1ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3674ae819SStefano Zampini #include <petscblaslapack.h> 4674ae819SStefano Zampini 5906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini 8b1b3d7a2SStefano Zampini #undef __FUNCT__ 9b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 1008122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 11b1b3d7a2SStefano Zampini { 12b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1308122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1408122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 1508122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 1608122e43SStefano Zampini PetscScalar *work,lwork; 1708122e43SStefano Zampini PetscScalar *St,*S,*eigv; 1808122e43SStefano Zampini PetscScalar *Sarray,*Starray; 1908122e43SStefano Zampini PetscReal *eigs,thresh; 201b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 21f6f667cfSStefano Zampini PetscBool allocated_S_St; 2208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2308122e43SStefano Zampini PetscReal *rwork; 2408122e43SStefano Zampini #endif 25b1b3d7a2SStefano Zampini PetscErrorCode ierr; 26b1b3d7a2SStefano Zampini 27b1b3d7a2SStefano Zampini PetscFunctionBegin; 2808122e43SStefano Zampini if (!sub_schurs->use_mumps) { 2908122e43SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 3008122e43SStefano Zampini } 3108122e43SStefano Zampini 32fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 33fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 34fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 35fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 36fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 37fd14bc51SStefano Zampini } 38fd14bc51SStefano Zampini 39e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 40e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 41e496cd5dSStefano Zampini } 42e496cd5dSStefano Zampini 43e496cd5dSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 44e496cd5dSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 45e496cd5dSStefano Zampini } 46e496cd5dSStefano Zampini 4708122e43SStefano Zampini /* max size of subsets */ 4808122e43SStefano Zampini mss = 0; 4908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5008122e43SStefano Zampini PetscInt subset_size; 51862806e4SStefano Zampini 5208122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5308122e43SStefano Zampini mss = PetscMax(mss,subset_size); 5408122e43SStefano Zampini } 5508122e43SStefano Zampini 5608122e43SStefano Zampini /* min/max and threshold */ 5708122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 58f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 5908122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 60f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 61f6f667cfSStefano Zampini if (nmin) { 62f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 63f6f667cfSStefano Zampini } 6408122e43SStefano Zampini 6508122e43SStefano Zampini /* allocate lapack workspace */ 6608122e43SStefano Zampini cum = cum2 = 0; 6708122e43SStefano Zampini maxneigs = 0; 6808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6908122e43SStefano Zampini PetscInt n,subset_size; 70f6f667cfSStefano Zampini 7108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7208122e43SStefano Zampini n = PetscMin(subset_size,nmax); 739162d606SStefano Zampini cum += subset_size; 749162d606SStefano Zampini cum2 += subset_size*n; 7508122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7608122e43SStefano Zampini } 7708122e43SStefano Zampini if (mss) { 789ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 7908122e43SStefano Zampini PetscBLASInt B_itype = 1; 8008122e43SStefano Zampini PetscBLASInt B_N = mss; 814c6709b3SStefano Zampini PetscReal zero = 0.0; 824c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8308122e43SStefano Zampini 8408122e43SStefano Zampini B_lwork = -1; 8508122e43SStefano Zampini S = NULL; 8608122e43SStefano Zampini St = NULL; 87a58a30b4SStefano Zampini eigs = NULL; 88a58a30b4SStefano Zampini eigv = NULL; 89a58a30b4SStefano Zampini B_iwork = NULL; 90a58a30b4SStefano Zampini B_ifail = NULL; 918bec7fa6SStefano Zampini thresh = 1.0; 9208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9408122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 9508122e43SStefano Zampini #else 9608122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 9708122e43SStefano Zampini #endif 9808122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 9908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10008122e43SStefano Zampini } else { 10108122e43SStefano Zampini /* TODO */ 10208122e43SStefano Zampini } 10308122e43SStefano Zampini } else { 10408122e43SStefano Zampini lwork = 0; 10508122e43SStefano Zampini } 10608122e43SStefano Zampini 10708122e43SStefano Zampini nv = 0; 108d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 109d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 11008122e43SStefano Zampini } 1114c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 112f6f667cfSStefano Zampini if (allocated_S_St) { 113f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 114f6f667cfSStefano Zampini } 115f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11708122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 11808122e43SStefano Zampini #endif 1199162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1209162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1219162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12208122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1239162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12408122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12508122e43SStefano Zampini 12608122e43SStefano Zampini maxneigs = 0; 12708122e43SStefano Zampini cum = cum2 = cumarray = 0; 1289162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1299162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 130d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13108122e43SStefano Zampini const PetscInt *idxs; 13208122e43SStefano Zampini 133d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 13408122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 13508122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 13708122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1389162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1399162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 14008122e43SStefano Zampini } 14108122e43SStefano Zampini cum2 = cum; 142d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14308122e43SStefano Zampini } 14408122e43SStefano Zampini 14508122e43SStefano Zampini if (mss) { /* multilevel */ 14608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 14708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 14808122e43SStefano Zampini } 14908122e43SStefano Zampini 15008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15108122e43SStefano Zampini 15208122e43SStefano Zampini const PetscInt *idxs; 153f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 154862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 15508122e43SStefano Zampini PetscBLASInt B_N; 156aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 15708122e43SStefano Zampini 158862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 159f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 160f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1619ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 162aff50787SStefano Zampini PetscInt j,k; 163aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 164aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 165aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 16608122e43SStefano Zampini } 16708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 168aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 169aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 170aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 171aff50787SStefano Zampini } 17208122e43SStefano Zampini } 17308122e43SStefano Zampini } else { 17408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 17608122e43SStefano Zampini } 1778bec7fa6SStefano Zampini } else { 178f6f667cfSStefano Zampini S = Sarray + cumarray; 179f6f667cfSStefano Zampini St = Starray + cumarray; 1808bec7fa6SStefano Zampini } 18108122e43SStefano Zampini 182f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 183aff50787SStefano Zampini /* see if we can save some work */ 184aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 185aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 186aff50787SStefano Zampini } 187aff50787SStefano Zampini 188aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 189aff50787SStefano Zampini B_neigs = 0; 190aff50787SStefano Zampini } else { 191aff50787SStefano Zampini /* Threshold: this is an heuristic for edges */ 192f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 193f6f667cfSStefano Zampini 1949ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 19508122e43SStefano Zampini PetscBLASInt B_itype = 1; 196f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1974c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1989552c7c7SStefano Zampini PetscInt nmin_s; 19908122e43SStefano Zampini 200f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 201fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 2028bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 203fd14bc51SStefano Zampini } 20408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 20508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 206f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 20708122e43SStefano Zampini #else 208f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 20908122e43SStefano Zampini #endif 21008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 21108122e43SStefano Zampini if (B_ierr) { 21208122e43SStefano Zampini if (B_ierr < 0 ) { 21308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 21408122e43SStefano Zampini } else if (B_ierr <= B_N) { 21508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 21608122e43SStefano Zampini } else { 2179552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 21808122e43SStefano Zampini } 21908122e43SStefano Zampini } 22008122e43SStefano Zampini 22108122e43SStefano Zampini if (B_neigs > nmax) { 222fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 223fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 224fd14bc51SStefano Zampini } 225f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 22608122e43SStefano Zampini B_neigs = nmax; 22708122e43SStefano Zampini } 22808122e43SStefano Zampini 2299552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2309552c7c7SStefano Zampini if (B_neigs < nmin_s) { 23108122e43SStefano Zampini PetscBLASInt B_neigs2; 23208122e43SStefano Zampini 233f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 234f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 235fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 236fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 237fd14bc51SStefano Zampini } 2389ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23908122e43SStefano Zampini PetscInt j; 24008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24108122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24208122e43SStefano Zampini } 24308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 24408122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 24508122e43SStefano Zampini } 24608122e43SStefano Zampini } else { 24708122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24808122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24908122e43SStefano Zampini } 25008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 25108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 252f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 25308122e43SStefano Zampini #else 254f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 25508122e43SStefano Zampini #endif 25608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 25708122e43SStefano Zampini B_neigs += B_neigs2; 25808122e43SStefano Zampini } 25908122e43SStefano Zampini if (B_ierr) { 26008122e43SStefano Zampini if (B_ierr < 0 ) { 26108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 26208122e43SStefano Zampini } else if (B_ierr <= B_N) { 26308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 26408122e43SStefano Zampini } else { 2659552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 26608122e43SStefano Zampini } 26708122e43SStefano Zampini } 268fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 269fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 27008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 27108122e43SStefano Zampini if (eigs[j] == 0.0) { 272fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 27308122e43SStefano Zampini } else { 274f6f667cfSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]); 275fd14bc51SStefano Zampini } 27608122e43SStefano Zampini } 27708122e43SStefano Zampini } 27808122e43SStefano Zampini } else { 27908122e43SStefano Zampini /* TODO */ 28008122e43SStefano Zampini } 281aff50787SStefano Zampini } 2828bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2838bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2849162d606SStefano Zampini if (B_neigs) { 2859162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 286fd14bc51SStefano Zampini 287fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2889552c7c7SStefano Zampini PetscInt ii; 2899552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 290fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2919552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2929162d606SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2939552c7c7SStefano Zampini } 2949552c7c7SStefano Zampini } 295fd14bc51SStefano Zampini } 29608122e43SStefano Zampini #if 0 2979162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29808122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29908122e43SStefano Zampini PetscScalar norm; 30008122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3019162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3029162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 30308122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 30408122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 30508122e43SStefano Zampini } else { 30608122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 30708122e43SStefano Zampini } 3089162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 309b1b3d7a2SStefano Zampini } 310b1b3d7a2SStefano Zampini #endif 3119162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3129162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3139162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3149162d606SStefano Zampini cum++; 31508122e43SStefano Zampini } 31608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 31708122e43SStefano Zampini /* shift for next computation */ 31808122e43SStefano Zampini cumarray += subset_size*subset_size; 31908122e43SStefano Zampini } 320fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 321fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 322fd14bc51SStefano Zampini } 32308122e43SStefano Zampini 32408122e43SStefano Zampini if (mss) { 32508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 32608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 327f6f667cfSStefano Zampini /* destroy matrices (junk) */ 328f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 329f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 33008122e43SStefano Zampini } 331f6f667cfSStefano Zampini if (allocated_S_St) { 332f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 333f6f667cfSStefano Zampini } 334f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 33508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 33608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33708122e43SStefano Zampini #endif 33808122e43SStefano Zampini if (pcbddc->dbg_flag) { 3391b968477SStefano Zampini PetscInt maxneigs_r; 34008122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3419b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 34208122e43SStefano Zampini } 34308122e43SStefano Zampini PetscFunctionReturn(0); 34408122e43SStefano Zampini } 345b1b3d7a2SStefano Zampini 346674ae819SStefano Zampini #undef __FUNCT__ 347c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 348c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 349c8587f34SStefano Zampini { 350c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3518629588bSStefano Zampini PetscScalar *coarse_submat_vals; 352c8587f34SStefano Zampini PetscErrorCode ierr; 353c8587f34SStefano Zampini 354c8587f34SStefano Zampini PetscFunctionBegin; 355f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3565e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 357c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 358c8587f34SStefano Zampini 359684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3600fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 361684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 362c8587f34SStefano Zampini 363c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 364b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 365c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 366c8587f34SStefano Zampini } 367c8587f34SStefano Zampini 3688629588bSStefano Zampini /* 3698629588bSStefano Zampini Setup local correction and local part of coarse basis. 3708629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3718629588bSStefano Zampini */ 37247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3738629588bSStefano Zampini 3748629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3758629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3768629588bSStefano Zampini 3778629588bSStefano Zampini /* free */ 3788629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 379c8587f34SStefano Zampini PetscFunctionReturn(0); 380c8587f34SStefano Zampini } 381c8587f34SStefano Zampini 382c8587f34SStefano Zampini #undef __FUNCT__ 383674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 384674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 385674ae819SStefano Zampini { 386674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 387674ae819SStefano Zampini PetscErrorCode ierr; 388674ae819SStefano Zampini 389674ae819SStefano Zampini PetscFunctionBegin; 390674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 391674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 392674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 393674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 394785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 395674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 396f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 397f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 398785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 40063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 401674ae819SStefano Zampini PetscFunctionReturn(0); 402674ae819SStefano Zampini } 403674ae819SStefano Zampini 404674ae819SStefano Zampini #undef __FUNCT__ 405674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 406674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 407674ae819SStefano Zampini { 408674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 409674ae819SStefano Zampini PetscErrorCode ierr; 410674ae819SStefano Zampini 411674ae819SStefano Zampini PetscFunctionBegin; 412b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 413674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 414674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 415674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 416b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 417674ae819SStefano Zampini PetscFunctionReturn(0); 418674ae819SStefano Zampini } 419674ae819SStefano Zampini 420674ae819SStefano Zampini #undef __FUNCT__ 421674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 422674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 423674ae819SStefano Zampini { 424674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 42506656605SStefano Zampini PetscScalar *array; 426674ae819SStefano Zampini PetscErrorCode ierr; 427674ae819SStefano Zampini 428674ae819SStefano Zampini PetscFunctionBegin; 429674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43058da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 43106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 43206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43358da7f69SStefano Zampini } 434674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 43615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 44106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 442674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 443674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4448ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 445674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 446674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 447674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 448f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 449f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 450f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 451f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 452727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4530e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 454f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 45570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4566e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 457674ae819SStefano Zampini PetscFunctionReturn(0); 458674ae819SStefano Zampini } 459674ae819SStefano Zampini 460674ae819SStefano Zampini #undef __FUNCT__ 461f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 462f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4636bfb1811SStefano Zampini { 4646bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4656bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4666bfb1811SStefano Zampini VecType impVecType; 467e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4686bfb1811SStefano Zampini PetscErrorCode ierr; 4696bfb1811SStefano Zampini 4706bfb1811SStefano Zampini PetscFunctionBegin; 471f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 472f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 473f4ddd8eeSStefano Zampini } 474e7b262bdSStefano Zampini /* get sizes */ 475b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 476b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4776bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 478e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 479e7b262bdSStefano Zampini /* R nodes */ 480e7b262bdSStefano Zampini old_size = -1; 481e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 482e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 483e7b262bdSStefano Zampini } 484e7b262bdSStefano Zampini if (n_R != old_size) { 485e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 486e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4876bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4886bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4896bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4906bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 491e7b262bdSStefano Zampini } 492e7b262bdSStefano Zampini /* local primal dofs */ 493e7b262bdSStefano Zampini old_size = -1; 494e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 495e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 496e7b262bdSStefano Zampini } 497e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 498e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 500e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5016bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 502e7b262bdSStefano Zampini } 503e7b262bdSStefano Zampini /* local explicit constraints */ 504e7b262bdSStefano Zampini old_size = -1; 505e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 506e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 507e7b262bdSStefano Zampini } 508e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 509e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 51083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 51183b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 51283b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 51383b7ccabSStefano Zampini } 5146bfb1811SStefano Zampini PetscFunctionReturn(0); 5156bfb1811SStefano Zampini } 5166bfb1811SStefano Zampini 5176bfb1811SStefano Zampini #undef __FUNCT__ 51847f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51947f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 52088ebb749SStefano Zampini { 52125084f0cSStefano Zampini PetscErrorCode ierr; 52225084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 52388ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 52488ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 525d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 52625084f0cSStefano Zampini /* submatrices of local problem */ 52780677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52806656605SStefano Zampini /* submatrices of local coarse problem */ 52906656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 53025084f0cSStefano Zampini /* working matrices */ 53106656605SStefano Zampini Mat C_CR; 53225084f0cSStefano Zampini /* additional working stuff */ 53306656605SStefano Zampini PC pc_R; 53406656605SStefano Zampini Mat F; 53506656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 53606656605SStefano Zampini 53725084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53806656605SStefano Zampini PetscScalar *work; 53906656605SStefano Zampini PetscInt *idx_V_B; 54006656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 54106656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 542b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 54345a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 54488ebb749SStefano Zampini MatType impMatType; 54525084f0cSStefano Zampini /* some shortcuts to scalars */ 54606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54788ebb749SStefano Zampini 54888ebb749SStefano Zampini PetscFunctionBegin; 549b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 55088ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 55188ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 552b371cd4fSStefano Zampini n_B = pcis->n_B; 553b371cd4fSStefano Zampini n_D = pcis->n - n_B; 55488ebb749SStefano Zampini n_R = pcis->n - n_vertices; 55588ebb749SStefano Zampini 55688ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55788ebb749SStefano Zampini impMatType = MATSEQDENSE; 55888ebb749SStefano Zampini 55988ebb749SStefano Zampini /* vertices in boundary numbering */ 560785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5610e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 56288ebb749SStefano Zampini if (i != n_vertices) { 56322d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 56488ebb749SStefano Zampini } 56588ebb749SStefano Zampini 56606656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 56706656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 57006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 57106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 57206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 57306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 57406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 57506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 57606656605SStefano Zampini 57706656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57806656605SStefano Zampini /* allocate workspace */ 57906656605SStefano Zampini n = 0; 58006656605SStefano Zampini if (n_constraints) { 58106656605SStefano Zampini n += n_R*n_constraints; 58206656605SStefano Zampini } 58306656605SStefano Zampini if (n_vertices) { 58406656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 58580677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 58606656605SStefano Zampini } 5873301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 58806656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58906656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 59006656605SStefano Zampini } 59106656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 59206656605SStefano Zampini 59306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 59406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 59506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 59606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 59706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59806656605SStefano Zampini if (isLU || isILU || isCHOL) { 59906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 600d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 601d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 602d62866d3SStefano Zampini MatFactorType type; 603d62866d3SStefano Zampini 6046816873aSStefano Zampini F = reuse_mumps->F; 6056816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 606d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 60706656605SStefano Zampini } else { 60806656605SStefano Zampini F = NULL; 60906656605SStefano Zampini } 61006656605SStefano Zampini 61188ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 61288ebb749SStefano Zampini if (n_constraints) { 61306656605SStefano Zampini Mat M1,M2,M3; 61480677318SStefano Zampini Mat auxmat; 61506656605SStefano Zampini IS is_aux; 61680677318SStefano Zampini PetscScalar *array,*array2; 61706656605SStefano Zampini 618f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 61980677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 62088ebb749SStefano Zampini 62125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 62225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6238ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 62480677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 62588ebb749SStefano Zampini 62680677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 62780677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 62806656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62988ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 63006656605SStefano Zampini const PetscScalar *row_cmat_values; 63106656605SStefano Zampini const PetscInt *row_cmat_indices; 63206656605SStefano Zampini PetscInt size_of_constraint,j; 63388ebb749SStefano Zampini 63406656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63506656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 63606656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 63706656605SStefano Zampini } 63806656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63906656605SStefano Zampini } 64080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 64106656605SStefano Zampini if (F) { 64206656605SStefano Zampini Mat B; 64306656605SStefano Zampini 64406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 64580677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 64606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64706656605SStefano Zampini } else { 64880677318SStefano Zampini PetscScalar *marr; 64980677318SStefano Zampini 65080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 65206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 65380677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 65406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 65506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 65606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65706656605SStefano Zampini } 65880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65906656605SStefano Zampini } 66080677318SStefano Zampini if (!pcbddc->switch_static) { 66180677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 66480677318SStefano Zampini for (i=0;i<n_constraints;i++) { 66580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 66680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 66780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 67080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 67180677318SStefano Zampini } 67280677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67380677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 67480677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67580677318SStefano Zampini } else { 67680677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 67780677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 67825084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67980677318SStefano Zampini } 68080677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 68180677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 68280677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 68306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 68406656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 68580677318SStefano Zampini if (isCHOL) { 68680677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 68780677318SStefano Zampini } else { 68825084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 68980677318SStefano Zampini } 69080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 69106656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 69225084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 69325084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 69425084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 69580677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 69680677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 69780677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 69806656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 69906656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 70088ebb749SStefano Zampini } 70188ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 70288ebb749SStefano Zampini if (n_vertices) { 70306656605SStefano Zampini 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; 1376d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1377785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13783a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1379d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1380d3df7717SStefano Zampini /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */ 13810e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1382d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13833a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13843a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13853a50541eSStefano Zampini break; 13863a50541eSStefano Zampini } 13873a50541eSStefano Zampini } 1388d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1389d3df7717SStefano Zampini } else { 1390d3df7717SStefano Zampini /* Verify directly the R set */ 1391d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1392d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1393d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1394d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1395d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1396d3df7717SStefano Zampini break; 1397d3df7717SStefano Zampini } 1398d3df7717SStefano Zampini } 1399d3df7717SStefano Zampini } 1400d3df7717SStefano Zampini } 14013a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 14023a50541eSStefano Zampini vbs = bs; 14033a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 14043a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 14053a50541eSStefano Zampini } 14063a50541eSStefano Zampini } 14073a50541eSStefano Zampini } 14083a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 140953892102SStefano Zampini if (sub_schurs->reuse_mumps) { 141053892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 141153892102SStefano Zampini 141253892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 141353892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 141453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 141553892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 141653892102SStefano Zampini } else { 14173a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 141853892102SStefano Zampini } 1419a64d13efSStefano Zampini 1420a64d13efSStefano Zampini /* print some info if requested */ 1421a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1422a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1423a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14240fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1425a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1426a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14273a50541eSStefano 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); 1428a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1429a64d13efSStefano Zampini } 1430a64d13efSStefano Zampini 1431a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14326816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14336816873aSStefano Zampini IS is_aux1,is_aux2; 14346816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14356816873aSStefano Zampini 14363a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1437854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1438854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1439a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14404641a718SStefano Zampini for (i=0; i<n_D; i++) { 14414641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14424641a718SStefano Zampini } 1443a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1444a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14454641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14464641a718SStefano Zampini aux_array1[j++] = i; 1447a64d13efSStefano Zampini } 1448a64d13efSStefano Zampini } 1449a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1450a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1451a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14524641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14534641a718SStefano Zampini aux_array2[j++] = i; 1454a64d13efSStefano Zampini } 1455a64d13efSStefano Zampini } 1456a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1457a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1458a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1459a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1460a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1461a64d13efSStefano Zampini 14628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1463785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1464a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14654641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14664641a718SStefano Zampini aux_array1[j++] = i; 1467a64d13efSStefano Zampini } 1468a64d13efSStefano Zampini } 1469a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1470a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1471a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1472a64d13efSStefano Zampini } 14734641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14743a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1475d62866d3SStefano Zampini } else { 147653892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14776816873aSStefano Zampini IS tis; 14786816873aSStefano Zampini PetscInt schur_size; 14796816873aSStefano Zampini 148053892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14816816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 148253892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14836816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14846816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14856816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14866816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14876816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1488d62866d3SStefano Zampini } 1489d62866d3SStefano Zampini } 1490a64d13efSStefano Zampini PetscFunctionReturn(0); 1491a64d13efSStefano Zampini } 1492a64d13efSStefano Zampini 1493304d26faSStefano Zampini 1494304d26faSStefano Zampini #undef __FUNCT__ 1495304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1496684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1497304d26faSStefano Zampini { 1498304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1499304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1500304d26faSStefano Zampini PC pc_temp; 1501304d26faSStefano Zampini Mat A_RR; 1502f4ddd8eeSStefano Zampini MatReuse reuse; 1503304d26faSStefano Zampini PetscScalar m_one = -1.0; 1504304d26faSStefano Zampini PetscReal value; 1505af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 15069577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1507304d26faSStefano Zampini PetscErrorCode ierr; 1508e604994aSStefano Zampini /* prefixes stuff */ 1509312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1510e604994aSStefano Zampini size_t len; 1511304d26faSStefano Zampini 1512304d26faSStefano Zampini PetscFunctionBegin; 1513304d26faSStefano Zampini 1514e604994aSStefano Zampini /* compute prefixes */ 1515e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1516e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1517e604994aSStefano Zampini if (!pcbddc->current_level) { 1518e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1519e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1520e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1521e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1522e604994aSStefano Zampini } else { 1523e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1524312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1525e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1526e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1527312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1528312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 152934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 153034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1531e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1532e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1533e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1534e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1535e604994aSStefano Zampini } 1536e604994aSStefano Zampini 1537304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1538684f6988SStefano Zampini if (dirichlet) { 1539d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15403301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15413301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1542964fefecSStefano Zampini } 1543ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1544964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1545304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1546304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1547304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1548304d26faSStefano Zampini /* default */ 1549304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1550e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15519577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1552304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15539577ea80SStefano Zampini if (issbaij) { 15549577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15559577ea80SStefano Zampini } else { 1556304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15579577ea80SStefano Zampini } 1558304d26faSStefano Zampini /* Allow user's customization */ 1559304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1560304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1561304d26faSStefano Zampini } 1562d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1563d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1564d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1565d62866d3SStefano Zampini 1566d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1567d5574798SStefano Zampini } 1568304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1569304d26faSStefano Zampini if (!n_D) { 1570304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1571304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1572304d26faSStefano Zampini } 1573304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1574304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1575304d26faSStefano Zampini /* set ksp_D into pcis data */ 1576304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1577304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1578304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1579684f6988SStefano Zampini } 1580304d26faSStefano Zampini 1581304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1582684f6988SStefano Zampini A_RR = 0; 1583684f6988SStefano Zampini if (neumann) { 1584d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1585f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15868ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1587f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1588f4ddd8eeSStefano Zampini PetscInt nn_R; 158981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1590f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1591f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1592f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1593f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1594f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1595f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1596f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1597727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1598f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1599f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1600f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1601f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1602f4ddd8eeSStefano Zampini } 1603f4ddd8eeSStefano Zampini } 1604f4ddd8eeSStefano Zampini /* last check */ 1605d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1606f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1607f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1608f4ddd8eeSStefano Zampini } 1609f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1610f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1611f4ddd8eeSStefano Zampini } 16126816873aSStefano Zampini /* extract A_RR */ 16136816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 16146816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 16156816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 16166816873aSStefano Zampini if (ibs != mbs) { 16176816873aSStefano Zampini Mat newmat; 16186816873aSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 16196816873aSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16206816873aSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 16216816873aSStefano Zampini } else { 16226816873aSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16236816873aSStefano Zampini } 16243301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16253301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16266816873aSStefano Zampini } 16276816873aSStefano Zampini } else { 16286816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16296816873aSStefano Zampini 16306816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16316816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16326816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16336816873aSStefano Zampini } 1634f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1635304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1636304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1637304d26faSStefano Zampini /* default */ 1638304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1639e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1640304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16419577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16429577ea80SStefano Zampini if (issbaij) { 16439577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16449577ea80SStefano Zampini } else { 1645304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16469577ea80SStefano Zampini } 1647304d26faSStefano Zampini /* Allow user's customization */ 1648304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1649304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1650304d26faSStefano Zampini } 1651d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1652304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1653304d26faSStefano Zampini if (!n_R) { 1654304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1655304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1656304d26faSStefano Zampini } 1657d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1658d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1659d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1660d62866d3SStefano Zampini 1661d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1662d62866d3SStefano Zampini } 1663304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1664304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1665684f6988SStefano Zampini } 16666816873aSStefano Zampini /* free Neumann problem's matrix */ 16676816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1668304d26faSStefano Zampini 1669304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16700fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1671684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1672684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1673684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1674684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1675684f6988SStefano Zampini } 1676684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16770fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16780fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16790fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16800fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16810fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1682304d26faSStefano Zampini /* need to be adapted? */ 1683b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1684b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1685b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1686304d26faSStefano Zampini /* print info */ 1687304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1688e604994aSStefano 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); 1689304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1690304d26faSStefano Zampini } 1691b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16928ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1693304d26faSStefano Zampini } 1694684f6988SStefano Zampini } 1695684f6988SStefano Zampini if (neumann) { /* Neumann */ 16966816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16970fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16980fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16990fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 17000fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 17010fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1702304d26faSStefano Zampini /* need to be adapted? */ 1703b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1704b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1705304d26faSStefano Zampini /* print info */ 1706304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1707e604994aSStefano 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); 1708304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1709304d26faSStefano Zampini } 1710b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 17118ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1712304d26faSStefano Zampini } 17130fccc4e9SStefano Zampini } 1714684f6988SStefano Zampini } 1715304d26faSStefano Zampini PetscFunctionReturn(0); 1716304d26faSStefano Zampini } 1717304d26faSStefano Zampini 1718304d26faSStefano Zampini #undef __FUNCT__ 1719ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 172080677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1721674ae819SStefano Zampini { 1722674ae819SStefano Zampini PetscErrorCode ierr; 1723674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1724be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1725674ae819SStefano Zampini 1726674ae819SStefano Zampini PetscFunctionBegin; 1727be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1729be83ff47SStefano Zampini } 173080677318SStefano Zampini if (!pcbddc->switch_static) { 173180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 173380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173420c7b377SStefano Zampini } 1735be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173820c7b377SStefano Zampini } else { 1739be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1740be83ff47SStefano Zampini 174153892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174253892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1743be83ff47SStefano Zampini } 1744be83ff47SStefano Zampini } else { 174580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 175080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 175180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 175280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175420c7b377SStefano Zampini } 1755674ae819SStefano Zampini } 1756be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 175780677318SStefano Zampini if (applytranspose) { 175880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 175980677318SStefano Zampini } else { 176080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176180677318SStefano Zampini } 176253892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1763be83ff47SStefano Zampini } else { 1764be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1765be83ff47SStefano Zampini 1766be83ff47SStefano Zampini if (applytranspose) { 176753892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1768be83ff47SStefano Zampini } else { 176953892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1770be83ff47SStefano Zampini } 177153892102SStefano Zampini #endif 1772be83ff47SStefano Zampini } 177380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 177480677318SStefano Zampini if (!pcbddc->switch_static) { 1775be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 177680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1778be83ff47SStefano Zampini } else { 1779be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1780be83ff47SStefano Zampini 178153892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178253892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1783be83ff47SStefano Zampini } 178480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 178780677318SStefano Zampini } 178880677318SStefano Zampini } else { 178980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 179480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 179580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 179680677318SStefano Zampini } 179780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180180677318SStefano Zampini } 1802674ae819SStefano Zampini PetscFunctionReturn(0); 1803674ae819SStefano Zampini } 1804674ae819SStefano Zampini 1805dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1806674ae819SStefano Zampini #undef __FUNCT__ 1807674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1808dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1809674ae819SStefano Zampini { 1810674ae819SStefano Zampini PetscErrorCode ierr; 1811674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1812674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1813674ae819SStefano Zampini const PetscScalar zero = 0.0; 1814674ae819SStefano Zampini 1815674ae819SStefano Zampini PetscFunctionBegin; 1816dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1817dc359a40SStefano Zampini if (applytranspose) { 1818674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18198eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1820dc359a40SStefano Zampini } else { 1821674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1822674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 182315aaf578SStefano Zampini } 182412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 182512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 182612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182812edc857SStefano Zampini 18299f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 183012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 183112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1832964fefecSStefano Zampini Vec rhs,sol; 1833964fefecSStefano Zampini 1834964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1835964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 183612edc857SStefano Zampini if (applytranspose) { 1837964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183812edc857SStefano Zampini } else { 1839964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 184012edc857SStefano Zampini } 184112edc857SStefano Zampini } 1842674ae819SStefano Zampini 1843674ae819SStefano Zampini /* Local solution on R nodes */ 184480677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 184580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18469f00e9b4SStefano Zampini } 1847674ae819SStefano Zampini 18489f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18499f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 185012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1851674ae819SStefano Zampini 1852674ae819SStefano Zampini /* Sum contributions from two levels */ 1853dc359a40SStefano Zampini if (applytranspose) { 1854dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1855dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1856dc359a40SStefano Zampini } else { 1857674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18588eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1859dc359a40SStefano Zampini } 1860674ae819SStefano Zampini PetscFunctionReturn(0); 1861674ae819SStefano Zampini } 1862674ae819SStefano Zampini 1863674ae819SStefano Zampini #undef __FUNCT__ 1864674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 186512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1866674ae819SStefano Zampini { 1867674ae819SStefano Zampini PetscErrorCode ierr; 1868674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 186958da7f69SStefano Zampini PetscScalar *array; 187012edc857SStefano Zampini Vec from,to; 1871674ae819SStefano Zampini 1872674ae819SStefano Zampini PetscFunctionBegin; 187312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 187412edc857SStefano Zampini from = pcbddc->coarse_vec; 187512edc857SStefano Zampini to = pcbddc->vec1_P; 187612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 187712edc857SStefano Zampini Vec tvec; 187858da7f69SStefano Zampini 187958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 188112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 188358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 188458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 188512edc857SStefano Zampini } 188612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 188712edc857SStefano Zampini from = pcbddc->vec1_P; 188812edc857SStefano Zampini to = pcbddc->coarse_vec; 188912edc857SStefano Zampini } 189012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1891674ae819SStefano Zampini PetscFunctionReturn(0); 1892674ae819SStefano Zampini } 1893674ae819SStefano Zampini 1894674ae819SStefano Zampini #undef __FUNCT__ 1895674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 189612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1897674ae819SStefano Zampini { 1898674ae819SStefano Zampini PetscErrorCode ierr; 1899674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 190058da7f69SStefano Zampini PetscScalar *array; 190112edc857SStefano Zampini Vec from,to; 1902674ae819SStefano Zampini 1903674ae819SStefano Zampini PetscFunctionBegin; 190412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 190512edc857SStefano Zampini from = pcbddc->coarse_vec; 190612edc857SStefano Zampini to = pcbddc->vec1_P; 190712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 190812edc857SStefano Zampini from = pcbddc->vec1_P; 190912edc857SStefano Zampini to = pcbddc->coarse_vec; 191012edc857SStefano Zampini } 191112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 191212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 191312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 191412edc857SStefano Zampini Vec tvec; 191558da7f69SStefano Zampini 191612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 191758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 191858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 191958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 192058da7f69SStefano Zampini } 192158da7f69SStefano Zampini } else { 192258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 192358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 192412edc857SStefano Zampini } 192512edc857SStefano Zampini } 1926674ae819SStefano Zampini PetscFunctionReturn(0); 1927674ae819SStefano Zampini } 1928674ae819SStefano Zampini 1929984c4197SStefano Zampini /* uncomment for testing purposes */ 1930984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1931674ae819SStefano Zampini #undef __FUNCT__ 1932674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1933674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1934674ae819SStefano Zampini { 1935674ae819SStefano Zampini PetscErrorCode ierr; 1936674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1937674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1938674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1939984c4197SStefano Zampini /* one and zero */ 1940984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1941984c4197SStefano Zampini /* space to store constraints and their local indices */ 19429162d606SStefano Zampini PetscScalar *constraints_data; 19439162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19449162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19459162d606SStefano Zampini PetscInt *constraints_n; 1946984c4197SStefano Zampini /* iterators */ 1947b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1948984c4197SStefano Zampini /* BLAS integers */ 1949e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1950e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1951c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1952727cdba6SStefano Zampini /* reuse */ 19530e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19540e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1955984c4197SStefano Zampini /* change of basis */ 1956b3d85658SStefano Zampini PetscBool qr_needed; 19579162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1958984c4197SStefano Zampini /* auxiliary stuff */ 195964efe560SStefano Zampini PetscInt *nnz,*is_indices; 19608a0068c3SStefano Zampini PetscInt ncc; 1961984c4197SStefano Zampini /* some quantities */ 196245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1963a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1964984c4197SStefano Zampini 1965674ae819SStefano Zampini PetscFunctionBegin; 19668e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19678e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19688e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1969088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1970088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19710e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19720e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19730e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19740e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19750e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1976088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1977cf5a6209SStefano Zampini 1978cf5a6209SStefano Zampini /* print some info */ 1979cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1980cf5a6209SStefano Zampini IS vertices; 1981cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1982cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1983cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1984cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1985cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1986cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1987cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1988fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1989fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1990cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1991cf5a6209SStefano Zampini } 1992cf5a6209SStefano Zampini 1993cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19949162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1995cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1996cf5a6209SStefano Zampini const Vec *nearnullvecs; 1997cf5a6209SStefano Zampini Vec *localnearnullsp; 1998cf5a6209SStefano Zampini PetscScalar *array; 1999cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 2000cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2001cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2002b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2003cf5a6209SStefano Zampini PetscScalar *work; 2004cf5a6209SStefano Zampini PetscReal *singular_vals; 2005cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2006cf5a6209SStefano Zampini PetscReal *rwork; 2007cf5a6209SStefano Zampini #endif 2008cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2009cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2010cf5a6209SStefano Zampini #else 2011964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2012964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2013cf5a6209SStefano Zampini #endif 2014cf5a6209SStefano Zampini 2015674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2016d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2017d06fc5fdSStefano Zampini /* free unneeded index sets */ 2018d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2019d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2020674ae819SStefano Zampini } 2021d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2022d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2023d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2024d06fc5fdSStefano Zampini } 2025d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2026d06fc5fdSStefano Zampini n_ISForEdges = 0; 2027d06fc5fdSStefano Zampini } 2028d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2029d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2030d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2031d06fc5fdSStefano Zampini } 2032d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2033d06fc5fdSStefano Zampini n_ISForFaces = 0; 2034d06fc5fdSStefano Zampini } 203570022509SStefano Zampini 203670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 203770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 203870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 203970022509SStefano Zampini if (pcbddc->NullSpace) { 204070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 204170022509SStefano Zampini 204270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2043b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2044d06fc5fdSStefano Zampini if (!ISForEdges) { 2045d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2046d06fc5fdSStefano Zampini } 2047b8ffe317SStefano Zampini } 2048d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2049d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2050d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2051d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2052d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 205398a51de6SStefano Zampini } 205470022509SStefano Zampini #endif 205508122e43SStefano Zampini 2056674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2057674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2058674ae819SStefano Zampini if (nearnullsp) { 2059674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2060f4ddd8eeSStefano Zampini /* remove any stored info */ 2061f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2062f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2063f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2064f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2065f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2066473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2067f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2068f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2069f4ddd8eeSStefano Zampini } 2070984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2071984c4197SStefano Zampini nnsp_size = 0; 2072674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2073674ae819SStefano Zampini } 2074984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2075984c4197SStefano Zampini max_constraints = nnsp_size; 2076984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2077984c4197SStefano Zampini 2078674ae819SStefano Zampini /* 2079674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20809162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20819162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20829162d606SStefano Zampini There can be multiple constraints per connected component 2083674ae819SStefano Zampini */ 2084674ae819SStefano Zampini n_vertices = 0; 2085674ae819SStefano Zampini if (ISForVertices) { 2086674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2087674ae819SStefano Zampini } 20889162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20899162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20909162d606SStefano Zampini 20919162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20929162d606SStefano Zampini total_counts *= max_constraints; 2093674ae819SStefano Zampini total_counts += n_vertices; 20944641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20959162d606SStefano Zampini 2096674ae819SStefano Zampini total_counts = 0; 2097674ae819SStefano Zampini max_size_of_constraint = 0; 2098674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20999162d606SStefano Zampini IS used_is; 2100674ae819SStefano Zampini if (i<n_ISForEdges) { 21019162d606SStefano Zampini used_is = ISForEdges[i]; 2102674ae819SStefano Zampini } else { 21039162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2104674ae819SStefano Zampini } 21059162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2106674ae819SStefano Zampini total_counts += j; 2107674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2108674ae819SStefano Zampini } 21099162d606SStefano 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); 21109162d606SStefano Zampini 2111984c4197SStefano Zampini /* get local part of global near null space vectors */ 2112785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2113984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2114984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2115984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2116984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2117984c4197SStefano Zampini } 2118674ae819SStefano Zampini 2119242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2120242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2121a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2122242a89d7SStefano Zampini 2123984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2124a773dcb8SStefano Zampini if (!skip_lapack) { 2125674ae819SStefano Zampini PetscScalar temp_work; 2126911cabfeSStefano Zampini 2127674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2128984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2129785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2130785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2131785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2132674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2133785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2134674ae819SStefano Zampini #endif 2135674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2136c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2137c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2138674ae819SStefano Zampini lwork = -1; 2139674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2140674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2141c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2142674ae819SStefano Zampini #else 2143c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2144674ae819SStefano Zampini #endif 2145674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2146984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2147674ae819SStefano Zampini #else /* on missing GESVD */ 2148674ae819SStefano Zampini /* SVD */ 2149674ae819SStefano Zampini PetscInt max_n,min_n; 2150674ae819SStefano Zampini max_n = max_size_of_constraint; 2151984c4197SStefano Zampini min_n = max_constraints; 2152984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2153674ae819SStefano Zampini min_n = max_size_of_constraint; 2154984c4197SStefano Zampini max_n = max_constraints; 2155674ae819SStefano Zampini } 2156785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2157674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2158785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2159674ae819SStefano Zampini #endif 2160674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2161674ae819SStefano Zampini lwork = -1; 2162e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2163e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2164b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2165674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2166674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21679162d606SStefano 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)); 2168674ae819SStefano Zampini #else 21699162d606SStefano 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)); 2170674ae819SStefano Zampini #endif 2171674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2172984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2173984c4197SStefano Zampini #endif /* on missing GESVD */ 2174674ae819SStefano Zampini /* Allocate optimal workspace */ 2175674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2176854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2177674ae819SStefano Zampini } 2178674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2179674ae819SStefano Zampini total_counts = 0; 21809162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21819162d606SStefano Zampini constraints_data_ptr[0] = 0; 2182674ae819SStefano Zampini /* vertices */ 21839162d606SStefano Zampini if (n_vertices) { 2184674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21859162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21869162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2187674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21889162d606SStefano Zampini constraints_n[total_counts] = 1; 21899162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21909162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21919162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2192674ae819SStefano Zampini total_counts++; 2193674ae819SStefano Zampini } 2194674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2195984c4197SStefano Zampini PetscBool used_vertex; 2196674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2197674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2198674ae819SStefano Zampini k = 0; 2199674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2200984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2201984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 22029162d606SStefano Zampini constraints_n[total_counts] = 1; 22039162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 22049162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22059162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22069162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2207674ae819SStefano Zampini total_counts++; 2208674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2209674ae819SStefano Zampini } 2210984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2211674ae819SStefano Zampini k++; 2212674ae819SStefano Zampini } 2213674ae819SStefano Zampini } 2214674ae819SStefano Zampini } 2215674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2216674ae819SStefano Zampini n_vertices = total_counts; 2217674ae819SStefano Zampini } 2218984c4197SStefano Zampini 2219674ae819SStefano Zampini /* edges and faces */ 22209162d606SStefano Zampini total_counts_cc = total_counts; 2221911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22229162d606SStefano Zampini IS used_is; 22239162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22249162d606SStefano Zampini 2225911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22269162d606SStefano Zampini used_is = ISForEdges[ncc]; 2227984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2228674ae819SStefano Zampini } else { 22299162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2230984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2231674ae819SStefano Zampini } 2232674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22339162d606SStefano Zampini 22349162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22359162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2236984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2237984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2238674ae819SStefano Zampini if (nnsp_has_cnst) { 22395b08dc53SStefano Zampini PetscScalar quad_value; 22409162d606SStefano Zampini 22419162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22429162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22439162d606SStefano Zampini 2244a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2245674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2246a773dcb8SStefano Zampini } else { 2247a773dcb8SStefano Zampini quad_value = 1.0; 2248a773dcb8SStefano Zampini } 2249674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22509162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2251674ae819SStefano Zampini } 22529162d606SStefano Zampini temp_constraints++; 2253674ae819SStefano Zampini total_counts++; 2254674ae819SStefano Zampini } 2255674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2256984c4197SStefano Zampini PetscReal real_value; 22579162d606SStefano Zampini PetscScalar *ptr_to_data; 22589162d606SStefano Zampini 2259984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22609162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2261674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22629162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2263674ae819SStefano Zampini } 2264984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2265984c4197SStefano Zampini /* check if array is null on the connected component */ 2266e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22679162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22685b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2269674ae819SStefano Zampini temp_constraints++; 2270674ae819SStefano Zampini total_counts++; 22719162d606SStefano Zampini if (!idxs_copied) { 22729162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22739162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2274674ae819SStefano Zampini } 2275674ae819SStefano Zampini } 22769162d606SStefano Zampini } 22779162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 227845a1bb75SStefano Zampini valid_constraints = temp_constraints; 2279eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2280a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22819162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22829162d606SStefano Zampini 22839162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2284a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22859162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2286a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22879162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2288a773dcb8SStefano Zampini } else { /* perform SVD */ 2289984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22909162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2291674ae819SStefano Zampini 2292674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2293984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2294984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2295984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2296984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2297984c4197SStefano Zampini from that computed using LAPACKgesvd 2298984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2299984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2300984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2301674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2302e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2303984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2304674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2305674ae819SStefano Zampini for (k=0;k<j+1;k++) { 23069162d606SStefano 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)); 2307674ae819SStefano Zampini } 2308674ae819SStefano Zampini } 2309e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2310e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2311e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2312674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2313c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2314674ae819SStefano Zampini #else 2315c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2316674ae819SStefano Zampini #endif 2317674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2318984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2319984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2320674ae819SStefano Zampini j = 0; 2321984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2322674ae819SStefano Zampini total_counts = total_counts-j; 232345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2324e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2325c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2326c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2327c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2328c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2329c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2330c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2331674ae819SStefano Zampini if (j<temp_constraints) { 2332984c4197SStefano Zampini PetscInt ii; 2333984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2334674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23359162d606SStefano 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)); 2336674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2337984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2338674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23399162d606SStefano 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]; 2340674ae819SStefano Zampini } 2341674ae819SStefano Zampini } 2342674ae819SStefano Zampini } 2343674ae819SStefano Zampini #else /* on missing GESVD */ 2344e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2345e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2346b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2347674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2348674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23499162d606SStefano 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)); 2350674ae819SStefano Zampini #else 23519162d606SStefano 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)); 2352674ae819SStefano Zampini #endif 2353984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2354674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2355984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2356e310c8b4SStefano Zampini k = temp_constraints; 2357e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2358674ae819SStefano Zampini j = 0; 2359e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 236045a1bb75SStefano Zampini valid_constraints = k-j; 2361911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2362984c4197SStefano Zampini #endif /* on missing GESVD */ 2363674ae819SStefano Zampini } 2364a773dcb8SStefano Zampini } 23659162d606SStefano Zampini /* update pointers information */ 23669162d606SStefano Zampini if (valid_constraints) { 23679162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23689162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23699162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23709162d606SStefano Zampini /* set change_of_basis flag */ 237145a1bb75SStefano Zampini if (boolforchange) { 2372b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23739162d606SStefano Zampini } 2374b3d85658SStefano Zampini total_counts_cc++; 237545a1bb75SStefano Zampini } 2376674ae819SStefano Zampini } 2377984c4197SStefano Zampini /* free workspace */ 23788f1c130eSStefano Zampini if (!skip_lapack) { 2379984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2380984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2381984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2382984c4197SStefano Zampini #endif 2383984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2384984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2385984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2386984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2387984c4197SStefano Zampini #endif 2388984c4197SStefano Zampini } 2389984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2390984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2391984c4197SStefano Zampini } 2392984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2393cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2394cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2395cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2396cf5a6209SStefano Zampini } 2397cf5a6209SStefano Zampini if (n_ISForFaces) { 2398cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2399cf5a6209SStefano Zampini } 2400cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2401cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2402cf5a6209SStefano Zampini } 2403cf5a6209SStefano Zampini if (n_ISForEdges) { 2404cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2405cf5a6209SStefano Zampini } 2406cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 240708122e43SStefano Zampini } else { 240808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 240908122e43SStefano Zampini 241008122e43SStefano Zampini total_counts = 0; 241108122e43SStefano Zampini n_vertices = 0; 2412d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2413d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 241408122e43SStefano Zampini } 241508122e43SStefano Zampini max_constraints = 0; 24169162d606SStefano Zampini total_counts_cc = 0; 241708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 241808122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24199162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 242008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 242108122e43SStefano Zampini } 24229162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24239162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24249162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24259162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 242674d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24279162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24289162d606SStefano Zampini total_counts_cc = 0; 24299162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24309162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24319162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 243208122e43SStefano Zampini } 243308122e43SStefano Zampini } 24349162d606SStefano Zampini #if 0 24359162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24369162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24379162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24389162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24399162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24409162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24419162d606SStefano Zampini } 24429162d606SStefano Zampini printf("\n"); 24439162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24449162d606SStefano Zampini } 24451b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24468bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24471b968477SStefano Zampini } 24481b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24498bec7fa6SStefano 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]); 24501b968477SStefano Zampini } 245108122e43SStefano Zampini #endif 245208122e43SStefano Zampini 24538bec7fa6SStefano Zampini max_size_of_constraint = 0; 24549162d606SStefano 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]); 24559162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 245608122e43SStefano Zampini /* Change of basis */ 2457b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 245808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 245908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 246008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2461b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 246208122e43SStefano Zampini } 246308122e43SStefano Zampini } 246408122e43SStefano Zampini } 246508122e43SStefano Zampini } 2466b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24670e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 246808122e43SStefano Zampini 24699162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24709162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24719162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24729162d606SStefano 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); 247308122e43SStefano Zampini } 2474984c4197SStefano Zampini 2475674ae819SStefano Zampini /* Create constraint matrix */ 2476674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 247716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2478984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2479984c4197SStefano Zampini 2480984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 248174d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 248274d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 248374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2484984c4197SStefano Zampini total_primal_vertices=0; 2485b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24869162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24879162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2488984c4197SStefano Zampini if (size_of_constraint == 1) { 24899162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2490b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 249164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24929162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24939162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24949162d606SStefano Zampini } 2495b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 249674d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 249774d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 249874d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 249974d5cdf7SStefano Zampini } 2500b3d85658SStefano Zampini } else { 2501b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2502984c4197SStefano Zampini } 2503984c4197SStefano Zampini } 2504b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2505b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 250670022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 250770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2508b3d85658SStefano Zampini 2509b3d85658SStefano 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); 25100e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25110e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 251270022509SStefano Zampini 2513984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 251474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2515785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2516984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 251774d5cdf7SStefano Zampini 2518984c4197SStefano Zampini j = total_primal_vertices; 251974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2520b3d85658SStefano Zampini cum = total_primal_vertices; 25219162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25224641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2523b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2524b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2525b3d85658SStefano Zampini cum++; 25269162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 252774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 252874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 252974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 253074d5cdf7SStefano Zampini } 25319162d606SStefano Zampini j += constraints_n[i]; 2532674ae819SStefano Zampini } 2533674ae819SStefano Zampini } 2534674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2535674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2536088faed8SStefano Zampini 2537674ae819SStefano Zampini /* set values in constraint matrix */ 2538984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25390e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2540674ae819SStefano Zampini } 2541984c4197SStefano Zampini total_counts = total_primal_vertices; 25429162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25434641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25449162d606SStefano Zampini PetscInt *cols; 25459162d606SStefano Zampini 25469162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25479162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25489162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25499162d606SStefano Zampini PetscInt row = total_counts+k; 25509162d606SStefano Zampini PetscScalar *vals; 25519162d606SStefano Zampini 25529162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25539162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25549162d606SStefano Zampini } 25559162d606SStefano Zampini total_counts += constraints_n[i]; 2556674ae819SStefano Zampini } 2557674ae819SStefano Zampini } 2558674ae819SStefano Zampini /* assembling */ 2559674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2560674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2561088faed8SStefano Zampini 2562984c4197SStefano Zampini /* 256345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2564984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2565984c4197SStefano Zampini */ 2566674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2567674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2568026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2569984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2570984c4197SStefano Zampini /* working stuff for GEQRF */ 257181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2572984c4197SStefano Zampini PetscBLASInt lqr_work; 2573984c4197SStefano Zampini /* working stuff for UNGQR */ 2574984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2575984c4197SStefano Zampini PetscBLASInt lgqr_work; 2576984c4197SStefano Zampini /* working stuff for TRTRS */ 2577984c4197SStefano Zampini PetscScalar *trs_rhs; 25783f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2579984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2580984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2581984c4197SStefano Zampini PetscScalar *start_vals; 2582984c4197SStefano Zampini /* working stuff for values insertion */ 25834641a718SStefano Zampini PetscBT is_primal; 258464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2585906d46d4SStefano Zampini /* matrix sizes */ 2586906d46d4SStefano Zampini PetscInt global_size,local_size; 2587906d46d4SStefano Zampini /* temporary change of basis */ 2588906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2589cf5a6209SStefano Zampini /* extra space for debugging */ 2590cf5a6209SStefano Zampini PetscScalar *dbg_work; 2591984c4197SStefano Zampini 2592906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2593906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 259416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2595bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2596906d46d4SStefano Zampini /* nonzeros for local mat */ 2597bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2598bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25999162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2600a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 26019162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2602a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 26039162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2604a717540cSStefano Zampini } else { 26059162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 26069162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2607a717540cSStefano Zampini } 2608a717540cSStefano Zampini } 2609a717540cSStefano Zampini } 2610906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2611bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2612a717540cSStefano Zampini /* Set initial identity in the matrix */ 2613bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2614906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2615a717540cSStefano Zampini } 2616a717540cSStefano Zampini 2617a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2618a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2619a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2620a717540cSStefano Zampini } 2621a717540cSStefano Zampini 2622a717540cSStefano Zampini 2623a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2624a717540cSStefano Zampini /* 2625a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2626a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2627a717540cSStefano Zampini 2628a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2629a717540cSStefano Zampini 2630a6b551f4SStefano 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) 2631a6b551f4SStefano Zampini 2632a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2633a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2634a717540cSStefano Zampini | ... | 2635a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2636a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2637a717540cSStefano Zampini 2638a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2639a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2640a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2641a6b551f4SStefano Zampini 2642a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2643a717540cSStefano Zampini */ 2644a717540cSStefano Zampini if (qr_needed) { 2645984c4197SStefano Zampini /* space to store Q */ 2646854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2647984c4197SStefano Zampini /* first we issue queries for optimal work */ 26483f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26493f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26503f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2651984c4197SStefano Zampini lqr_work = -1; 26523f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2653984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2654984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2655785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2656984c4197SStefano Zampini lgqr_work = -1; 26573f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26583f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26593f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26603f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26613f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26623f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2663984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2664984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2665785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2666984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2667785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2668984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2669785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2670a717540cSStefano Zampini /* allocating workspace for check */ 2671a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2672cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2673a717540cSStefano Zampini } 2674a717540cSStefano Zampini } 2675984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26764641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2677473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26780e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 267939e2fb2aSStefano Zampini if (i != total_primal_vertices) { 268039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26814641a718SStefano Zampini } 268239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 268339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 268439e2fb2aSStefano Zampini } 268539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2686984c4197SStefano Zampini 2687a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26889162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26899162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26904641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2691984c4197SStefano Zampini /* get constraint info */ 26929162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2693984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2694984c4197SStefano Zampini 2695984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26969162d606SStefano 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); 2697674ae819SStefano Zampini } 2698984c4197SStefano Zampini 2699fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2700a717540cSStefano Zampini 2701a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2702a717540cSStefano Zampini if (pcbddc->dbg_flag) { 27039162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2704a717540cSStefano Zampini } 2705984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 27069162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2707984c4197SStefano Zampini 2708984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27093f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27103f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2712674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27133f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2714984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2715674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2716984c4197SStefano Zampini 2717984c4197SStefano Zampini /* explictly compute R^-T */ 2718984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2719984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27203f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27213f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2724984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27253f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2726984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2727984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2728984c4197SStefano Zampini 2729a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27313f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27323f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27333f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2734984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27353f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2736984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2737984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2738984c4197SStefano Zampini 2739984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2740984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2741984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27423f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27443f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27453f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27473f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2748984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27499162d606SStefano 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)); 2750984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27519162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2752984c4197SStefano Zampini 2753984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27549162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2755984c4197SStefano Zampini /* insert cols for primal dofs */ 2756984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2757984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27589162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2759906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2760984c4197SStefano Zampini } 2761984c4197SStefano Zampini /* insert cols for dual dofs */ 2762984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27639162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2764984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27659162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2766906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2767984c4197SStefano Zampini j++; 2768674ae819SStefano Zampini } 2769674ae819SStefano Zampini } 2770984c4197SStefano Zampini 2771984c4197SStefano Zampini /* check change of basis */ 2772984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2773984c4197SStefano Zampini PetscInt ii,jj; 2774984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2775c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2776c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2777c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2778c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2779c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2780c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2781984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2782cf5a6209SStefano 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)); 2783984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2784984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2785984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2786cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2787cf5a6209SStefano 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; 2788674ae819SStefano Zampini } 2789674ae819SStefano Zampini } 2790984c4197SStefano Zampini if (!valid_qr) { 279122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2792984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2793984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2794cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2795cf5a6209SStefano 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])); 2796674ae819SStefano Zampini } 2797cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2798cf5a6209SStefano 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])); 2799984c4197SStefano Zampini } 2800984c4197SStefano Zampini } 2801984c4197SStefano Zampini } 2802674ae819SStefano Zampini } else { 280322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2804674ae819SStefano Zampini } 2805674ae819SStefano Zampini } 2806a717540cSStefano Zampini } else { /* simple transformation block */ 2807a717540cSStefano Zampini PetscInt row,col; 2808a6b551f4SStefano Zampini PetscScalar val,norm; 2809a6b551f4SStefano Zampini 2810a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28119162d606SStefano 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)); 2812a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28139162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28149162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2815bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28169162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2817906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28189162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2819a717540cSStefano Zampini } else { 2820a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28219162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2822a717540cSStefano Zampini if (row != col) { 28239162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2824a717540cSStefano Zampini } else { 28259162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2826a717540cSStefano Zampini } 2827906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2828a717540cSStefano Zampini } 2829a717540cSStefano Zampini } 2830a717540cSStefano Zampini } 283198a51de6SStefano Zampini if (pcbddc->dbg_flag) { 283222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2833a717540cSStefano Zampini } 2834674ae819SStefano Zampini } 2835984c4197SStefano Zampini } else { 2836984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28379162d606SStefano 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); 2838674ae819SStefano Zampini } 2839674ae819SStefano Zampini } 2840674ae819SStefano Zampini } 2841a717540cSStefano Zampini 2842a717540cSStefano Zampini /* free workspace */ 2843a717540cSStefano Zampini if (qr_needed) { 2844984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2845cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2846984c4197SStefano Zampini } 2847984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2848984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2849984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2850984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2851984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2852674ae819SStefano Zampini } 2853a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2854906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2855906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2856906d46d4SStefano Zampini 2857906d46d4SStefano Zampini /* assembling of global change of variable */ 2858bbb9e6c6SStefano Zampini { 2859bbb9e6c6SStefano Zampini Mat tmat; 286016f15bc4SStefano Zampini PetscInt bs; 286116f15bc4SStefano Zampini 2862906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2863906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2866bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2867bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 286816f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 286916f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2870906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2871bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2872bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2873bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2874bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2875bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2878bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2879bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2880906d46d4SStefano Zampini } 2881906d46d4SStefano Zampini /* check */ 2882906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2883906d46d4SStefano Zampini PetscReal error; 2884906d46d4SStefano Zampini Vec x,x_change; 2885906d46d4SStefano Zampini 2886906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2887906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2888906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2889906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2890bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2891bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2892bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2893bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2894bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2895906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2896906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2897906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2898906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2899bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2900906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2901906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2902906d46d4SStefano Zampini } 2903b96c3477SStefano Zampini 2904b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2905b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2906b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2907b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2908ac632422SStefano Zampini Mat S_new,tmat; 2909bbb9e6c6SStefano Zampini IS is_all_N; 2910bbb9e6c6SStefano Zampini 2911bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29126816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2913bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2914ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2915b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2916ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2917ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2918ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2919ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2920ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2921b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2922ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2923ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2924ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2925ac632422SStefano Zampini } 2926b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2927b96c3477SStefano Zampini } 2928b96c3477SStefano Zampini } 2929906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2930906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2931b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2932b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2933b9b85e73SStefano Zampini } 2934906d46d4SStefano Zampini 2935906d46d4SStefano Zampini /* set up change of basis context */ 2936906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2937906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2938906d46d4SStefano Zampini 2939906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2940906d46d4SStefano Zampini PetscInt global_size,local_size; 2941906d46d4SStefano Zampini 2942906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2943906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2944906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2945906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2946906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2947906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2950906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2951906d46d4SStefano Zampini } else { 2952906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2953906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2954906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2955906d46d4SStefano Zampini } 2956906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2957906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2958906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2959906d46d4SStefano Zampini } else { 2960906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2961906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2962906d46d4SStefano Zampini } 2963906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2964906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2965906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2966906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2967b9b85e73SStefano Zampini } 2968a717540cSStefano Zampini 2969727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2970727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2971727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2972aff50787SStefano 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); 29730e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29740e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2975aff50787SStefano 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); 2976c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2977727cdba6SStefano Zampini } 29780e6343abSStefano Zampini } 29790e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2980727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2981727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2982727cdba6SStefano Zampini 2983a717540cSStefano Zampini /* flush dbg viewer */ 2984b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2985b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2986b8ffe317SStefano Zampini } 2987a717540cSStefano Zampini 2988e310c8b4SStefano Zampini /* free workspace */ 2989a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29904641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 299108122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29929162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29939162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 299408122e43SStefano Zampini } else { 29959162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29969162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29979162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 299808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 299908122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 30009162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 30019162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 300208122e43SStefano Zampini } 3003674ae819SStefano Zampini PetscFunctionReturn(0); 3004674ae819SStefano Zampini } 3005674ae819SStefano Zampini 3006674ae819SStefano Zampini #undef __FUNCT__ 3007674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3008674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3009674ae819SStefano Zampini { 3010674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3011674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3012674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30137fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3014674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3015674ae819SStefano Zampini 3016674ae819SStefano Zampini PetscFunctionBegin; 30178e61c736SStefano Zampini /* Reset previously computed graph */ 30188e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3019674ae819SStefano Zampini /* Init local Graph struct */ 30207fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30217fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3022674ae819SStefano Zampini 3023575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3024575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3025575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3026575ad6abSStefano Zampini } 30279577ea80SStefano Zampini 3028674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30294d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30304d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30314d379d7bSStefano Zampini PetscInt nvtxs; 3032e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3033674ae819SStefano Zampini 30344d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30352fffb893SStefano Zampini 30362fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30372fffb893SStefano Zampini if (flg_row) { 30384d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3039b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30402fffb893SStefano Zampini } 30412fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30429b28b941SStefano 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 */ 30434d379d7bSStefano Zampini IS is_dummy; 30444d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30454d379d7bSStefano Zampini PetscInt j,sum; 30464d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30474d379d7bSStefano Zampini const PetscInt *idxs; 30484d379d7bSStefano Zampini PCBDDCGraph graph; 30494d379d7bSStefano Zampini PetscBT is_on_boundary; 30504d379d7bSStefano Zampini 30514d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30524d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30534d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30544d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30557fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30564d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3057e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3058e496cd5dSStefano Zampini if (flg_row) { 30594d379d7bSStefano Zampini graph->xadj = xadj; 30604d379d7bSStefano Zampini graph->adjncy = adjncy; 3061e496cd5dSStefano Zampini } 30624d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30634d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3064e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30654d379d7bSStefano Zampini 30664d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30679b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30684d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30694d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30704d379d7bSStefano Zampini } 30714d379d7bSStefano Zampini } 30724d379d7bSStefano Zampini 3073e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30744d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30754d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30764d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30774d379d7bSStefano Zampini } 30784d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30794d379d7bSStefano Zampini 3080e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30814d379d7bSStefano Zampini sum = 0; 30824d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30834d379d7bSStefano Zampini PetscInt sizecc = 0; 30844d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30854d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30864d379d7bSStefano Zampini sizecc++; 30874d379d7bSStefano Zampini } 30884d379d7bSStefano Zampini } 30894d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30904d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30914d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30924d379d7bSStefano Zampini } 30934d379d7bSStefano Zampini } 30944d379d7bSStefano Zampini sum += sizecc*sizecc; 30954d379d7bSStefano Zampini } 30964d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30974d379d7bSStefano Zampini sum = 0; 3098e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30994d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 31004d379d7bSStefano Zampini cxadj[i] = sum; 31014d379d7bSStefano Zampini sum += temp; 31024d379d7bSStefano Zampini } 3103e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 31044d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31054d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31064d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31074d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31084d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31094d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31104d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31114d379d7bSStefano Zampini sizecc++; 31124d379d7bSStefano Zampini } 31134d379d7bSStefano Zampini } 31144d379d7bSStefano Zampini } 31154d379d7bSStefano Zampini } 31164d379d7bSStefano Zampini } 31179b28b941SStefano Zampini if (sum) { 3118e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31194d379d7bSStefano Zampini } else { 31204d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31214d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31224d379d7bSStefano Zampini } 31234d379d7bSStefano Zampini graph->xadj = 0; 31244d379d7bSStefano Zampini graph->adjncy = 0; 31254d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31264d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31274d379d7bSStefano Zampini } 3128674ae819SStefano Zampini } 31299b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31309b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31319b28b941SStefano Zampini } 3132674ae819SStefano Zampini 313363602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3134674ae819SStefano Zampini vertex_size = 1; 313563602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 313663602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 313795ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3139a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 314063602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3141674ae819SStefano Zampini } 314263602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 314363602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 314463602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3145674ae819SStefano Zampini } 314663602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3147674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 314863602bcaSStefano Zampini } else { 314963602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 315063602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3151854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 315263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 315363602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 315463602bcaSStefano Zampini } 315563602bcaSStefano Zampini } 3156674ae819SStefano Zampini } 3157674ae819SStefano Zampini 3158674ae819SStefano Zampini /* Setup of Graph */ 3159785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3160a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3161785d1243SStefano Zampini } 3162785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3163a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3164785d1243SStefano Zampini } 3165302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3166674ae819SStefano Zampini 3167674ae819SStefano Zampini /* Graph's connected components analysis */ 3168674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3169674ae819SStefano Zampini 3170674ae819SStefano Zampini /* print some info to stdout */ 3171674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3172302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3173674ae819SStefano Zampini } 3174fb180af4SStefano Zampini 3175fb180af4SStefano Zampini /* mark topography has done */ 3176fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3177674ae819SStefano Zampini PetscFunctionReturn(0); 3178674ae819SStefano Zampini } 3179674ae819SStefano Zampini 3180*dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3181674ae819SStefano Zampini #undef __FUNCT__ 3182674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3183*dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3184674ae819SStefano Zampini { 3185*dc456d91SStefano Zampini PetscSF sf; 3186*dc456d91SStefano Zampini PetscLayout map; 3187*dc456d91SStefano Zampini const PetscInt *idxs; 3188*dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3189*dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3190*dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3191*dc456d91SStefano Zampini PetscMPIInt commsize; 3192674ae819SStefano Zampini PetscBool first_found; 3193674ae819SStefano Zampini PetscErrorCode ierr; 3194674ae819SStefano Zampini 3195674ae819SStefano Zampini PetscFunctionBegin; 3196*dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3197*dc456d91SStefano Zampini if (subset_mult) { 3198*dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3199*dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3200*dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3201674ae819SStefano Zampini } 3202*dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3203*dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3204*dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3205*dc456d91SStefano Zampini for (i=0;i<n;i++) { 3206*dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3207*dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3208674ae819SStefano Zampini } 3209*dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3210*dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3211*dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3212*dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3213*dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3214*dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3215*dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3216*dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3217*dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3218*dc456d91SStefano Zampini 3219*dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3220*dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3221*dc456d91SStefano Zampini if (subset_mult) { 3222*dc456d91SStefano Zampini const PetscInt* idxs_mult; 3223*dc456d91SStefano Zampini 3224*dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3225*dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3226*dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3227674ae819SStefano Zampini } else { 3228*dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3229674ae819SStefano Zampini } 3230*dc456d91SStefano Zampini /* local size of new subset */ 3231*dc456d91SStefano Zampini n_n = 0; 3232*dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3233*dc456d91SStefano Zampini 3234*dc456d91SStefano Zampini /* global indexes in layout */ 3235*dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3236*dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3237*dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3238*dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3239*dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3240*dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3241*dc456d91SStefano Zampini 3242*dc456d91SStefano Zampini /* reduce from leaves to roots */ 3243*dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 3244*dc456d91SStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPIU_MAX);CHKERRQ(ierr); 3245*dc456d91SStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPIU_MAX);CHKERRQ(ierr); 3246*dc456d91SStefano Zampini 3247*dc456d91SStefano Zampini /* count indexes in local part of layout */ 3248674ae819SStefano Zampini nlocals = 0; 3249674ae819SStefano Zampini first_index = -1; 3250674ae819SStefano Zampini first_found = PETSC_FALSE; 3251*dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3252*dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3253674ae819SStefano Zampini first_found = PETSC_TRUE; 3254674ae819SStefano Zampini first_index = i; 3255674ae819SStefano Zampini } 3256*dc456d91SStefano Zampini nlocals += root_data[i]; 3257674ae819SStefano Zampini } 3258*dc456d91SStefano Zampini 3259*dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 3260*dc456d91SStefano Zampini start = 0; 3261*dc456d91SStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPIU_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3262*dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3263*dc456d91SStefano Zampini *N_n = start + nlocals; 3264*dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3265*dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3266674ae819SStefano Zampini } 3267*dc456d91SStefano Zampini /* adapt roof data with cumulative */ 3268674ae819SStefano Zampini if (first_found) { 3269*dc456d91SStefano Zampini PetscInt old_index; 3270*dc456d91SStefano Zampini 3271*dc456d91SStefano Zampini root_data[first_index] += start; 3272674ae819SStefano Zampini old_index = first_index; 3273*dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3274*dc456d91SStefano Zampini if (root_data[i]) { 3275*dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3276674ae819SStefano Zampini old_index = i; 3277674ae819SStefano Zampini } 3278674ae819SStefano Zampini } 3279674ae819SStefano Zampini } 3280*dc456d91SStefano Zampini 3281*dc456d91SStefano Zampini /* from roots to leaves */ 3282*dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3283*dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3284*dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3285*dc456d91SStefano Zampini 3286*dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3287*dc456d91SStefano Zampini if (subset_mult) { 3288*dc456d91SStefano Zampini const PetscInt* idxs_mult; 3289*dc456d91SStefano Zampini PetscInt cum; 3290*dc456d91SStefano Zampini 3291*dc456d91SStefano Zampini cum = 0; 3292*dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3293*dc456d91SStefano Zampini for (i=0;i<n;i++) { 3294*dc456d91SStefano Zampini PetscInt j; 3295*dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3296674ae819SStefano Zampini } 3297*dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3298674ae819SStefano Zampini } else { 3299*dc456d91SStefano Zampini for (i=0;i<n;i++) { 3300*dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3301674ae819SStefano Zampini } 3302674ae819SStefano Zampini } 3303*dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3304*dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3305674ae819SStefano Zampini PetscFunctionReturn(0); 3306674ae819SStefano Zampini } 33079a7d3425SStefano Zampini 33089a7d3425SStefano Zampini #undef __FUNCT__ 33099a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33109a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33119a7d3425SStefano Zampini { 33129a7d3425SStefano Zampini PetscInt i,j; 33139a7d3425SStefano Zampini PetscScalar *alphas; 33149a7d3425SStefano Zampini PetscErrorCode ierr; 33159a7d3425SStefano Zampini 33169a7d3425SStefano Zampini PetscFunctionBegin; 33179a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3318785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33199a7d3425SStefano Zampini for (i=0;i<n;i++) { 33209a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33219a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33229a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33239a7d3425SStefano Zampini } 33249a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33259a7d3425SStefano Zampini PetscFunctionReturn(0); 33269a7d3425SStefano Zampini } 33279a7d3425SStefano Zampini 3328e7931f94SStefano Zampini #undef __FUNCT__ 332970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 333028143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3331e7931f94SStefano Zampini { 3332e7931f94SStefano Zampini Mat subdomain_adj; 3333e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3334e7931f94SStefano Zampini MatPartitioning partitioner; 3335e7931f94SStefano Zampini Mat_IS *matis; 3336e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3337e7931f94SStefano Zampini PetscInt prank; 3338e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3339e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3340e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33413837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3342e7931f94SStefano Zampini PetscErrorCode ierr; 33432b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3344e7931f94SStefano Zampini PetscSubcomm subcomm; 3345a57a6d2fSStefano Zampini 3346e7931f94SStefano Zampini PetscFunctionBegin; 33472b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33482b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33492b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3350e7931f94SStefano Zampini 3351e7931f94SStefano Zampini /* Get info on mapping */ 3352e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3353e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3354e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3355e7931f94SStefano Zampini 3356e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3357785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3358e7931f94SStefano Zampini xadj[0] = 0; 3359e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3360785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3361785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3362e7931f94SStefano Zampini 33632b510759SStefano Zampini if (threshold) { 3364d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33652b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3366d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3367d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3368d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3369d023bfaeSStefano Zampini xadj_count++; 3370e7931f94SStefano Zampini } 3371e7931f94SStefano Zampini } 3372d023bfaeSStefano Zampini xadj[1] = xadj_count; 3373c8587f34SStefano Zampini } else { 3374e7931f94SStefano Zampini if (xadj[1]) { 3375e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3376e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3377c8587f34SStefano Zampini } 3378e7931f94SStefano Zampini } 3379e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3380e7931f94SStefano Zampini if (use_square) { 3381e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3382e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3383e7931f94SStefano Zampini } 3384e7931f94SStefano Zampini } 3385e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3386e7931f94SStefano Zampini 33873837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3388e7931f94SStefano Zampini 3389e7931f94SStefano Zampini /* 3390e7931f94SStefano Zampini Restrict work on active processes only. 3391e7931f94SStefano Zampini */ 3392e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3393e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3394e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33952b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3396d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3397e7931f94SStefano Zampini if (color) { 3398e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3399e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3400e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3401c8587f34SStefano Zampini } else { 340228143c3dSStefano Zampini PetscInt coarsening_ratio; 3403306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3404785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3405e7931f94SStefano Zampini prank = rank; 3406306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34078002ef2cSStefano Zampini /* 3408e7931f94SStefano Zampini for (i=0;i<size;i++) { 3409e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3410c8587f34SStefano Zampini } 34118002ef2cSStefano Zampini */ 3412e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3413e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3414c8587f34SStefano Zampini } 3415e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3416306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 341722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3418e7931f94SStefano Zampini 3419e7931f94SStefano Zampini /* Partition */ 3420306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3421e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3422e7931f94SStefano Zampini if (use_vwgt) { 34233837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3424e7931f94SStefano Zampini v_wgt[0] = local_size; 3425e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3426c8587f34SStefano Zampini } 342728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 342828143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 342928143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3430e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3431e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 343222b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3433e7931f94SStefano Zampini 3434e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 343528143c3dSStefano Zampini if (contiguous) { 343628143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 343728143c3dSStefano Zampini } else { 343884ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 343928143c3dSStefano Zampini } 3440e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3441e7931f94SStefano Zampini /* clean up */ 3442e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3443e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3444e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3445e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3446e7931f94SStefano Zampini } 3447e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3448e7931f94SStefano Zampini 3449e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3450e7931f94SStefano Zampini i = 1; 3451e7931f94SStefano Zampini if (color) i=0; 3452e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3453e7931f94SStefano Zampini 3454e7931f94SStefano Zampini /* get back IS */ 3455e7931f94SStefano Zampini *is_sends = ranks_send_to; 3456e7931f94SStefano Zampini PetscFunctionReturn(0); 3457e7931f94SStefano Zampini } 3458e7931f94SStefano Zampini 3459e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3460e7931f94SStefano Zampini 3461e7931f94SStefano Zampini #undef __FUNCT__ 3462e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 346353a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3464e7931f94SStefano Zampini { 346570cf5478SStefano Zampini Mat local_mat; 3466e7931f94SStefano Zampini Mat_IS *matis; 3467e7931f94SStefano Zampini IS is_sends_internal; 34689d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 346928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34709d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3471e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3472e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3473e7931f94SStefano Zampini const PetscInt* is_indices; 3474e7931f94SStefano Zampini MatType new_local_type; 3475e7931f94SStefano Zampini /* buffers */ 3476e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 347728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34789d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3479e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3480e7931f94SStefano Zampini /* MPI */ 348128143c3dSStefano Zampini MPI_Comm comm,comm_n; 348228143c3dSStefano Zampini PetscSubcomm subcomm; 3483e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 348428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 348528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 348628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 348728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 348828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3489e7931f94SStefano Zampini PetscErrorCode ierr; 3490e7931f94SStefano Zampini 3491e7931f94SStefano Zampini PetscFunctionBegin; 349228143c3dSStefano Zampini /* TODO: add missing checks */ 349328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 349428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 349528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 349628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3497e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 349828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3499e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3500e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3501e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3502e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3503e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 350428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 350570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 350670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 350728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 350870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 350970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 351070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 351170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 351270cf5478SStefano Zampini } 3513e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3514e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3515e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3516e7931f94SStefano Zampini if (!is_sends) { 351728143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 351828143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 351928143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3520c8587f34SStefano Zampini } else { 3521e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3522e7931f94SStefano Zampini is_sends_internal = is_sends; 3523c8587f34SStefano Zampini } 3524e7931f94SStefano Zampini 3525e7931f94SStefano Zampini /* get pointer of MATIS data */ 3526e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3527e7931f94SStefano Zampini 3528e7931f94SStefano Zampini /* get comm */ 3529a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3530e7931f94SStefano Zampini 3531e7931f94SStefano Zampini /* compute number of sends */ 3532e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3533e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3534e7931f94SStefano Zampini 3535e7931f94SStefano Zampini /* compute number of receives */ 3536e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3537785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3538e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3539e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3540e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3541e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3542e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3543e7931f94SStefano Zampini 354428143c3dSStefano Zampini /* restrict comm if requested */ 354528143c3dSStefano Zampini subcomm = 0; 354628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 354728143c3dSStefano Zampini if (restrict_comm) { 3548779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3549779c1cceSStefano Zampini 355028143c3dSStefano Zampini color = 0; 355153a05cb3SStefano Zampini if (restrict_full) { 355253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 355353a05cb3SStefano Zampini } else { 355453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 355553a05cb3SStefano Zampini } 355628143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 355728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 355828143c3dSStefano Zampini /* check if reuse has been requested */ 355928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 356028143c3dSStefano Zampini if (*mat_n) { 356128143c3dSStefano Zampini PetscMPIInt subcommsize2; 356228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 356328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 356428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 356528143c3dSStefano Zampini } else { 356628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 356728143c3dSStefano Zampini } 356828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3569779c1cceSStefano Zampini PetscMPIInt rank; 3570779c1cceSStefano Zampini 3571779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 357228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 357328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 357428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3575306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 357628143c3dSStefano Zampini } 357728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 357828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 357928143c3dSStefano Zampini } else { 358028143c3dSStefano Zampini comm_n = comm; 358128143c3dSStefano Zampini } 358228143c3dSStefano Zampini 3583e7931f94SStefano Zampini /* prepare send/receive buffers */ 3584785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3585e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3586785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3587e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 358828143c3dSStefano Zampini if (nis) { 3589854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 359028143c3dSStefano Zampini } 3591e7931f94SStefano Zampini 359228143c3dSStefano Zampini /* Get data from local matrices */ 3593e7931f94SStefano Zampini if (!isdense) { 3594a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3595e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3596e7931f94SStefano Zampini /* 3597e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3598e7931f94SStefano Zampini send_buffer_idxs should contain: 3599e7931f94SStefano Zampini - MatType_PRIVATE type 3600e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3601e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3602e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3603e7931f94SStefano Zampini */ 3604e7931f94SStefano Zampini } else { 3605e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3606e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3607854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3608e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3609e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3610e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3611e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3612e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3613e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3614e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3615e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3616e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3617c8587f34SStefano Zampini } 3618c8587f34SStefano Zampini } 3619e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 362028143c3dSStefano Zampini /* additional is (if any) */ 362128143c3dSStefano Zampini if (nis) { 362228143c3dSStefano Zampini PetscMPIInt psum; 362328143c3dSStefano Zampini PetscInt j; 362428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 362528143c3dSStefano Zampini PetscInt plen; 362628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 362728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 362828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 362928143c3dSStefano Zampini } 3630854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 363128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 363228143c3dSStefano Zampini PetscInt plen; 363328143c3dSStefano Zampini const PetscInt *is_array_idxs; 363428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 363528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 363628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 363728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 363828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 363928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 364028143c3dSStefano Zampini } 364128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 364228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 364328143c3dSStefano Zampini } 364428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 364528143c3dSStefano Zampini } 364628143c3dSStefano Zampini 3647e7931f94SStefano Zampini buf_size_idxs = 0; 3648e7931f94SStefano Zampini buf_size_vals = 0; 364928143c3dSStefano Zampini buf_size_idxs_is = 0; 3650e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3651e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3652e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 365328143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3654e7931f94SStefano Zampini } 3655785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3656785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 365795ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3658e7931f94SStefano Zampini 3659e7931f94SStefano Zampini /* get new tags for clean communications */ 3660e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3661e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 366228143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3663e7931f94SStefano Zampini 3664e7931f94SStefano Zampini /* allocate for requests */ 3665785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3666785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 366795ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3668785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3669785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 367095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3671e7931f94SStefano Zampini 3672e7931f94SStefano Zampini /* communications */ 3673e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3674e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 367528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3676e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3677e7931f94SStefano Zampini source_dest = onodes[i]; 3678e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3679e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3680e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3681e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 368228143c3dSStefano Zampini if (nis) { 368328143c3dSStefano 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); 368428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 368528143c3dSStefano Zampini } 3686e7931f94SStefano Zampini } 3687e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3688e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3689e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3690e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 369128143c3dSStefano Zampini if (nis) { 369228143c3dSStefano 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); 369328143c3dSStefano Zampini } 3694e7931f94SStefano Zampini } 3695e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3696e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3697e7931f94SStefano Zampini 3698e7931f94SStefano Zampini /* assemble new l2g map */ 3699e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3700e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37019d30be91SStefano Zampini new_local_rows = 0; 3702e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37039d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3704e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3705e7931f94SStefano Zampini } 37069d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3707e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37089d30be91SStefano Zampini new_local_rows = 0; 3709e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37109d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37119d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3712e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3713e7931f94SStefano Zampini } 37149d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37159d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3716e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3717e7931f94SStefano Zampini 3718e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3719e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3720e7931f94SStefano 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) */ 3721e7931f94SStefano Zampini if (n_recvs) { 372228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3723e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3724e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3725e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3726e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3727e7931f94SStefano Zampini break; 3728e7931f94SStefano Zampini } 3729e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3730e7931f94SStefano Zampini } 3731e7931f94SStefano Zampini switch (new_local_type_private) { 373228143c3dSStefano Zampini case MATDENSE_PRIVATE: 373328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3734e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3735e7931f94SStefano Zampini bs = 1; 373628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 373728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 373828143c3dSStefano Zampini bs = 1; 373928143c3dSStefano Zampini } 3740e7931f94SStefano Zampini break; 3741e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3742e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3743e7931f94SStefano Zampini bs = 1; 3744e7931f94SStefano Zampini break; 3745e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3746e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3747e7931f94SStefano Zampini break; 3748e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3749e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3750e7931f94SStefano Zampini break; 3751e7931f94SStefano Zampini default: 37529d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3753e7931f94SStefano Zampini break; 3754e7931f94SStefano Zampini } 375528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 375628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 375728143c3dSStefano Zampini bs = 1; 3758e7931f94SStefano Zampini } 3759e7931f94SStefano Zampini 376070cf5478SStefano Zampini /* create MATIS object if needed */ 376170cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3762e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 376328143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 376470cf5478SStefano Zampini } else { 376570cf5478SStefano Zampini /* it also destroys the local matrices */ 376670cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 376770cf5478SStefano Zampini } 376870cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3769e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37709d30be91SStefano Zampini 37719d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37729d30be91SStefano Zampini 37739d30be91SStefano Zampini /* Global to local map of received indices */ 37749d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37759d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37769d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37779d30be91SStefano Zampini 37789d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37799d30be91SStefano Zampini buf_size_idxs = 0; 37809d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37819d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37829d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37839d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37849d30be91SStefano Zampini } 37859d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37869d30be91SStefano Zampini 37879d30be91SStefano Zampini /* set preallocation */ 37889d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37899d30be91SStefano Zampini if (!newisdense) { 37909d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37919d30be91SStefano Zampini 37929d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37939d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37949d30be91SStefano Zampini if (n_recvs) { 37959d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37969d30be91SStefano Zampini } 37979d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37989d30be91SStefano Zampini PetscInt j; 37999d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38009d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38019d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38029d30be91SStefano Zampini } 38039d30be91SStefano Zampini } else { 38049d30be91SStefano Zampini /* TODO */ 38059d30be91SStefano Zampini } 38069d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38079d30be91SStefano Zampini } 38089d30be91SStefano Zampini if (new_local_nnz) { 38099d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38109d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38119d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38129d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38139d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38149d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38159d30be91SStefano Zampini } else { 38169d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38179d30be91SStefano Zampini } 38189d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38199d30be91SStefano Zampini } else { 38209d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38219d30be91SStefano Zampini } 3822e7931f94SStefano Zampini 3823e7931f94SStefano Zampini /* set values */ 3824e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38259d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3826e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3827e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3828e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38299d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3830e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3831e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3832e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 383328143c3dSStefano Zampini } else { 383428143c3dSStefano Zampini /* TODO */ 3835e7931f94SStefano Zampini } 3836e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3837e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3838e7931f94SStefano Zampini } 3839e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3840e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 384170cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 384270cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38439d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38449d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3845e7931f94SStefano Zampini 3846dfd14d43SStefano Zampini #if 0 384728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3848e7931f94SStefano Zampini Vec lvec,rvec; 3849e7931f94SStefano Zampini PetscReal infty_error; 3850e7931f94SStefano Zampini 38512a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3852e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3853e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3854e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 385570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3856e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3857e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3858e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3859e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3860e7931f94SStefano Zampini } 386128143c3dSStefano Zampini #endif 3862e7931f94SStefano Zampini 386328143c3dSStefano Zampini /* assemble new additional is (if any) */ 386428143c3dSStefano Zampini if (nis) { 386528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 386628143c3dSStefano Zampini 386728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3868854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 386928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 387028143c3dSStefano Zampini psum = 0; 387128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 387228143c3dSStefano Zampini for (j=0;j<nis;j++) { 387328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 387428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 387528143c3dSStefano Zampini psum += plen; 387628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 387728143c3dSStefano Zampini } 387828143c3dSStefano Zampini } 3879854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3880854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 388128143c3dSStefano Zampini for (i=1;i<nis;i++) { 388228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 388328143c3dSStefano Zampini } 388428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 388528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 388628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 388728143c3dSStefano Zampini for (j=0;j<nis;j++) { 388828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 388928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 389028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 389128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 389228143c3dSStefano Zampini } 389328143c3dSStefano Zampini } 389428143c3dSStefano Zampini for (i=0;i<nis;i++) { 389528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 389628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 389728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 389828143c3dSStefano Zampini } 389928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 390028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 390128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 390228143c3dSStefano Zampini } 3903e7931f94SStefano Zampini /* free workspace */ 390428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3905e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3906e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3907e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3908e7931f94SStefano Zampini if (isdense) { 3909e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3910e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3911e7931f94SStefano Zampini } else { 3912e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3913e7931f94SStefano Zampini } 391428143c3dSStefano Zampini if (nis) { 391528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 391628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 391728143c3dSStefano Zampini } 3918e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3919e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 392028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3921e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3922e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 392328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3924e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3925e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3926e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3927e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3928e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 392928143c3dSStefano Zampini if (nis) { 393028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 393128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 393228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 393328143c3dSStefano Zampini } 393428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 393528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 393628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 393728143c3dSStefano Zampini for (i=0;i<nis;i++) { 393828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 393928143c3dSStefano Zampini } 394053a05cb3SStefano Zampini *mat_n = NULL; 394128143c3dSStefano Zampini } 3942e7931f94SStefano Zampini PetscFunctionReturn(0); 3943e7931f94SStefano Zampini } 3944a57a6d2fSStefano Zampini 394512edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3946af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 394712edc857SStefano Zampini 3948c8587f34SStefano Zampini #undef __FUNCT__ 3949c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3950c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3951c8587f34SStefano Zampini { 3952c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3953c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 395420a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39559881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 395620a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39576e683305SStefano Zampini IS coarse_is,*isarray; 39586e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39596e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3960f9eb5b7dSStefano Zampini PC pc_temp; 3961c8587f34SStefano Zampini PCType coarse_pc_type; 3962c8587f34SStefano Zampini KSPType coarse_ksp_type; 3963f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39644f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39656e683305SStefano Zampini Mat t_coarse_mat_is; 39666e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39676e683305SStefano Zampini PetscMPIInt all_procs; 396874e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 396968457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 397022bc73bbSStefano Zampini PetscScalar *array; 39719881197aSStefano Zampini PetscErrorCode ierr; 3972fdc09c96SStefano Zampini 3973c8587f34SStefano Zampini PetscFunctionBegin; 3974c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 397568457ee5SStefano 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 */ 3976fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39775a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3978fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3979f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3980f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3981f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3982fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3983fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3984727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3985fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3986fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3987fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3988f4ddd8eeSStefano Zampini } 3989fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3990fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3991f4ddd8eeSStefano Zampini } 399270cf5478SStefano Zampini /* reset any subassembling information */ 399370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39946e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39956e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3996fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3997f4ddd8eeSStefano Zampini } 3998c8587f34SStefano Zampini 39996e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40002b510759SStefano Zampini im_active = !!(pcis->n); 40012b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40026e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40036e683305SStefano Zampini void_procs = all_procs-active_procs; 40046e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 400574e2c79eSStefano Zampini redist = PETSC_FALSE; 400622bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40076e683305SStefano Zampini csin_ml = PETSC_TRUE; 40086e683305SStefano Zampini ncoarse_ml = void_procs; 4009779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4010779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 401118a45a71SStefano Zampini csin_ds = PETSC_TRUE; 401218a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 401318a45a71SStefano Zampini redist = PETSC_TRUE; 401418a45a71SStefano Zampini } else { 40156e683305SStefano Zampini csin_ds = PETSC_TRUE; 4016779c1cceSStefano Zampini ncoarse_ds = active_procs; 4017779c1cceSStefano Zampini redist = PETSC_TRUE; 401818a45a71SStefano Zampini } 40196e683305SStefano Zampini } else { 40206e683305SStefano Zampini csin_ml = PETSC_FALSE; 40216e683305SStefano Zampini ncoarse_ml = all_procs; 40226e683305SStefano Zampini if (void_procs) { 40236e683305SStefano Zampini csin_ds = PETSC_TRUE; 40246e683305SStefano Zampini ncoarse_ds = void_procs; 40256e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40266e683305SStefano Zampini } else { 4027779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 402874e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 402974e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 403074e2c79eSStefano Zampini redist = PETSC_TRUE; 403174e2c79eSStefano Zampini } else { 40326e683305SStefano Zampini csin_ds = PETSC_FALSE; 40336e683305SStefano Zampini ncoarse_ds = all_procs; 40346e683305SStefano Zampini } 40356e683305SStefano Zampini } 403674e2c79eSStefano Zampini } 40376e683305SStefano Zampini 40386e683305SStefano Zampini /* 40396e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40406e683305SStefano Zampini - we have not exceeded the number of levels requested 40416e683305SStefano Zampini - we can actually subassemble the active processes 40426e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40436e683305SStefano Zampini */ 40446e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40456e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40466e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40476e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40486e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4049f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40502b510759SStefano Zampini } else { 4051f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4052c8587f34SStefano Zampini } 4053c8587f34SStefano Zampini } 40546e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40556e683305SStefano Zampini if (multilevel_allowed) { 40566e683305SStefano Zampini ncoarse = ncoarse_ml; 40576e683305SStefano Zampini csin = csin_ml; 405858da7f69SStefano Zampini redist = PETSC_FALSE; 40596e683305SStefano Zampini } else { 40606e683305SStefano Zampini ncoarse = ncoarse_ds; 40616e683305SStefano Zampini csin = csin_ds; 40626e683305SStefano Zampini } 4063e7931f94SStefano Zampini 4064abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4065abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4066abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4067abbbba34SStefano Zampini 4068abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 406922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 407022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 407122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 407222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4073b9b85e73SStefano Zampini #if 0 4074b9b85e73SStefano Zampini { 4075b9b85e73SStefano Zampini PetscViewer viewer; 4076b9b85e73SStefano Zampini char filename[256]; 4077b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4078b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4079b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4080b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4081b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4082b9b85e73SStefano Zampini } 4083b9b85e73SStefano Zampini #endif 40846e683305SStefano 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); 40856e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40866e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40876e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4088abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4089abbbba34SStefano Zampini 40906e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40916e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40926e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40936e683305SStefano Zampini const PetscInt *idxs; 40946e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40956e683305SStefano Zampini 40966e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40970be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40986e683305SStefano Zampini /* allocate space for temporary storage */ 4099854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4100854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41016e683305SStefano Zampini /* allocate for IS array */ 41026e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41036e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41046e683305SStefano Zampini nis = nisdofs + nisneu; 4105854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41066e683305SStefano Zampini /* dofs splitting */ 41076e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41086e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41096e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41106e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41116e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41126e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41136e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41146e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41156e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41166e683305SStefano Zampini } 41176e683305SStefano Zampini /* neumann boundaries */ 41186e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41196e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41206e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41216e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41226e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41236e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41246e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41256e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41266e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41276e683305SStefano Zampini } 41286e683305SStefano Zampini /* free memory */ 41296e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41306e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41316e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41326e683305SStefano Zampini } else { 41336e683305SStefano Zampini nis = 0; 41346e683305SStefano Zampini nisdofs = 0; 41356e683305SStefano Zampini nisneu = 0; 41366e683305SStefano Zampini isarray = NULL; 41376e683305SStefano Zampini } 41386e683305SStefano Zampini /* destroy no longer needed map */ 41396e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41406e683305SStefano Zampini 41416e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41426e683305SStefano Zampini coarse_mat_is = NULL; 41436e683305SStefano Zampini if (csin) { 41446e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 414574e2c79eSStefano Zampini if (redist) { 414674e2c79eSStefano Zampini PetscMPIInt rank; 4147779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 414874e2c79eSStefano Zampini 414974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 415058da7f69SStefano Zampini spc = active_procs/ncoarse; 415158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4152779c1cceSStefano Zampini if (im_active) { 4153779c1cceSStefano Zampini destsize = 1; 415474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 415574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 415674e2c79eSStefano Zampini } else { 415774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 415874e2c79eSStefano Zampini } 415974e2c79eSStefano Zampini } else { 4160779c1cceSStefano Zampini destsize = 0; 41616e683305SStefano Zampini } 4162779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4163779c1cceSStefano Zampini } else if (csin_type_simple) { 41646e683305SStefano Zampini PetscMPIInt rank; 41656e683305SStefano Zampini PetscInt issize,isidx; 4166779c1cceSStefano Zampini 41676e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41686e683305SStefano Zampini if (im_active) { 41696e683305SStefano Zampini issize = 1; 41706e683305SStefano Zampini isidx = (PetscInt)rank; 41716e683305SStefano Zampini } else { 41726e683305SStefano Zampini issize = 0; 41736e683305SStefano Zampini isidx = -1; 41746e683305SStefano Zampini } 41756e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4176779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41776e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41786e683305SStefano Zampini } 4179779c1cceSStefano Zampini 4180779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4181779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4182779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4183779c1cceSStefano Zampini PetscInt *coarse_candidates; 4184779c1cceSStefano Zampini const PetscInt* tisindices; 4185779c1cceSStefano Zampini 4186779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4187779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4188779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4189779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4190779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4191779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4192779c1cceSStefano Zampini } 4193779c1cceSStefano Zampini } 4194779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4195779c1cceSStefano Zampini 4196779c1cceSStefano Zampini 41976e683305SStefano Zampini if (pcbddc->dbg_flag) { 41986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42006e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4202779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42046e683305SStefano Zampini } 42056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42066e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42076e683305SStefano Zampini } 42086e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42096e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42106e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4211854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42126e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42136e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42146e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42156e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42166e683305SStefano Zampini } 42176e683305SStefano Zampini if (pcbddc->dbg_flag) { 42186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42206e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42216e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42226e683305SStefano Zampini } 4223779c1cceSStefano Zampini } 42246e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 422553a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 422653a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 422753a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 422853a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 422953a05cb3SStefano Zampini } 42306e683305SStefano Zampini } else { 42316e683305SStefano Zampini if (pcbddc->dbg_flag) { 42326e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42346e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42356e683305SStefano Zampini } 42366e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42376e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42386e683305SStefano Zampini } 42396e683305SStefano Zampini 42406e683305SStefano Zampini /* create local to global scatters for coarse problem */ 424168457ee5SStefano Zampini if (compute_vecs) { 42426e683305SStefano Zampini PetscInt lrows; 42436e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42446e683305SStefano Zampini if (coarse_mat_is) { 42456e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42466e683305SStefano Zampini } else { 42476e683305SStefano Zampini lrows = 0; 42486e683305SStefano Zampini } 42496e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42506e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42516e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42526e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42536e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42546e683305SStefano Zampini } 42556e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42566e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4257c8587f34SStefano Zampini 4258f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4259f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4260f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4261f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4262f9eb5b7dSStefano Zampini } else { 4263f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4264f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4265c8587f34SStefano Zampini } 4266c8587f34SStefano Zampini 42676e683305SStefano Zampini /* print some info if requested */ 42686e683305SStefano Zampini if (pcbddc->dbg_flag) { 42696e683305SStefano Zampini if (!multilevel_allowed) { 42706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42716e683305SStefano Zampini if (multilevel_requested) { 42726e683305SStefano 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); 42736e683305SStefano Zampini } else if (pcbddc->max_levels) { 42746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42756e683305SStefano Zampini } 42766e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42776e683305SStefano Zampini } 42786e683305SStefano Zampini } 42796e683305SStefano Zampini 4280f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42816e683305SStefano Zampini if (coarse_mat_is) { 42826e683305SStefano Zampini MatReuse coarse_mat_reuse; 42836a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42846e683305SStefano Zampini if (pcbddc->dbg_flag) { 42856e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42866e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42876e683305SStefano Zampini } 4288f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4289312be037SStefano Zampini char prefix[256],str_level[16]; 4290e604994aSStefano Zampini size_t len; 42916e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4292c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4293f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42945f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4295c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42966e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4297c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4298c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42993e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4300e604994aSStefano Zampini /* prefix */ 4301e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4302e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4303e604994aSStefano Zampini if (!pcbddc->current_level) { 4304e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4305e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4306c8587f34SStefano Zampini } else { 4307e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4308312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4309312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 431034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4311312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4312e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4313e604994aSStefano Zampini } 4314e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43153e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43163e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43173e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43183e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4319f9eb5b7dSStefano Zampini /* allow user customization */ 4320f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43213e3c6dadSStefano Zampini } 43223e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43233e3c6dadSStefano Zampini if (nisdofs) { 43243e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43253e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43263e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43273e3c6dadSStefano Zampini } 43283e3c6dadSStefano Zampini } 43293e3c6dadSStefano Zampini if (nisneu) { 43303e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43313e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4332312be037SStefano Zampini } 4333f9eb5b7dSStefano Zampini 4334f9eb5b7dSStefano Zampini /* get some info after set from options */ 4335f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4336f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4337f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43384f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43396e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4340f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4341f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4342f9eb5b7dSStefano Zampini } 43434f3a063dSStefano Zampini if (isredundant) { 43444f3a063dSStefano Zampini KSP inner_ksp; 43454f3a063dSStefano Zampini PC inner_pc; 43464f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43474f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43484f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43494f3a063dSStefano Zampini } 4350f9eb5b7dSStefano Zampini 4351f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4352fa7f1dd8SStefano Zampini if (coarse_reuse) { 435381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4354fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43556e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4356fa7f1dd8SStefano Zampini } else { 43576e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4358fa7f1dd8SStefano Zampini } 4359c8587f34SStefano Zampini if (isbddc || isnn) { 436022bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 436170cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 436228143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 436322b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43666e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43676e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 436822b6e8a2SStefano Zampini } 436970cf5478SStefano Zampini } 437053a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 437170cf5478SStefano Zampini } else { 437222bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 437322bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 437422bc73bbSStefano Zampini } 437522bc73bbSStefano Zampini } else { 43762e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4377c8587f34SStefano Zampini } 4378c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4379c8587f34SStefano Zampini 43803301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 43815a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 43823301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 43833301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 43843301b35fSStefano Zampini } 43853301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 43863301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 43873301b35fSStefano Zampini } 43883301b35fSStefano Zampini if (pc->pmat->spd_set) { 43893301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 43903301b35fSStefano Zampini } 43916e683305SStefano Zampini /* set operators */ 43925f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43936e683305SStefano Zampini if (pcbddc->dbg_flag) { 43946e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43956e683305SStefano Zampini } 43966e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43976e683305SStefano Zampini coarse_mat = 0; 43986e683305SStefano Zampini } 43996e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4400b9b85e73SStefano Zampini #if 0 4401b9b85e73SStefano Zampini { 4402b9b85e73SStefano Zampini PetscViewer viewer; 4403b9b85e73SStefano Zampini char filename[256]; 4404b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4405b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4406b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4407b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4408b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4409b9b85e73SStefano Zampini } 4410b9b85e73SStefano Zampini #endif 4411c8587f34SStefano Zampini 4412c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4413c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4414c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 441598a51de6SStefano Zampini } 441698a51de6SStefano Zampini 441798a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 441898a51de6SStefano Zampini Vec crhs,csol; 441998a51de6SStefano Zampini PetscBool ispreonly; 442098a51de6SStefano Zampini if (CoarseNullSpace) { 4421c8587f34SStefano Zampini if (isbddc) { 4422c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4423c8587f34SStefano Zampini } else { 4424c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4425c8587f34SStefano Zampini } 4426c8587f34SStefano Zampini } 4427f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4428f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4429f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4430f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44316e683305SStefano Zampini /* hack */ 4432f347579bSStefano Zampini if (!csol) { 44332a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4434f9eb5b7dSStefano Zampini } 4435f347579bSStefano Zampini if (!crhs) { 44362a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4437f347579bSStefano Zampini } 4438cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4439cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44406e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4441c8587f34SStefano Zampini KSP check_ksp; 44422b510759SStefano Zampini KSPType check_ksp_type; 4443c8587f34SStefano Zampini PC check_pc; 44446e683305SStefano Zampini Vec check_vec,coarse_vec; 44456a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44462b510759SStefano Zampini PetscInt its; 44476e683305SStefano Zampini PetscBool compute_eigs; 44486e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44496e683305SStefano Zampini PetscInt neigs; 44508e185a42SStefano Zampini const char *prefix; 4451c8587f34SStefano Zampini 44522b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44536e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 445423ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4455f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44562b510759SStefano Zampini if (ispreonly) { 44572b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44586e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44592b510759SStefano Zampini } else { 4460cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44616e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4462c8587f34SStefano Zampini } 4463c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44646e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44656e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44666e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4467a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4468a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4469a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4470a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4471c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4472c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4473c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4474c8587f34SStefano Zampini /* create random vec */ 44756e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44766e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4477c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4478c8587f34SStefano Zampini if (CoarseNullSpace) { 4479c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4480c8587f34SStefano Zampini } 44816e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4482c8587f34SStefano Zampini /* solve coarse problem */ 44836e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4484c8587f34SStefano Zampini if (CoarseNullSpace) { 44856e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4486c8587f34SStefano Zampini } 4487cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44886e683305SStefano Zampini if (compute_eigs) { 4489854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4490854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44916e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44926e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44936e683305SStefano Zampini lambda_min = eigs_r[0]; 44946e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44956e683305SStefano Zampini if (lambda_max>lambda_min) { 4496cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4497cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4498cbcc2c2aSStefano Zampini } 4499c8587f34SStefano Zampini } 4500c8587f34SStefano Zampini } 4501cbcc2c2aSStefano Zampini 4502c8587f34SStefano Zampini /* check coarse problem residual error */ 45036e683305SStefano Zampini if (pcbddc->dbg_flag) { 45046e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45056e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45066e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4507c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45086e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45096e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4510c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4511779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45126e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45136e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45146e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45156e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45166e683305SStefano Zampini if (compute_eigs) { 45176e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4518deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4519c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45206e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45216e683305SStefano 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); 45226e683305SStefano Zampini for (i=0;i<neigs;i++) { 45236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4524c8587f34SStefano Zampini } 45256e683305SStefano Zampini } 45266e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45276e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45286e683305SStefano Zampini } 4529c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45306e683305SStefano Zampini if (compute_eigs) { 45316e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45326e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4533c8587f34SStefano Zampini } 45346e683305SStefano Zampini } 45356e683305SStefano Zampini } 4536cbcc2c2aSStefano Zampini /* print additional info */ 4537cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45386e683305SStefano Zampini /* waits until all processes reaches this point */ 45396e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4540cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4541cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4542cbcc2c2aSStefano Zampini } 4543cbcc2c2aSStefano Zampini 45442b510759SStefano Zampini /* free memory */ 4545c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4546fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4547c8587f34SStefano Zampini PetscFunctionReturn(0); 4548c8587f34SStefano Zampini } 4549674ae819SStefano Zampini 4550f34684f1SStefano Zampini #undef __FUNCT__ 4551f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4552f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4553f34684f1SStefano Zampini { 4554f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4555f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4556f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4557*dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4558*dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 455973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4560*dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4561f34684f1SStefano Zampini PetscErrorCode ierr; 4562f34684f1SStefano Zampini 4563f34684f1SStefano Zampini PetscFunctionBegin; 4564f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45650e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45660e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4567727cdba6SStefano Zampini } 4568*dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 4569*dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,subset_n,&subset);CHKERRQ(ierr); 4570*dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4571*dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4572*dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4573*dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4574*dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4575*dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 45760e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45770e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45780e6343abSStefano Zampini } 4579*dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4580*dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4581*dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4582*dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4583*dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4584f34684f1SStefano Zampini 4585f34684f1SStefano Zampini /* check numbering */ 4586f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4587f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4588*dc456d91SStefano Zampini PetscInt i; 4589b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4590f34684f1SStefano Zampini 4591f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4592f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4593f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45940fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4595f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4596f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4597727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4598f34684f1SStefano Zampini } 4599f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4600f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4601f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4602f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4603f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4604f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4605f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4606f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4607f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4608f34684f1SStefano Zampini if (array[i] == 1.0) { 4609b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4610f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4611f34684f1SStefano Zampini } 4612f34684f1SStefano Zampini } 4613b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4614f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4615f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4616f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4617f34684f1SStefano Zampini } 4618f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4619f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4620f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4621f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4622f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4623f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4624b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4625ca8b9ea9SStefano Zampini PetscInt *gidxs; 4626ca8b9ea9SStefano Zampini 4627ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4628ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4629f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4630f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4631f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4632f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46334bc2dc4bSStefano 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); 4634f34684f1SStefano Zampini } 4635f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4636ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4637f34684f1SStefano Zampini } 4638f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4639302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4640f34684f1SStefano Zampini } 46418bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4642f34684f1SStefano Zampini /* get back data */ 4643f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4644f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4645674ae819SStefano Zampini PetscFunctionReturn(0); 4646674ae819SStefano Zampini } 4647674ae819SStefano Zampini 4648e456f2a8SStefano Zampini #undef __FUNCT__ 4649e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4650a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4651e456f2a8SStefano Zampini { 4652e456f2a8SStefano Zampini IS localis_t; 4653a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4654e456f2a8SStefano Zampini PetscScalar *vals; 4655e456f2a8SStefano Zampini PetscErrorCode ierr; 4656e456f2a8SStefano Zampini 4657e456f2a8SStefano Zampini PetscFunctionBegin; 4658a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4659e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4660854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4661e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4662e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4663a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4664a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46651035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4666a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46671035eff8SStefano Zampini } 4668a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4669e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4670e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4671a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4672a7dc3881SStefano Zampini /* now compute set in local ordering */ 4673a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4674a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4675a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4676a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4677a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4678ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4679e456f2a8SStefano Zampini lsize++; 4680e456f2a8SStefano Zampini } 4681e456f2a8SStefano Zampini } 4682854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4683a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4684ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4685e456f2a8SStefano Zampini idxs[lsize++] = i; 4686e456f2a8SStefano Zampini } 4687e456f2a8SStefano Zampini } 4688a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4689a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4690e456f2a8SStefano Zampini *localis = localis_t; 4691e456f2a8SStefano Zampini PetscFunctionReturn(0); 4692e456f2a8SStefano Zampini } 4693906d46d4SStefano Zampini 4694906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4695906d46d4SStefano Zampini #undef __FUNCT__ 4696906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4697906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4698906d46d4SStefano Zampini { 4699906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4700906d46d4SStefano Zampini PetscErrorCode ierr; 4701906d46d4SStefano Zampini 4702906d46d4SStefano Zampini PetscFunctionBegin; 4703906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4704906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4705906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4706906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4707906d46d4SStefano Zampini PetscFunctionReturn(0); 4708906d46d4SStefano Zampini } 4709906d46d4SStefano Zampini 4710906d46d4SStefano Zampini #undef __FUNCT__ 4711906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4712906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4713906d46d4SStefano Zampini { 4714906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4715906d46d4SStefano Zampini PetscErrorCode ierr; 4716906d46d4SStefano Zampini 4717906d46d4SStefano Zampini PetscFunctionBegin; 4718906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4719906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4720906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4721906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4722906d46d4SStefano Zampini PetscFunctionReturn(0); 4723906d46d4SStefano Zampini } 4724b96c3477SStefano Zampini 4725b96c3477SStefano Zampini #undef __FUNCT__ 4726b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 472708122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4728b96c3477SStefano Zampini { 4729a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4730b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4731b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4732a64f4aa4SStefano Zampini Mat S_j; 4733b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4734b96c3477SStefano Zampini PetscBool free_used_adj; 4735b96c3477SStefano Zampini PetscErrorCode ierr; 4736b96c3477SStefano Zampini 4737b96c3477SStefano Zampini PetscFunctionBegin; 4738b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4739b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 474008122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4741b96c3477SStefano Zampini used_xadj = NULL; 4742b96c3477SStefano Zampini used_adjncy = NULL; 4743b96c3477SStefano Zampini } else { 474408122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 474508122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 474608122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 474708122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4748b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4749b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4750b96c3477SStefano Zampini } else { 47512fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4752b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4753b96c3477SStefano Zampini PetscInt nvtxs; 4754b96c3477SStefano Zampini 47552fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47562fffb893SStefano Zampini if (flg_row) { 4757b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4758b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4759b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4760b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47612fffb893SStefano Zampini } else { 47622fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47632fffb893SStefano Zampini used_xadj = NULL; 47642fffb893SStefano Zampini used_adjncy = NULL; 47652fffb893SStefano Zampini } 47662fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4767b96c3477SStefano Zampini } 4768b96c3477SStefano Zampini } 4769d5574798SStefano Zampini 4770d5574798SStefano Zampini /* setup sub_schurs data */ 4771a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4772a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4773a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4774a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4775862806e4SStefano 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); 4776a64f4aa4SStefano Zampini } else { 47776816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 47785feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 47795feab87aSStefano Zampini PetscInt n_vertices; 47805feab87aSStefano Zampini 47815feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 47822034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 47835feab87aSStefano Zampini } 4784862806e4SStefano 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); 4785a64f4aa4SStefano Zampini } 4786a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4787b96c3477SStefano Zampini 4788b96c3477SStefano Zampini /* free adjacency */ 4789b96c3477SStefano Zampini if (free_used_adj) { 4790b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4791b96c3477SStefano Zampini } 4792b96c3477SStefano Zampini PetscFunctionReturn(0); 4793b96c3477SStefano Zampini } 4794b96c3477SStefano Zampini 4795b96c3477SStefano Zampini #undef __FUNCT__ 4796b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 479708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4798b96c3477SStefano Zampini { 4799b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4800b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4801b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4802b96c3477SStefano Zampini PCBDDCGraph graph; 4803b96c3477SStefano Zampini PetscErrorCode ierr; 4804b96c3477SStefano Zampini 4805b96c3477SStefano Zampini PetscFunctionBegin; 4806b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 480708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48083301b35fSStefano Zampini IS verticesIS,verticescomm; 48093301b35fSStefano Zampini PetscInt vsize,*idxs; 4810b96c3477SStefano Zampini 4811b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48123301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48133301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48143301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48153301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48163301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4817b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48187fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48193301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48203301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4821b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4822b96c3477SStefano Zampini /* 4823b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4824b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4825b96c3477SStefano Zampini } 4826b96c3477SStefano Zampini */ 4827b96c3477SStefano Zampini } else { 4828b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4829b96c3477SStefano Zampini } 4830b96c3477SStefano Zampini 4831b96c3477SStefano Zampini /* sub_schurs init */ 4832a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4833a64f4aa4SStefano Zampini 4834b96c3477SStefano Zampini /* free graph struct */ 483508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4836b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4837b96c3477SStefano Zampini } 4838b96c3477SStefano Zampini PetscFunctionReturn(0); 4839b96c3477SStefano Zampini } 4840