini /* update pointers information */ 23529162d606SStefano Zampini if (valid_constraints) { 23539162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23549162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23559162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23569162d606SStefano Zampini /* set change_of_basis flag */ 235745a1bb75SStefano Zampini if (boolforchange) { 2358b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23599162d606SStefano Zampini } 2360b3d85658SStefano Zampini total_counts_cc++; 236145a1bb75SStefano Zampini } 2362674ae819SStefano Zampini } 2363984c4197SStefano Zampini /* free workspace */ 23648f1c130eSStefano Zampini if (!skip_lapack) { 2365984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2366984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2367984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2368984c4197SStefano Zampini #endif 2369984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2370984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2371984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2372984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2373984c4197SStefano Zampini #endif 2374984c4197SStefano Zampini } 2375984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2376984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2377984c4197SStefano Zampini } 2378984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2379cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2380cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2381cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2382cf5a6209SStefano Zampini } 2383cf5a6209SStefano Zampini if (n_ISForFaces) { 2384cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2385cf5a6209SStefano Zampini } 2386cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2387cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2388cf5a6209SStefano Zampini } 2389cf5a6209SStefano Zampini if (n_ISForEdges) { 2390cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2391cf5a6209SStefano Zampini } 2392cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 239308122e43SStefano Zampini } else { 239408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 239508122e43SStefano Zampini 239608122e43SStefano Zampini total_counts = 0; 239708122e43SStefano Zampini n_vertices = 0; 2398d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2399d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 240008122e43SStefano Zampini } 240108122e43SStefano Zampini max_constraints = 0; 24029162d606SStefano Zampini total_counts_cc = 0; 240308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 240408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24059162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 240608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 240708122e43SStefano Zampini } 24089162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24099162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24109162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24119162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 241274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24139162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24149162d606SStefano Zampini total_counts_cc = 0; 24159162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24169162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24179162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 241808122e43SStefano Zampini } 241908122e43SStefano Zampini } 24209162d606SStefano Zampini #if 0 24219162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24229162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24239162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24249162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24259162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24269162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24279162d606SStefano Zampini } 24289162d606SStefano Zampini printf("\n"); 24299162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24309162d606SStefano Zampini } 24311b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24328bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24331b968477SStefano Zampini } 24341b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24358bec7fa6SStefano 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]); 24361b968477SStefano Zampini } 243708122e43SStefano Zampini #endif 243808122e43SStefano Zampini 24398bec7fa6SStefano Zampini max_size_of_constraint = 0; 24409162d606SStefano 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]); 24419162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 244208122e43SStefano Zampini /* Change of basis */ 2443b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 244408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 244508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 244608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2447b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 244808122e43SStefano Zampini } 244908122e43SStefano Zampini } 245008122e43SStefano Zampini } 245108122e43SStefano Zampini } 2452b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24530e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 245408122e43SStefano Zampini 24559162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24569162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24579162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24589162d606SStefano 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); 245908122e43SStefano Zampini } 2460984c4197SStefano Zampini 2461674ae819SStefano Zampini /* Create constraint matrix */ 2462674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 246316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2464984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2465984c4197SStefano Zampini 2466984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 246774d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 246874d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 246974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2470984c4197SStefano Zampini total_primal_vertices=0; 2471b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24729162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24739162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2474984c4197SStefano Zampini if (size_of_constraint == 1) { 24759162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2476b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 247764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24789162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24799162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24809162d606SStefano Zampini } 2481b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 248274d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 248374d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 248474d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 248574d5cdf7SStefano Zampini } 2486b3d85658SStefano Zampini } else { 2487b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2488984c4197SStefano Zampini } 2489984c4197SStefano Zampini } 2490b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2491b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 249270022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 249370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2494b3d85658SStefano Zampini 2495b3d85658SStefano 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); 24960e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24970e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 249870022509SStefano Zampini 2499984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 250074d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2501785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2502984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 250374d5cdf7SStefano Zampini 2504984c4197SStefano Zampini j = total_primal_vertices; 250574d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2506b3d85658SStefano Zampini cum = total_primal_vertices; 25079162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25084641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2509b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2510b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2511b3d85658SStefano Zampini cum++; 25129162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 251374d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 251474d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 251574d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 251674d5cdf7SStefano Zampini } 25179162d606SStefano Zampini j += constraints_n[i]; 2518674ae819SStefano Zampini } 2519674ae819SStefano Zampini } 2520674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2521674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2522088faed8SStefano Zampini 2523674ae819SStefano Zampini /* set values in constraint matrix */ 2524984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25250e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2526674ae819SStefano Zampini } 2527984c4197SStefano Zampini total_counts = total_primal_vertices; 25289162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25294641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25309162d606SStefano Zampini PetscInt *cols; 25319162d606SStefano Zampini 25329162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25339162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25349162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25359162d606SStefano Zampini PetscInt row = total_counts+k; 25369162d606SStefano Zampini PetscScalar *vals; 25379162d606SStefano Zampini 25389162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25399162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25409162d606SStefano Zampini } 25419162d606SStefano Zampini total_counts += constraints_n[i]; 2542674ae819SStefano Zampini } 2543674ae819SStefano Zampini } 2544674ae819SStefano Zampini /* assembling */ 2545674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2546674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2547088faed8SStefano Zampini 2548984c4197SStefano Zampini /* 254945a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2550984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2551984c4197SStefano Zampini */ 2552674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2553674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2554026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2555984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2556984c4197SStefano Zampini /* working stuff for GEQRF */ 255781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2558984c4197SStefano Zampini PetscBLASInt lqr_work; 2559984c4197SStefano Zampini /* working stuff for UNGQR */ 2560984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2561984c4197SStefano Zampini PetscBLASInt lgqr_work; 2562984c4197SStefano Zampini /* working stuff for TRTRS */ 2563984c4197SStefano Zampini PetscScalar *trs_rhs; 25643f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2565984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2566984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2567984c4197SStefano Zampini PetscScalar *start_vals; 2568984c4197SStefano Zampini /* working stuff for values insertion */ 25694641a718SStefano Zampini PetscBT is_primal; 257064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2571906d46d4SStefano Zampini /* matrix sizes */ 2572906d46d4SStefano Zampini PetscInt global_size,local_size; 2573906d46d4SStefano Zampini /* temporary change of basis */ 2574906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2575cf5a6209SStefano Zampini /* extra space for debugging */ 2576cf5a6209SStefano Zampini PetscScalar *dbg_work; 2577984c4197SStefano Zampini 2578906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2579906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 258016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2581bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2582906d46d4SStefano Zampini /* nonzeros for local mat */ 2583bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2584bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25859162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2586a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25879162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2588a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25899162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2590a717540cSStefano Zampini } else { 25919162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25929162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2593a717540cSStefano Zampini } 2594a717540cSStefano Zampini } 2595a717540cSStefano Zampini } 2596906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2597bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2598a717540cSStefano Zampini /* Set initial identity in the matrix */ 2599bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2600906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2601a717540cSStefano Zampini } 2602a717540cSStefano Zampini 2603a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2604a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2605a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2606a717540cSStefano Zampini } 2607a717540cSStefano Zampini 2608a717540cSStefano Zampini 2609a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2610a717540cSStefano Zampini /* 2611a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2612a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2613a717540cSStefano Zampini 2614a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2615a717540cSStefano Zampini 2616a6b551f4SStefano 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) 2617a6b551f4SStefano Zampini 2618a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2619a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2620a717540cSStefano Zampini | ... | 2621a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2622a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2623a717540cSStefano Zampini 2624a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2625a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2626a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2627a6b551f4SStefano Zampini 2628a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2629a717540cSStefano Zampini */ 2630a717540cSStefano Zampini if (qr_needed) { 2631984c4197SStefano Zampini /* space to store Q */ 2632854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2633984c4197SStefano Zampini /* first we issue queries for optimal work */ 26343f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26353f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26363f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2637984c4197SStefano Zampini lqr_work = -1; 26383f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2639984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2640984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2641785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2642984c4197SStefano Zampini lgqr_work = -1; 26433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26443f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26453f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26463f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26473f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26483f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2649984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2650984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2651785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2652984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2653785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2654984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2655785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2656a717540cSStefano Zampini /* allocating workspace for check */ 2657a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2658cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2659a717540cSStefano Zampini } 2660a717540cSStefano Zampini } 2661984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26624641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2663473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26640e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 266539e2fb2aSStefano Zampini if (i != total_primal_vertices) { 266639e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26674641a718SStefano Zampini } 266839e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 266939e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 267039e2fb2aSStefano Zampini } 267139e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2672984c4197SStefano Zampini 2673a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26749162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26759162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26764641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2677984c4197SStefano Zampini /* get constraint info */ 26789162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2679984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2680984c4197SStefano Zampini 2681984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26829162d606SStefano 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); 2683674ae819SStefano Zampini } 2684984c4197SStefano Zampini 2685fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2686a717540cSStefano Zampini 2687a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2688a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26899162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2690a717540cSStefano Zampini } 2691984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26929162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2693984c4197SStefano Zampini 2694984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26953f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26963f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26973f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2698674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26993f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2700984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2701674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2702984c4197SStefano Zampini 2703984c4197SStefano Zampini /* explictly compute R^-T */ 2704984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2705984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27063f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27073f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27093f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2710984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27113f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2712984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2713984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2714984c4197SStefano Zampini 2715a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27183f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27193f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2720984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27213f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2722984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2723984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2724984c4197SStefano Zampini 2725984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2726984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2727984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27283f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27293f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27303f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27323f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27333f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2734984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27359162d606SStefano 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)); 2736984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27379162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2738984c4197SStefano Zampini 2739984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27409162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2741984c4197SStefano Zampini /* insert cols for primal dofs */ 2742984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2743984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27449162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2745906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2746984c4197SStefano Zampini } 2747984c4197SStefano Zampini /* insert cols for dual dofs */ 2748984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27499162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2750984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27519162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2752906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2753984c4197SStefano Zampini j++; 2754674ae819SStefano Zampini } 2755674ae819SStefano Zampini } 2756984c4197SStefano Zampini 2757984c4197SStefano Zampini /* check change of basis */ 2758984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2759984c4197SStefano Zampini PetscInt ii,jj; 2760984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2761c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2762c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2763c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2764c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2765c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2766c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2767984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2768cf5a6209SStefano 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)); 2769984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2770984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2771984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2772cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2773cf5a6209SStefano 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; 2774674ae819SStefano Zampini } 2775674ae819SStefano Zampini } 2776984c4197SStefano Zampini if (!valid_qr) { 277722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2778984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2779984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2780cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2781cf5a6209SStefano 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])); 2782674ae819SStefano Zampini } 2783cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2784cf5a6209SStefano 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])); 2785984c4197SStefano Zampini } 2786984c4197SStefano Zampini } 2787984c4197SStefano Zampini } 2788674ae819SStefano Zampini } else { 278922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2790674ae819SStefano Zampini } 2791674ae819SStefano Zampini } 2792a717540cSStefano Zampini } else { /* simple transformation block */ 2793a717540cSStefano Zampini PetscInt row,col; 2794a6b551f4SStefano Zampini PetscScalar val,norm; 2795a6b551f4SStefano Zampini 2796a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27979162d606SStefano 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)); 2798a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27999162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28009162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2801bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28029162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2803906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28049162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2805a717540cSStefano Zampini } else { 2806a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28079162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2808a717540cSStefano Zampini if (row != col) { 28099162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2810a717540cSStefano Zampini } else { 28119162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2812a717540cSStefano Zampini } 2813906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2814a717540cSStefano Zampini } 2815a717540cSStefano Zampini } 2816a717540cSStefano Zampini } 281798a51de6SStefano Zampini if (pcbddc->dbg_flag) { 281822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2819a717540cSStefano Zampini } 2820674ae819SStefano Zampini } 2821984c4197SStefano Zampini } else { 2822984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28239162d606SStefano 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); 2824674ae819SStefano Zampini } 2825674ae819SStefano Zampini } 2826674ae819SStefano Zampini } 2827a717540cSStefano Zampini 2828a717540cSStefano Zampini /* free workspace */ 2829a717540cSStefano Zampini if (qr_needed) { 2830984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2831cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2832984c4197SStefano Zampini } 2833984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2834984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2835984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2836984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2837984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2838674ae819SStefano Zampini } 2839a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2840906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2841906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2842906d46d4SStefano Zampini 2843906d46d4SStefano Zampini /* assembling of global change of variable */ 2844bbb9e6c6SStefano Zampini { 2845bbb9e6c6SStefano Zampini Mat tmat; 284616f15bc4SStefano Zampini PetscInt bs; 284716f15bc4SStefano Zampini 2848906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2849906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2850bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2851bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2852bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2853bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 285416f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 285516f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2856906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2857bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2858bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2859bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2860bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2861bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2862bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2863bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2866906d46d4SStefano Zampini } 2867906d46d4SStefano Zampini /* check */ 2868906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2869906d46d4SStefano Zampini PetscReal error; 2870906d46d4SStefano Zampini Vec x,x_change; 2871906d46d4SStefano Zampini 2872906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2873906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2874906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2875906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2878bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2879bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2880bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2881906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2882906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2883906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2884906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2885bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2886906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2887906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2888906d46d4SStefano Zampini } 2889b96c3477SStefano Zampini 2890b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2891b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2892b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2893b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2894ac632422SStefano Zampini Mat S_new,tmat; 2895bbb9e6c6SStefano Zampini IS is_all_N; 2896bbb9e6c6SStefano Zampini 2897bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28986816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2899bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2900ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2901b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2902ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2903ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2904ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2905ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2906ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2907b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2908ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2909ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2910ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2911ac632422SStefano Zampini } 2912b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2913b96c3477SStefano Zampini } 2914b96c3477SStefano Zampini } 2915906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2916906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2917b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2918b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2919b9b85e73SStefano Zampini } 2920906d46d4SStefano Zampini 2921906d46d4SStefano Zampini /* set up change of basis context */ 2922906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2923906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2924906d46d4SStefano Zampini 2925906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2926906d46d4SStefano Zampini PetscInt global_size,local_size; 2927906d46d4SStefano Zampini 2928906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2929906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2930906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2931906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2932906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2933906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2934906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2935906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2936906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2937906d46d4SStefano Zampini } else { 2938906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2939906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2940906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2941906d46d4SStefano Zampini } 2942906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2943906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2944906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2945906d46d4SStefano Zampini } else { 2946906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2947906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2948906d46d4SStefano Zampini } 2949906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2951906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2952906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2953b9b85e73SStefano Zampini } 2954a717540cSStefano Zampini 2955727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2956727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2957727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2958aff50787SStefano 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); 29590e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29600e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2961aff50787SStefano 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); 2962c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2963727cdba6SStefano Zampini } 29640e6343abSStefano Zampini } 29650e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2966727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2967727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2968727cdba6SStefano Zampini 2969a717540cSStefano Zampini /* flush dbg viewer */ 2970b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2971b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2972b8ffe317SStefano Zampini } 2973a717540cSStefano Zampini 2974e310c8b4SStefano Zampini /* free workspace */ 2975a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29764641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 297708122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29789162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29799162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 298008122e43SStefano Zampini } else { 29819162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29829162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29839162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 298408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 298508122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29869162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29879162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 298808122e43SStefano Zampini } 2989674ae819SStefano Zampini PetscFunctionReturn(0); 2990674ae819SStefano Zampini } 2991674ae819SStefano Zampini 2992674ae819SStefano Zampini #undef __FUNCT__ 2993674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2994674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2995674ae819SStefano Zampini { 2996674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2997674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2998674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29997fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3000674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3001674ae819SStefano Zampini 3002674ae819SStefano Zampini PetscFunctionBegin; 30038e61c736SStefano Zampini /* Reset previously computed graph */ 30048e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3005674ae819SStefano Zampini /* Init local Graph struct */ 30067fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30077fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3008674ae819SStefano Zampini 3009575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3010575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3011575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3012575ad6abSStefano Zampini } 30139577ea80SStefano Zampini 3014674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30154d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30164d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30174d379d7bSStefano Zampini PetscInt nvtxs; 3018e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3019674ae819SStefano Zampini 30204d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30212fffb893SStefano Zampini 30222fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30232fffb893SStefano Zampini if (flg_row) { 30244d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3025b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30262fffb893SStefano Zampini } 30272fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30289b28b941SStefano 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 */ 30294d379d7bSStefano Zampini IS is_dummy; 30304d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30314d379d7bSStefano Zampini PetscInt j,sum; 30324d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30334d379d7bSStefano Zampini const PetscInt *idxs; 30344d379d7bSStefano Zampini PCBDDCGraph graph; 30354d379d7bSStefano Zampini PetscBT is_on_boundary; 30364d379d7bSStefano Zampini 30374d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30384d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30394d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30404d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30417fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30424d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3043e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3044e496cd5dSStefano Zampini if (flg_row) { 30454d379d7bSStefano Zampini graph->xadj = xadj; 30464d379d7bSStefano Zampini graph->adjncy = adjncy; 3047e496cd5dSStefano Zampini } 30484d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30494d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3050e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30514d379d7bSStefano Zampini 30524d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30539b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30544d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30554d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30564d379d7bSStefano Zampini } 30574d379d7bSStefano Zampini } 30584d379d7bSStefano Zampini 3059e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30604d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30614d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30624d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30634d379d7bSStefano Zampini } 30644d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30654d379d7bSStefano Zampini 3066e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30674d379d7bSStefano Zampini sum = 0; 30684d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30694d379d7bSStefano Zampini PetscInt sizecc = 0; 30704d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30714d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30724d379d7bSStefano Zampini sizecc++; 30734d379d7bSStefano Zampini } 30744d379d7bSStefano Zampini } 30754d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30764d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30774d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30784d379d7bSStefano Zampini } 30794d379d7bSStefano Zampini } 30804d379d7bSStefano Zampini sum += sizecc*sizecc; 30814d379d7bSStefano Zampini } 30824d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30834d379d7bSStefano Zampini sum = 0; 3084e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30854d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30864d379d7bSStefano Zampini cxadj[i] = sum; 30874d379d7bSStefano Zampini sum += temp; 30884d379d7bSStefano Zampini } 3089e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30904d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30914d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30924d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30934d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30944d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30954d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30964d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30974d379d7bSStefano Zampini sizecc++; 30984d379d7bSStefano Zampini } 30994d379d7bSStefano Zampini } 31004d379d7bSStefano Zampini } 31014d379d7bSStefano Zampini } 31024d379d7bSStefano Zampini } 31039b28b941SStefano Zampini if (sum) { 3104e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31054d379d7bSStefano Zampini } else { 31064d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31074d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31084d379d7bSStefano Zampini } 31094d379d7bSStefano Zampini graph->xadj = 0; 31104d379d7bSStefano Zampini graph->adjncy = 0; 31114d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31124d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31134d379d7bSStefano Zampini } 3114674ae819SStefano Zampini } 31159b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31169b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31179b28b941SStefano Zampini } 3118674ae819SStefano Zampini 311963602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3120674ae819SStefano Zampini vertex_size = 1; 312163602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 312263602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 312395ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 312463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3125a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 312663602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3127674ae819SStefano Zampini } 312863602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 312963602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 313063602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3131674ae819SStefano Zampini } 313263602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3133674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 313463602bcaSStefano Zampini } else { 313563602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 313663602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3137854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 313963602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314063602bcaSStefano Zampini } 314163602bcaSStefano Zampini } 3142674ae819SStefano Zampini } 3143674ae819SStefano Zampini 3144674ae819SStefano Zampini /* Setup of Graph */ 3145785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3146a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3147785d1243SStefano Zampini } 3148785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3149a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3150785d1243SStefano Zampini } 3151302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3152674ae819SStefano Zampini 3153674ae819SStefano Zampini /* Graph's connected components analysis */ 3154674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3155674ae819SStefano Zampini 3156674ae819SStefano Zampini /* print some info to stdout */ 3157674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3158302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3159674ae819SStefano Zampini } 3160fb180af4SStefano Zampini 3161fb180af4SStefano Zampini /* mark topography has done */ 3162fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3163674ae819SStefano Zampini PetscFunctionReturn(0); 3164674ae819SStefano Zampini } 3165674ae819SStefano Zampini 3166674ae819SStefano Zampini #undef __FUNCT__ 3167674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3168674ae819SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(MPI_Comm comm,ISLocalToGlobalMapping l2gmap, PetscInt n_local_dofs, PetscInt local_dofs[], PetscInt local_dofs_mult[], PetscInt* n_global_subset, PetscInt* global_numbering_subset[]) 3169674ae819SStefano Zampini { 3170674ae819SStefano Zampini Vec local_vec,global_vec; 3171674ae819SStefano Zampini IS seqis,paris; 3172674ae819SStefano Zampini VecScatter scatter_ctx; 3173674ae819SStefano Zampini PetscScalar *array; 3174674ae819SStefano Zampini PetscInt *temp_global_dofs; 3175674ae819SStefano Zampini PetscScalar globalsum; 3176674ae819SStefano Zampini PetscInt i,j,s; 317717ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 317817ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 317917ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3180674ae819SStefano Zampini PetscBool first_found; 3181674ae819SStefano Zampini PetscErrorCode ierr; 3182674ae819SStefano Zampini 3183674ae819SStefano Zampini PetscFunctionBegin; 3184674ae819SStefano Zampini /* mpi buffers */ 3185b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3186b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3187674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 318817ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3189674ae819SStefano Zampini /* get maximum size of subset */ 3190785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3191674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3192674ae819SStefano Zampini max_local = 0; 3193b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3194674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3195674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3196674ae819SStefano Zampini } 3197674ae819SStefano Zampini } 3198b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3199674ae819SStefano Zampini max_global++; 3200674ae819SStefano Zampini max_local = 0; 3201b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3202674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3203674ae819SStefano Zampini max_local = local_dofs[i]; 3204674ae819SStefano Zampini } 3205674ae819SStefano Zampini } 3206674ae819SStefano Zampini max_local++; 3207674ae819SStefano Zampini /* allocate workspace */ 3208674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3209674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3210674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3211674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3212674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3213674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3214674ae819SStefano Zampini /* create scatter */ 3215674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3216674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3217674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3218674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3219674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3220674ae819SStefano Zampini /* init array */ 3221674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3222674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3223674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3224674ae819SStefano Zampini if (local_dofs_mult) { 3225674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3226674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3227674ae819SStefano Zampini } 3228674ae819SStefano Zampini } else { 3229674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3230674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3231674ae819SStefano Zampini } 3232674ae819SStefano Zampini } 3233674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3234674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3235674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3236674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3237674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 32385b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3239674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3240674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3241674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3242674ae819SStefano Zampini nlocals = 0; 3243674ae819SStefano Zampini first_index = -1; 3244674ae819SStefano Zampini first_found = PETSC_FALSE; 3245674ae819SStefano Zampini for (i=0;i<s;i++) { 3246b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3247674ae819SStefano Zampini first_found = PETSC_TRUE; 3248674ae819SStefano Zampini first_index = i; 3249674ae819SStefano Zampini } 32505b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3251674ae819SStefano Zampini } 3252674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3253674ae819SStefano Zampini if (!rank_prec_comm) { 3254674ae819SStefano Zampini dof_displs[0]=0; 3255674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3256674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3257674ae819SStefano Zampini } 3258674ae819SStefano Zampini } 3259674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3260674ae819SStefano Zampini if (first_found) { 3261674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3262674ae819SStefano Zampini old_index = first_index; 3263674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3264b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3265674ae819SStefano Zampini array[i] += array[old_index]; 3266674ae819SStefano Zampini old_index = i; 3267674ae819SStefano Zampini } 3268674ae819SStefano Zampini } 3269674ae819SStefano Zampini } 3270674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3271674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3272674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3273674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3274674ae819SStefano Zampini /* get global ordering of local dofs */ 32752b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3276674ae819SStefano Zampini if (local_dofs_mult) { 3277674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32785b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3279674ae819SStefano Zampini } 3280674ae819SStefano Zampini } else { 3281674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32825b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3283674ae819SStefano Zampini } 3284674ae819SStefano Zampini } 32852b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3286674ae819SStefano Zampini /* free workspace */ 3287674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3288674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3289674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 329017ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3291674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3292674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3293674ae819SStefano Zampini PetscFunctionReturn(0); 3294674ae819SStefano Zampini } 32959a7d3425SStefano Zampini 32969a7d3425SStefano Zampini #undef __FUNCT__ 32979a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32989a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32999a7d3425SStefano Zampini { 33009a7d3425SStefano Zampini PetscInt i,j; 33019a7d3425SStefano Zampini PetscScalar *alphas; 33029a7d3425SStefano Zampini PetscErrorCode ierr; 33039a7d3425SStefano Zampini 33049a7d3425SStefano Zampini PetscFunctionBegin; 33059a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3306785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33079a7d3425SStefano Zampini for (i=0;i<n;i++) { 33089a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33099a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33109a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33119a7d3425SStefano Zampini } 33129a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33139a7d3425SStefano Zampini PetscFunctionReturn(0); 33149a7d3425SStefano Zampini } 33159a7d3425SStefano Zampini 3316e7931f94SStefano Zampini #undef __FUNCT__ 331770cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 331828143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3319e7931f94SStefano Zampini { 3320e7931f94SStefano Zampini Mat subdomain_adj; 3321e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3322e7931f94SStefano Zampini MatPartitioning partitioner; 3323e7931f94SStefano Zampini Mat_IS *matis; 3324e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3325e7931f94SStefano Zampini PetscInt prank; 3326e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3327e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3328e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33293837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3330e7931f94SStefano Zampini PetscErrorCode ierr; 33312b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3332e7931f94SStefano Zampini PetscSubcomm subcomm; 3333a57a6d2fSStefano Zampini 3334e7931f94SStefano Zampini PetscFunctionBegin; 33352b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33362b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33372b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3338e7931f94SStefano Zampini 3339e7931f94SStefano Zampini /* Get info on mapping */ 3340e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3341e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3342e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3343e7931f94SStefano Zampini 3344e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3345785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3346e7931f94SStefano Zampini xadj[0] = 0; 3347e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3348785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3349785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3350e7931f94SStefano Zampini 33512b510759SStefano Zampini if (threshold) { 3352d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33532b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3354d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3355d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3356d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3357d023bfaeSStefano Zampini xadj_count++; 3358e7931f94SStefano Zampini } 3359e7931f94SStefano Zampini } 3360d023bfaeSStefano Zampini xadj[1] = xadj_count; 3361c8587f34SStefano Zampini } else { 3362e7931f94SStefano Zampini if (xadj[1]) { 3363e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3364e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3365c8587f34SStefano Zampini } 3366e7931f94SStefano Zampini } 3367e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3368e7931f94SStefano Zampini if (use_square) { 3369e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3370e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3371e7931f94SStefano Zampini } 3372e7931f94SStefano Zampini } 3373e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3374e7931f94SStefano Zampini 33753837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3376e7931f94SStefano Zampini 3377e7931f94SStefano Zampini /* 3378e7931f94SStefano Zampini Restrict work on active processes only. 3379e7931f94SStefano Zampini */ 3380e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3381e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3382e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33832b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3384d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3385e7931f94SStefano Zampini if (color) { 3386e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3387e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3388e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3389c8587f34SStefano Zampini } else { 339028143c3dSStefano Zampini PetscInt coarsening_ratio; 3391306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3392785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3393e7931f94SStefano Zampini prank = rank; 3394306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33958002ef2cSStefano Zampini /* 3396e7931f94SStefano Zampini for (i=0;i<size;i++) { 3397e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3398c8587f34SStefano Zampini } 33998002ef2cSStefano Zampini */ 3400e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3401e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3402c8587f34SStefano Zampini } 3403e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3404306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 340522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3406e7931f94SStefano Zampini 3407e7931f94SStefano Zampini /* Partition */ 3408306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3409e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3410e7931f94SStefano Zampini if (use_vwgt) { 34113837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3412e7931f94SStefano Zampini v_wgt[0] = local_size; 3413e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3414c8587f34SStefano Zampini } 341528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 341628143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 341728143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3418e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3419e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 342022b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3421e7931f94SStefano Zampini 3422e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 342328143c3dSStefano Zampini if (contiguous) { 342428143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 342528143c3dSStefano Zampini } else { 342684ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 342728143c3dSStefano Zampini } 3428e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3429e7931f94SStefano Zampini /* clean up */ 3430e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3431e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3432e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3433e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3434e7931f94SStefano Zampini } 3435e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3436e7931f94SStefano Zampini 3437e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3438e7931f94SStefano Zampini i = 1; 3439e7931f94SStefano Zampini if (color) i=0; 3440e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3441e7931f94SStefano Zampini 3442e7931f94SStefano Zampini /* get back IS */ 3443e7931f94SStefano Zampini *is_sends = ranks_send_to; 3444e7931f94SStefano Zampini PetscFunctionReturn(0); 3445e7931f94SStefano Zampini } 3446e7931f94SStefano Zampini 3447e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3448e7931f94SStefano Zampini 3449e7931f94SStefano Zampini #undef __FUNCT__ 3450e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 345128143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3452e7931f94SStefano Zampini { 345370cf5478SStefano Zampini Mat local_mat; 3454e7931f94SStefano Zampini Mat_IS *matis; 3455e7931f94SStefano Zampini IS is_sends_internal; 34569d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 345728143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34589d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3459e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3460e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3461e7931f94SStefano Zampini const PetscInt* is_indices; 3462e7931f94SStefano Zampini MatType new_local_type; 3463e7931f94SStefano Zampini /* buffers */ 3464e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 346528143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34669d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3467e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3468e7931f94SStefano Zampini /* MPI */ 346928143c3dSStefano Zampini MPI_Comm comm,comm_n; 347028143c3dSStefano Zampini PetscSubcomm subcomm; 3471e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 347228143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 347328143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 347428143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 347528143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 347628143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3477e7931f94SStefano Zampini PetscErrorCode ierr; 3478e7931f94SStefano Zampini 3479e7931f94SStefano Zampini PetscFunctionBegin; 348028143c3dSStefano Zampini /* TODO: add missing checks */ 348128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 348228143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 348328143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 348428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3485e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 348628143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3487e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3488e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3489e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3490e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3491e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 349228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 349370cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 349470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 349528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 349670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 349770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 349870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 349970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 350070cf5478SStefano Zampini } 3501e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3502e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3503e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3504e7931f94SStefano Zampini if (!is_sends) { 350528143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 350628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 350728143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3508c8587f34SStefano Zampini } else { 3509e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3510e7931f94SStefano Zampini is_sends_internal = is_sends; 3511c8587f34SStefano Zampini } 3512e7931f94SStefano Zampini 3513e7931f94SStefano Zampini /* get pointer of MATIS data */ 3514e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3515e7931f94SStefano Zampini 3516e7931f94SStefano Zampini /* get comm */ 3517a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3518e7931f94SStefano Zampini 3519e7931f94SStefano Zampini /* compute number of sends */ 3520e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3521e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3522e7931f94SStefano Zampini 3523e7931f94SStefano Zampini /* compute number of receives */ 3524e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3525785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3526e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3527e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3528e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3529e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3530e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3531e7931f94SStefano Zampini 353228143c3dSStefano Zampini /* restrict comm if requested */ 353328143c3dSStefano Zampini subcomm = 0; 353428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 353528143c3dSStefano Zampini if (restrict_comm) { 3536779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3537779c1cceSStefano Zampini 353828143c3dSStefano Zampini color = 0; 35396366ad86SStefano Zampini if (!n_recvs && n_sends) color = 1; /* processes sending and not receiving anything will not partecipate in new comm */ 354028143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 354128143c3dSStefano Zampini subcommsize = commsize - subcommsize; 354228143c3dSStefano Zampini /* check if reuse has been requested */ 354328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 354428143c3dSStefano Zampini if (*mat_n) { 354528143c3dSStefano Zampini PetscMPIInt subcommsize2; 354628143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 354728143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 354828143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 354928143c3dSStefano Zampini } else { 355028143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 355128143c3dSStefano Zampini } 355228143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3553779c1cceSStefano Zampini PetscMPIInt rank; 3554779c1cceSStefano Zampini 3555779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 355628143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 355728143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 355828143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3559306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 356028143c3dSStefano Zampini } 356128143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 356228143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 356328143c3dSStefano Zampini } else { 356428143c3dSStefano Zampini comm_n = comm; 356528143c3dSStefano Zampini } 356628143c3dSStefano Zampini 3567e7931f94SStefano Zampini /* prepare send/receive buffers */ 3568785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3569e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3570785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3571e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 357228143c3dSStefano Zampini if (nis) { 3573854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 357428143c3dSStefano Zampini } 3575e7931f94SStefano Zampini 357628143c3dSStefano Zampini /* Get data from local matrices */ 3577e7931f94SStefano Zampini if (!isdense) { 3578a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3579e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3580e7931f94SStefano Zampini /* 3581e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3582e7931f94SStefano Zampini send_buffer_idxs should contain: 3583e7931f94SStefano Zampini - MatType_PRIVATE type 3584e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3585e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3586e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3587e7931f94SStefano Zampini */ 3588e7931f94SStefano Zampini } else { 3589e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3590e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3591854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3592e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3593e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3594e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3595e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3596e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3597e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3598e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3599e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3600e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3601c8587f34SStefano Zampini } 3602c8587f34SStefano Zampini } 3603e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 360428143c3dSStefano Zampini /* additional is (if any) */ 360528143c3dSStefano Zampini if (nis) { 360628143c3dSStefano Zampini PetscMPIInt psum; 360728143c3dSStefano Zampini PetscInt j; 360828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 360928143c3dSStefano Zampini PetscInt plen; 361028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 361128143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 361228143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 361328143c3dSStefano Zampini } 3614854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 361528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 361628143c3dSStefano Zampini PetscInt plen; 361728143c3dSStefano Zampini const PetscInt *is_array_idxs; 361828143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 361928143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 362028143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362128143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 362228143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 362328143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 362428143c3dSStefano Zampini } 362528143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 362628143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 362728143c3dSStefano Zampini } 362828143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 362928143c3dSStefano Zampini } 363028143c3dSStefano Zampini 3631e7931f94SStefano Zampini buf_size_idxs = 0; 3632e7931f94SStefano Zampini buf_size_vals = 0; 363328143c3dSStefano Zampini buf_size_idxs_is = 0; 3634e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3635e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3636e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 363728143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3638e7931f94SStefano Zampini } 3639785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3640785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 364195ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3642e7931f94SStefano Zampini 3643e7931f94SStefano Zampini /* get new tags for clean communications */ 3644e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3645e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 364628143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3647e7931f94SStefano Zampini 3648e7931f94SStefano Zampini /* allocate for requests */ 3649785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3650785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 365195ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3652785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3653785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 365495ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3655e7931f94SStefano Zampini 3656e7931f94SStefano Zampini /* communications */ 3657e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3658e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 365928143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3660e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3661e7931f94SStefano Zampini source_dest = onodes[i]; 3662e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3663e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3664e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3665e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 366628143c3dSStefano Zampini if (nis) { 366728143c3dSStefano 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); 366828143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 366928143c3dSStefano Zampini } 3670e7931f94SStefano Zampini } 3671e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3672e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3673e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3674e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 367528143c3dSStefano Zampini if (nis) { 367628143c3dSStefano 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); 367728143c3dSStefano Zampini } 3678e7931f94SStefano Zampini } 3679e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3680e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3681e7931f94SStefano Zampini 3682e7931f94SStefano Zampini /* assemble new l2g map */ 3683e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3684e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36859d30be91SStefano Zampini new_local_rows = 0; 3686e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36879d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3688e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3689e7931f94SStefano Zampini } 36909d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3691e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36929d30be91SStefano Zampini new_local_rows = 0; 3693e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36949d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 36959d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3696e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3697e7931f94SStefano Zampini } 36989d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 36999d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3700e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3701e7931f94SStefano Zampini 3702e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3703e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3704e7931f94SStefano 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) */ 3705e7931f94SStefano Zampini if (n_recvs) { 370628143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3707e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3708e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3709e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3710e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3711e7931f94SStefano Zampini break; 3712e7931f94SStefano Zampini } 3713e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3714e7931f94SStefano Zampini } 3715e7931f94SStefano Zampini switch (new_local_type_private) { 371628143c3dSStefano Zampini case MATDENSE_PRIVATE: 371728143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3718e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3719e7931f94SStefano Zampini bs = 1; 372028143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 372128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 372228143c3dSStefano Zampini bs = 1; 372328143c3dSStefano Zampini } 3724e7931f94SStefano Zampini break; 3725e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3726e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3727e7931f94SStefano Zampini bs = 1; 3728e7931f94SStefano Zampini break; 3729e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3730e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3731e7931f94SStefano Zampini break; 3732e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3733e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3734e7931f94SStefano Zampini break; 3735e7931f94SStefano Zampini default: 37369d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3737e7931f94SStefano Zampini break; 3738e7931f94SStefano Zampini } 373928143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 374028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 374128143c3dSStefano Zampini bs = 1; 3742e7931f94SStefano Zampini } 3743e7931f94SStefano Zampini 374470cf5478SStefano Zampini /* create MATIS object if needed */ 374570cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3746e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 374728143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 374870cf5478SStefano Zampini } else { 374970cf5478SStefano Zampini /* it also destroys the local matrices */ 375070cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 375170cf5478SStefano Zampini } 375270cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3753e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37549d30be91SStefano Zampini 37559d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37569d30be91SStefano Zampini 37579d30be91SStefano Zampini /* Global to local map of received indices */ 37589d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37599d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37609d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37619d30be91SStefano Zampini 37629d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37639d30be91SStefano Zampini buf_size_idxs = 0; 37649d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37659d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37669d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37679d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37689d30be91SStefano Zampini } 37699d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37709d30be91SStefano Zampini 37719d30be91SStefano Zampini /* set preallocation */ 37729d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37739d30be91SStefano Zampini if (!newisdense) { 37749d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37759d30be91SStefano Zampini 37769d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37779d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37789d30be91SStefano Zampini if (n_recvs) { 37799d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37809d30be91SStefano Zampini } 37819d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37829d30be91SStefano Zampini PetscInt j; 37839d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37849d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37859d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37869d30be91SStefano Zampini } 37879d30be91SStefano Zampini } else { 37889d30be91SStefano Zampini /* TODO */ 37899d30be91SStefano Zampini } 37909d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37919d30be91SStefano Zampini } 37929d30be91SStefano Zampini if (new_local_nnz) { 37939d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 37949d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 37959d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 37969d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37979d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 37989d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37999d30be91SStefano Zampini } else { 38009d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38019d30be91SStefano Zampini } 38029d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38039d30be91SStefano Zampini } else { 38049d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38059d30be91SStefano Zampini } 3806e7931f94SStefano Zampini 3807e7931f94SStefano Zampini /* set values */ 3808e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38099d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3810e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3811e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3812e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38139d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3814e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3815e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3816e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 381728143c3dSStefano Zampini } else { 381828143c3dSStefano Zampini /* TODO */ 3819e7931f94SStefano Zampini } 3820e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3821e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3822e7931f94SStefano Zampini } 3823e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3824e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382570cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 382670cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38279d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38289d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3829e7931f94SStefano Zampini 3830dfd14d43SStefano Zampini #if 0 383128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3832e7931f94SStefano Zampini Vec lvec,rvec; 3833e7931f94SStefano Zampini PetscReal infty_error; 3834e7931f94SStefano Zampini 38352a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3836e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3837e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3838e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 383970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3840e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3841e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3842e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3843e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3844e7931f94SStefano Zampini } 384528143c3dSStefano Zampini #endif 3846e7931f94SStefano Zampini 384728143c3dSStefano Zampini /* assemble new additional is (if any) */ 384828143c3dSStefano Zampini if (nis) { 384928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 385028143c3dSStefano Zampini 385128143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3852854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 385328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 385428143c3dSStefano Zampini psum = 0; 385528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 385628143c3dSStefano Zampini for (j=0;j<nis;j++) { 385728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 385828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 385928143c3dSStefano Zampini psum += plen; 386028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 386128143c3dSStefano Zampini } 386228143c3dSStefano Zampini } 3863854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3864854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 386528143c3dSStefano Zampini for (i=1;i<nis;i++) { 386628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 386728143c3dSStefano Zampini } 386828143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 386928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 387028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 387128143c3dSStefano Zampini for (j=0;j<nis;j++) { 387228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 387328143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 387428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 387528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 387628143c3dSStefano Zampini } 387728143c3dSStefano Zampini } 387828143c3dSStefano Zampini for (i=0;i<nis;i++) { 387928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 388028143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 388128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 388228143c3dSStefano Zampini } 388328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 388428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 388528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 388628143c3dSStefano Zampini } 3887e7931f94SStefano Zampini /* free workspace */ 388828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3889e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3890e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3891e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3892e7931f94SStefano Zampini if (isdense) { 3893e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3894e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3895e7931f94SStefano Zampini } else { 3896e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3897e7931f94SStefano Zampini } 389828143c3dSStefano Zampini if (nis) { 389928143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 390028143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 390128143c3dSStefano Zampini } 3902e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3903e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 390428143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3906e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 390728143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3908e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3909e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3910e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3911e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3912e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 391328143c3dSStefano Zampini if (nis) { 391428143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 391528143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 391628143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 391728143c3dSStefano Zampini } 391828143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 391928143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 392028143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 392128143c3dSStefano Zampini for (i=0;i<nis;i++) { 392228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 392328143c3dSStefano Zampini } 392428143c3dSStefano Zampini } 3925e7931f94SStefano Zampini PetscFunctionReturn(0); 3926e7931f94SStefano Zampini } 3927a57a6d2fSStefano Zampini 392812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3929af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 393012edc857SStefano Zampini 3931c8587f34SStefano Zampini #undef __FUNCT__ 3932c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3933c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3934c8587f34SStefano Zampini { 3935c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3936c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 393720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39389881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 393920a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39406e683305SStefano Zampini IS coarse_is,*isarray; 39416e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39426e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3943f9eb5b7dSStefano Zampini PC pc_temp; 3944c8587f34SStefano Zampini PCType coarse_pc_type; 3945c8587f34SStefano Zampini KSPType coarse_ksp_type; 3946f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39474f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39486e683305SStefano Zampini Mat t_coarse_mat_is; 39496e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39506e683305SStefano Zampini PetscMPIInt all_procs; 395174e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 395268457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 395322bc73bbSStefano Zampini PetscScalar *array; 39549881197aSStefano Zampini PetscErrorCode ierr; 3955fdc09c96SStefano Zampini 3956c8587f34SStefano Zampini PetscFunctionBegin; 3957c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 395868457ee5SStefano 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 */ 3959fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39605a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3961fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3962f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3963f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3964f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3965fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3966fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3967727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3968fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3969fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3970fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3971f4ddd8eeSStefano Zampini } 3972fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3973fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3974f4ddd8eeSStefano Zampini } 397570cf5478SStefano Zampini /* reset any subassembling information */ 397670cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39776e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39786e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3979fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3980f4ddd8eeSStefano Zampini } 3981c8587f34SStefano Zampini 39826e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39832b510759SStefano Zampini im_active = !!(pcis->n); 39842b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39856e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39866e683305SStefano Zampini void_procs = all_procs-active_procs; 39876e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 398874e2c79eSStefano Zampini redist = PETSC_FALSE; 398922bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39906e683305SStefano Zampini csin_ml = PETSC_TRUE; 39916e683305SStefano Zampini ncoarse_ml = void_procs; 3992779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3993779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 399418a45a71SStefano Zampini csin_ds = PETSC_TRUE; 399518a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 399618a45a71SStefano Zampini redist = PETSC_TRUE; 399718a45a71SStefano Zampini } else { 39986e683305SStefano Zampini csin_ds = PETSC_TRUE; 3999779c1cceSStefano Zampini ncoarse_ds = active_procs; 4000779c1cceSStefano Zampini redist = PETSC_TRUE; 400118a45a71SStefano Zampini } 40026e683305SStefano Zampini } else { 40036e683305SStefano Zampini csin_ml = PETSC_FALSE; 40046e683305SStefano Zampini ncoarse_ml = all_procs; 40056e683305SStefano Zampini if (void_procs) { 40066e683305SStefano Zampini csin_ds = PETSC_TRUE; 40076e683305SStefano Zampini ncoarse_ds = void_procs; 40086e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40096e683305SStefano Zampini } else { 4010779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 401174e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 401274e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 401374e2c79eSStefano Zampini redist = PETSC_TRUE; 401474e2c79eSStefano Zampini } else { 40156e683305SStefano Zampini csin_ds = PETSC_FALSE; 40166e683305SStefano Zampini ncoarse_ds = all_procs; 40176e683305SStefano Zampini } 40186e683305SStefano Zampini } 401974e2c79eSStefano Zampini } 40206e683305SStefano Zampini 40216e683305SStefano Zampini /* 40226e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40236e683305SStefano Zampini - we have not exceeded the number of levels requested 40246e683305SStefano Zampini - we can actually subassemble the active processes 40256e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40266e683305SStefano Zampini */ 40276e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40286e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40296e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40306e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40316e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4032f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40332b510759SStefano Zampini } else { 4034f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4035c8587f34SStefano Zampini } 4036c8587f34SStefano Zampini } 40376e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40386e683305SStefano Zampini if (multilevel_allowed) { 40396e683305SStefano Zampini ncoarse = ncoarse_ml; 40406e683305SStefano Zampini csin = csin_ml; 404158da7f69SStefano Zampini redist = PETSC_FALSE; 40426e683305SStefano Zampini } else { 40436e683305SStefano Zampini ncoarse = ncoarse_ds; 40446e683305SStefano Zampini csin = csin_ds; 40456e683305SStefano Zampini } 4046e7931f94SStefano Zampini 4047abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4048abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4049abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4050abbbba34SStefano Zampini 4051abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 405222bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 405322bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 405422bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 405522bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4056b9b85e73SStefano Zampini #if 0 4057b9b85e73SStefano Zampini { 4058b9b85e73SStefano Zampini PetscViewer viewer; 4059b9b85e73SStefano Zampini char filename[256]; 4060b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4061b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4062b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4063b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4064b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4065b9b85e73SStefano Zampini } 4066b9b85e73SStefano Zampini #endif 40676e683305SStefano 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); 40686e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40696e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40706e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4071abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4072abbbba34SStefano Zampini 40736e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40746e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40756e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40766e683305SStefano Zampini const PetscInt *idxs; 40776e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40786e683305SStefano Zampini 40796e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40800be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40816e683305SStefano Zampini /* allocate space for temporary storage */ 4082854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4083854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40846e683305SStefano Zampini /* allocate for IS array */ 40856e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40866e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40876e683305SStefano Zampini nis = nisdofs + nisneu; 4088854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40896e683305SStefano Zampini /* dofs splitting */ 40906e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40916e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40926e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40936e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40946e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40956e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40966e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40976e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 40986e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 40996e683305SStefano Zampini } 41006e683305SStefano Zampini /* neumann boundaries */ 41016e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41026e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41036e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41046e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41056e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41066e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41076e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41086e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41096e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41106e683305SStefano Zampini } 41116e683305SStefano Zampini /* free memory */ 41126e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41136e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41146e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41156e683305SStefano Zampini } else { 41166e683305SStefano Zampini nis = 0; 41176e683305SStefano Zampini nisdofs = 0; 41186e683305SStefano Zampini nisneu = 0; 41196e683305SStefano Zampini isarray = NULL; 41206e683305SStefano Zampini } 41216e683305SStefano Zampini /* destroy no longer needed map */ 41226e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41236e683305SStefano Zampini 41246e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41256e683305SStefano Zampini coarse_mat_is = NULL; 41266e683305SStefano Zampini if (csin) { 41276e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 412874e2c79eSStefano Zampini if (redist) { 412974e2c79eSStefano Zampini PetscMPIInt rank; 4130779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 413174e2c79eSStefano Zampini 413274e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 413358da7f69SStefano Zampini spc = active_procs/ncoarse; 413458da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4135779c1cceSStefano Zampini if (im_active) { 4136779c1cceSStefano Zampini destsize = 1; 413774e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 413874e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 413974e2c79eSStefano Zampini } else { 414074e2c79eSStefano Zampini dest[0] = rank/(spc+1); 414174e2c79eSStefano Zampini } 414274e2c79eSStefano Zampini } else { 4143779c1cceSStefano Zampini destsize = 0; 41446e683305SStefano Zampini } 4145779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4146779c1cceSStefano Zampini } else if (csin_type_simple) { 41476e683305SStefano Zampini PetscMPIInt rank; 41486e683305SStefano Zampini PetscInt issize,isidx; 4149779c1cceSStefano Zampini 41506e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41516e683305SStefano Zampini if (im_active) { 41526e683305SStefano Zampini issize = 1; 41536e683305SStefano Zampini isidx = (PetscInt)rank; 41546e683305SStefano Zampini } else { 41556e683305SStefano Zampini issize = 0; 41566e683305SStefano Zampini isidx = -1; 41576e683305SStefano Zampini } 41586e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4159779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41606e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41616e683305SStefano Zampini } 4162779c1cceSStefano Zampini 4163779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4164779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4165779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4166779c1cceSStefano Zampini PetscInt *coarse_candidates; 4167779c1cceSStefano Zampini const PetscInt* tisindices; 4168779c1cceSStefano Zampini 4169779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4170779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4171779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4172779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4173779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4174779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4175779c1cceSStefano Zampini } 4176779c1cceSStefano Zampini } 4177779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4178779c1cceSStefano Zampini 4179779c1cceSStefano Zampini 41806e683305SStefano Zampini if (pcbddc->dbg_flag) { 41816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41836e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41846e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4185779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41876e683305SStefano Zampini } 41886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41896e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41906e683305SStefano Zampini } 41916e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41926e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41936e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4194854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 41956e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 41966e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 41976e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 41986e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 41996e683305SStefano Zampini } 42006e683305SStefano Zampini if (pcbddc->dbg_flag) { 42016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42036e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42046e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42056e683305SStefano Zampini } 4206779c1cceSStefano Zampini } 42076e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 42086e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 42096e683305SStefano Zampini } else { 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 not needed\n");CHKERRQ(ierr); 42136e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42146e683305SStefano Zampini } 42156e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42166e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42176e683305SStefano Zampini } 42186e683305SStefano Zampini 42196e683305SStefano Zampini /* create local to global scatters for coarse problem */ 422068457ee5SStefano Zampini if (compute_vecs) { 42216e683305SStefano Zampini PetscInt lrows; 42226e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42236e683305SStefano Zampini if (coarse_mat_is) { 42246e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42256e683305SStefano Zampini } else { 42266e683305SStefano Zampini lrows = 0; 42276e683305SStefano Zampini } 42286e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42296e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42306e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42316e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42326e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42336e683305SStefano Zampini } 42346e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42356e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4236c8587f34SStefano Zampini 4237f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4238f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4239f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4240f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4241f9eb5b7dSStefano Zampini } else { 4242f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4243f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4244c8587f34SStefano Zampini } 4245c8587f34SStefano Zampini 42466e683305SStefano Zampini /* print some info if requested */ 42476e683305SStefano Zampini if (pcbddc->dbg_flag) { 42486e683305SStefano Zampini if (!multilevel_allowed) { 42496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42506e683305SStefano Zampini if (multilevel_requested) { 42516e683305SStefano 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); 42526e683305SStefano Zampini } else if (pcbddc->max_levels) { 42536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42546e683305SStefano Zampini } 42556e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42566e683305SStefano Zampini } 42576e683305SStefano Zampini } 42586e683305SStefano Zampini 4259f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42606e683305SStefano Zampini if (coarse_mat_is) { 42616e683305SStefano Zampini MatReuse coarse_mat_reuse; 42626a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42636e683305SStefano Zampini if (pcbddc->dbg_flag) { 42646e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42656e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42666e683305SStefano Zampini } 4267f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4268312be037SStefano Zampini char prefix[256],str_level[16]; 4269e604994aSStefano Zampini size_t len; 42706e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4271c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4272f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42735f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4274c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42756e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4276c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4277c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42783e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4279e604994aSStefano Zampini /* prefix */ 4280e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4281e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4282e604994aSStefano Zampini if (!pcbddc->current_level) { 4283e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4284e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4285c8587f34SStefano Zampini } else { 4286e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4287312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4288312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 428934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4290312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4291e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4292e604994aSStefano Zampini } 4293e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 42943e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42953e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 42963e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 42973e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4298f9eb5b7dSStefano Zampini /* allow user customization */ 4299f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43003e3c6dadSStefano Zampini } 43013e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43023e3c6dadSStefano Zampini if (nisdofs) { 43033e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43043e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43053e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43063e3c6dadSStefano Zampini } 43073e3c6dadSStefano Zampini } 43083e3c6dadSStefano Zampini if (nisneu) { 43093e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43103e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4311312be037SStefano Zampini } 4312f9eb5b7dSStefano Zampini 4313f9eb5b7dSStefano Zampini /* get some info after set from options */ 4314f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4315f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4316f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43174f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43186e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4319f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4320f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4321f9eb5b7dSStefano Zampini } 43224f3a063dSStefano Zampini if (isredundant) { 43234f3a063dSStefano Zampini KSP inner_ksp; 43244f3a063dSStefano Zampini PC inner_pc; 43254f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43264f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43274f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43284f3a063dSStefano Zampini } 4329f9eb5b7dSStefano Zampini 4330f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4331fa7f1dd8SStefano Zampini if (coarse_reuse) { 433281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4333fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43346e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4335fa7f1dd8SStefano Zampini } else { 43366e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4337fa7f1dd8SStefano Zampini } 4338c8587f34SStefano Zampini if (isbddc || isnn) { 433922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 434070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 434128143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 434222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43456e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43466e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 434722b6e8a2SStefano Zampini } 434870cf5478SStefano Zampini } 43496e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 435070cf5478SStefano Zampini } else { 435122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 435222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 435322bc73bbSStefano Zampini } 435422bc73bbSStefano Zampini } else { 43552e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4356c8587f34SStefano Zampini } 4357c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4358c8587f34SStefano Zampini 43593301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 43605a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 43613301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 43623301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 43633301b35fSStefano Zampini } 43643301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 43653301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 43663301b35fSStefano Zampini } 43673301b35fSStefano Zampini if (pc->pmat->spd_set) { 43683301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 43693301b35fSStefano Zampini } 43706e683305SStefano Zampini /* set operators */ 43715f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43726e683305SStefano Zampini if (pcbddc->dbg_flag) { 43736e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43746e683305SStefano Zampini } 43756e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43766e683305SStefano Zampini coarse_mat = 0; 43776e683305SStefano Zampini } 43786e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4379b9b85e73SStefano Zampini #if 0 4380b9b85e73SStefano Zampini { 4381b9b85e73SStefano Zampini PetscViewer viewer; 4382b9b85e73SStefano Zampini char filename[256]; 4383b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4384b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4385b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4386b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4387b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4388b9b85e73SStefano Zampini } 4389b9b85e73SStefano Zampini #endif 4390c8587f34SStefano Zampini 4391c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4392c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4393c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 439498a51de6SStefano Zampini } 439598a51de6SStefano Zampini 439698a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 439798a51de6SStefano Zampini Vec crhs,csol; 439898a51de6SStefano Zampini PetscBool ispreonly; 439998a51de6SStefano Zampini if (CoarseNullSpace) { 4400c8587f34SStefano Zampini if (isbddc) { 4401c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4402c8587f34SStefano Zampini } else { 4403c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4404c8587f34SStefano Zampini } 4405c8587f34SStefano Zampini } 4406f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4407f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4408f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4409f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44106e683305SStefano Zampini /* hack */ 4411f347579bSStefano Zampini if (!csol) { 44122a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4413f9eb5b7dSStefano Zampini } 4414f347579bSStefano Zampini if (!crhs) { 44152a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4416f347579bSStefano Zampini } 4417cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4418cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44196e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4420c8587f34SStefano Zampini KSP check_ksp; 44212b510759SStefano Zampini KSPType check_ksp_type; 4422c8587f34SStefano Zampini PC check_pc; 44236e683305SStefano Zampini Vec check_vec,coarse_vec; 44246a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44252b510759SStefano Zampini PetscInt its; 44266e683305SStefano Zampini PetscBool compute_eigs; 44276e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44286e683305SStefano Zampini PetscInt neigs; 44298e185a42SStefano Zampini const char *prefix; 4430c8587f34SStefano Zampini 44312b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44326e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 443323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4434f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44352b510759SStefano Zampini if (ispreonly) { 44362b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44376e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44382b510759SStefano Zampini } else { 4439cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44406e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4441c8587f34SStefano Zampini } 4442c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44436e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44446e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44456e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4446a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4447a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4448a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4449a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4450c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4451c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4452c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4453c8587f34SStefano Zampini /* create random vec */ 44546e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44556e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4456c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4457c8587f34SStefano Zampini if (CoarseNullSpace) { 4458c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4459c8587f34SStefano Zampini } 44606e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4461c8587f34SStefano Zampini /* solve coarse problem */ 44626e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4463c8587f34SStefano Zampini if (CoarseNullSpace) { 44646e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4465c8587f34SStefano Zampini } 4466cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44676e683305SStefano Zampini if (compute_eigs) { 4468854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4469854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44706e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44716e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44726e683305SStefano Zampini lambda_min = eigs_r[0]; 44736e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44746e683305SStefano Zampini if (lambda_max>lambda_min) { 4475cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4476cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4477cbcc2c2aSStefano Zampini } 4478c8587f34SStefano Zampini } 4479c8587f34SStefano Zampini } 4480cbcc2c2aSStefano Zampini 4481c8587f34SStefano Zampini /* check coarse problem residual error */ 44826e683305SStefano Zampini if (pcbddc->dbg_flag) { 44836e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44846e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44856e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4486c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44876e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44886e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4489c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4490779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44916e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44926e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44936e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 44946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 44956e683305SStefano Zampini if (compute_eigs) { 44966e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4497deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4498c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 44996e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45006e683305SStefano 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); 45016e683305SStefano Zampini for (i=0;i<neigs;i++) { 45026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4503c8587f34SStefano Zampini } 45046e683305SStefano Zampini } 45056e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45066e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45076e683305SStefano Zampini } 4508c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45096e683305SStefano Zampini if (compute_eigs) { 45106e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45116e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4512c8587f34SStefano Zampini } 45136e683305SStefano Zampini } 45146e683305SStefano Zampini } 4515cbcc2c2aSStefano Zampini /* print additional info */ 4516cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45176e683305SStefano Zampini /* waits until all processes reaches this point */ 45186e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4519cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4520cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4521cbcc2c2aSStefano Zampini } 4522cbcc2c2aSStefano Zampini 45232b510759SStefano Zampini /* free memory */ 4524c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4525fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4526c8587f34SStefano Zampini PetscFunctionReturn(0); 4527c8587f34SStefano Zampini } 4528674ae819SStefano Zampini 4529f34684f1SStefano Zampini #undef __FUNCT__ 4530f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4531f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4532f34684f1SStefano Zampini { 4533f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4534f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4535f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 45360e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 453773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 45380e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4539f34684f1SStefano Zampini PetscErrorCode ierr; 4540f34684f1SStefano Zampini 4541f34684f1SStefano Zampini PetscFunctionBegin; 4542f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45430e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45440e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4545727cdba6SStefano Zampini } 45460e6343abSStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(pc->pmat)),matis->mapping,pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult,&coarse_size,&t_local_primal_indices);CHKERRQ(ierr); 45470e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 45480e6343abSStefano Zampini local_size = 0; 45490e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 45500e6343abSStefano Zampini PetscInt j; 45510e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 45520e6343abSStefano Zampini } 45530e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 45540e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45550e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45560e6343abSStefano Zampini } 4557f34684f1SStefano Zampini 4558f34684f1SStefano Zampini /* check numbering */ 4559f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4560f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4561b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4562f34684f1SStefano Zampini 4563f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4564f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4565f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45660fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4567f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4568f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4569727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4570f34684f1SStefano Zampini } 4571f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4572f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4573f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4574f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4575f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4576f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4577f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4578f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4579f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4580f34684f1SStefano Zampini if (array[i] == 1.0) { 4581b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4582f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4583f34684f1SStefano Zampini } 4584f34684f1SStefano Zampini } 4585b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4586f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4587f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4588f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4589f34684f1SStefano Zampini } 4590f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4591f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4592f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4593f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4594f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4595f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4596b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4597ca8b9ea9SStefano Zampini PetscInt *gidxs; 4598ca8b9ea9SStefano Zampini 4599ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4600ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4601f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4602f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4603f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4604f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46054bc2dc4bSStefano 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); 4606f34684f1SStefano Zampini } 4607f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4608ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4609f34684f1SStefano Zampini } 4610f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4611302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4612f34684f1SStefano Zampini } 46138bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4614f34684f1SStefano Zampini /* get back data */ 4615f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4616f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4617674ae819SStefano Zampini PetscFunctionReturn(0); 4618674ae819SStefano Zampini } 4619674ae819SStefano Zampini 4620e456f2a8SStefano Zampini #undef __FUNCT__ 4621e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4622a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4623e456f2a8SStefano Zampini { 4624e456f2a8SStefano Zampini IS localis_t; 4625a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4626e456f2a8SStefano Zampini PetscScalar *vals; 4627e456f2a8SStefano Zampini PetscErrorCode ierr; 4628e456f2a8SStefano Zampini 4629e456f2a8SStefano Zampini PetscFunctionBegin; 4630a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4631e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4632854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4633e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4634e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4635a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4636a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46371035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4638a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46391035eff8SStefano Zampini } 4640a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4641e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4642e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4643a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4644a7dc3881SStefano Zampini /* now compute set in local ordering */ 4645a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4646a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4647a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4648a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4649a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4650ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4651e456f2a8SStefano Zampini lsize++; 4652e456f2a8SStefano Zampini } 4653e456f2a8SStefano Zampini } 4654854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4655a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4656ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4657e456f2a8SStefano Zampini idxs[lsize++] = i; 4658e456f2a8SStefano Zampini } 4659e456f2a8SStefano Zampini } 4660a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4661a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4662e456f2a8SStefano Zampini *localis = localis_t; 4663e456f2a8SStefano Zampini PetscFunctionReturn(0); 4664e456f2a8SStefano Zampini } 4665906d46d4SStefano Zampini 4666906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4667906d46d4SStefano Zampini #undef __FUNCT__ 4668906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4669906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4670906d46d4SStefano Zampini { 4671906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4672906d46d4SStefano Zampini PetscErrorCode ierr; 4673906d46d4SStefano Zampini 4674906d46d4SStefano Zampini PetscFunctionBegin; 4675906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4676906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4677906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4678906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4679906d46d4SStefano Zampini PetscFunctionReturn(0); 4680906d46d4SStefano Zampini } 4681906d46d4SStefano Zampini 4682906d46d4SStefano Zampini #undef __FUNCT__ 4683906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4684906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4685906d46d4SStefano Zampini { 4686906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4687906d46d4SStefano Zampini PetscErrorCode ierr; 4688906d46d4SStefano Zampini 4689906d46d4SStefano Zampini PetscFunctionBegin; 4690906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4691906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4692906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4693906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4694906d46d4SStefano Zampini PetscFunctionReturn(0); 4695906d46d4SStefano Zampini } 4696b96c3477SStefano Zampini 4697b96c3477SStefano Zampini #undef __FUNCT__ 4698b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 469908122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4700b96c3477SStefano Zampini { 4701a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4702b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4703b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4704a64f4aa4SStefano Zampini Mat S_j; 4705b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4706b96c3477SStefano Zampini PetscBool free_used_adj; 4707b96c3477SStefano Zampini PetscErrorCode ierr; 4708b96c3477SStefano Zampini 4709b96c3477SStefano Zampini PetscFunctionBegin; 4710b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4711b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 471208122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4713b96c3477SStefano Zampini used_xadj = NULL; 4714b96c3477SStefano Zampini used_adjncy = NULL; 4715b96c3477SStefano Zampini } else { 471608122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 471708122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 471808122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 471908122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4720b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4721b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4722b96c3477SStefano Zampini } else { 47232fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4724b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4725b96c3477SStefano Zampini PetscInt nvtxs; 4726b96c3477SStefano Zampini 47272fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47282fffb893SStefano Zampini if (flg_row) { 4729b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4730b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4731b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4732b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47332fffb893SStefano Zampini } else { 47342fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47352fffb893SStefano Zampini used_xadj = NULL; 47362fffb893SStefano Zampini used_adjncy = NULL; 47372fffb893SStefano Zampini } 47382fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4739b96c3477SStefano Zampini } 4740b96c3477SStefano Zampini } 4741d5574798SStefano Zampini 4742d5574798SStefano Zampini /* setup sub_schurs data */ 4743a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4744a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4745a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4746a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4747*862806e4SStefano 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); 4748a64f4aa4SStefano Zampini } else { 47496816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 4750*862806e4SStefano 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); 4751a64f4aa4SStefano Zampini } 4752a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4753b96c3477SStefano Zampini 4754b96c3477SStefano Zampini /* free adjacency */ 4755b96c3477SStefano Zampini if (free_used_adj) { 4756b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4757b96c3477SStefano Zampini } 4758b96c3477SStefano Zampini PetscFunctionReturn(0); 4759b96c3477SStefano Zampini } 4760b96c3477SStefano Zampini 4761b96c3477SStefano Zampini #undef __FUNCT__ 4762b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 476308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4764b96c3477SStefano Zampini { 4765b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4766b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4767b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4768b96c3477SStefano Zampini PCBDDCGraph graph; 4769b96c3477SStefano Zampini PetscErrorCode ierr; 4770b96c3477SStefano Zampini 4771b96c3477SStefano Zampini PetscFunctionBegin; 4772b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 477308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 47743301b35fSStefano Zampini IS verticesIS,verticescomm; 47753301b35fSStefano Zampini PetscInt vsize,*idxs; 4776b96c3477SStefano Zampini 4777b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 47783301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 47793301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 47803301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 47813301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 47823301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4783b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 47847fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 47853301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 47863301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4787b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4788b96c3477SStefano Zampini /* 4789b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4790b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4791b96c3477SStefano Zampini } 4792b96c3477SStefano Zampini */ 4793b96c3477SStefano Zampini } else { 4794b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4795b96c3477SStefano Zampini } 4796b96c3477SStefano Zampini 4797b96c3477SStefano Zampini /* sub_schurs init */ 4798a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4799a64f4aa4SStefano Zampini 4800b96c3477SStefano Zampini /* free graph struct */ 480108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4802b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4803b96c3477SStefano Zampini } 4804b96c3477SStefano Zampini PetscFunctionReturn(0); 4805b96c3477SStefano Zampini } 4806