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) { 269ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 27008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 27108122e43SStefano Zampini if (eigs[j] == 0.0) { 272ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 27308122e43SStefano Zampini } else { 274ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 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++) { 290ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 2919552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 292ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 293ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 294ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 295ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 296ac47001eSStefano Zampini #else 297ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 298ac47001eSStefano Zampini #endif 2999552c7c7SStefano Zampini } 3009552c7c7SStefano Zampini } 301fd14bc51SStefano Zampini } 30208122e43SStefano Zampini #if 0 3039162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 30408122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 30508122e43SStefano Zampini PetscScalar norm; 30608122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 3079162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 3089162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 30908122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 31008122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 31108122e43SStefano Zampini } else { 31208122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 31308122e43SStefano Zampini } 3149162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 315b1b3d7a2SStefano Zampini } 316b1b3d7a2SStefano Zampini #endif 3179162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3189162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3199162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3209162d606SStefano Zampini cum++; 32108122e43SStefano Zampini } 32208122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 32308122e43SStefano Zampini /* shift for next computation */ 32408122e43SStefano Zampini cumarray += subset_size*subset_size; 32508122e43SStefano Zampini } 326fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 327fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 328fd14bc51SStefano Zampini } 32908122e43SStefano Zampini 33008122e43SStefano Zampini if (mss) { 33108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 33208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 333f6f667cfSStefano Zampini /* destroy matrices (junk) */ 334f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 335f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 33608122e43SStefano Zampini } 337f6f667cfSStefano Zampini if (allocated_S_St) { 338f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 339f6f667cfSStefano Zampini } 340f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 34108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 34208122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 34308122e43SStefano Zampini #endif 34408122e43SStefano Zampini if (pcbddc->dbg_flag) { 3451b968477SStefano Zampini PetscInt maxneigs_r; 34608122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3479b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 34808122e43SStefano Zampini } 34908122e43SStefano Zampini PetscFunctionReturn(0); 35008122e43SStefano Zampini } 351b1b3d7a2SStefano Zampini 352674ae819SStefano Zampini #undef __FUNCT__ 353c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 354c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 355c8587f34SStefano Zampini { 356c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3578629588bSStefano Zampini PetscScalar *coarse_submat_vals; 358c8587f34SStefano Zampini PetscErrorCode ierr; 359c8587f34SStefano Zampini 360c8587f34SStefano Zampini PetscFunctionBegin; 361f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3625e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 363c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 364c8587f34SStefano Zampini 365684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3660fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 367684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 368c8587f34SStefano Zampini 369c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 370b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 371c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 372c8587f34SStefano Zampini } 373c8587f34SStefano Zampini 3748629588bSStefano Zampini /* 3758629588bSStefano Zampini Setup local correction and local part of coarse basis. 3768629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3778629588bSStefano Zampini */ 37847f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3798629588bSStefano Zampini 3808629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3818629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3828629588bSStefano Zampini 3838629588bSStefano Zampini /* free */ 3848629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 385c8587f34SStefano Zampini PetscFunctionReturn(0); 386c8587f34SStefano Zampini } 387c8587f34SStefano Zampini 388c8587f34SStefano Zampini #undef __FUNCT__ 389674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 390674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 391674ae819SStefano Zampini { 392674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 393674ae819SStefano Zampini PetscErrorCode ierr; 394674ae819SStefano Zampini 395674ae819SStefano Zampini PetscFunctionBegin; 396674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 397674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 398674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 399674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 400785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 401674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 402f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 403f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 404785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 40563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 40663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 407674ae819SStefano Zampini PetscFunctionReturn(0); 408674ae819SStefano Zampini } 409674ae819SStefano Zampini 410674ae819SStefano Zampini #undef __FUNCT__ 411674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 412674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 413674ae819SStefano Zampini { 414674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 415674ae819SStefano Zampini PetscErrorCode ierr; 416674ae819SStefano Zampini 417674ae819SStefano Zampini PetscFunctionBegin; 418b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 419674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 420674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 421674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 422b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 423674ae819SStefano Zampini PetscFunctionReturn(0); 424674ae819SStefano Zampini } 425674ae819SStefano Zampini 426674ae819SStefano Zampini #undef __FUNCT__ 427674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 428674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 429674ae819SStefano Zampini { 430674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 43106656605SStefano Zampini PetscScalar *array; 432674ae819SStefano Zampini PetscErrorCode ierr; 433674ae819SStefano Zampini 434674ae819SStefano Zampini PetscFunctionBegin; 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 43658da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 43706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 43806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 43958da7f69SStefano Zampini } 440674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 441674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 44215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 44315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 444674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 445674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 446674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 44706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 448674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 449674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4508ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 451674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 452674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 453674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 454f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 455f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 456f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 457f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 458727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4590e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 460f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 46170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4626e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 463674ae819SStefano Zampini PetscFunctionReturn(0); 464674ae819SStefano Zampini } 465674ae819SStefano Zampini 466674ae819SStefano Zampini #undef __FUNCT__ 467f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 468f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4696bfb1811SStefano Zampini { 4706bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4716bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4726bfb1811SStefano Zampini VecType impVecType; 473e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4746bfb1811SStefano Zampini PetscErrorCode ierr; 4756bfb1811SStefano Zampini 4766bfb1811SStefano Zampini PetscFunctionBegin; 477f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 478f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 479f4ddd8eeSStefano Zampini } 480e7b262bdSStefano Zampini /* get sizes */ 481b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 482b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4836bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 484e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 485e7b262bdSStefano Zampini /* R nodes */ 486e7b262bdSStefano Zampini old_size = -1; 487e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 488e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 489e7b262bdSStefano Zampini } 490e7b262bdSStefano Zampini if (n_R != old_size) { 491e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 492e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4936bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4946bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4956bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4966bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 497e7b262bdSStefano Zampini } 498e7b262bdSStefano Zampini /* local primal dofs */ 499e7b262bdSStefano Zampini old_size = -1; 500e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 501e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 502e7b262bdSStefano Zampini } 503e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 504e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 506e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 5076bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 508e7b262bdSStefano Zampini } 509e7b262bdSStefano Zampini /* local explicit constraints */ 510e7b262bdSStefano Zampini old_size = -1; 511e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 512e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 513e7b262bdSStefano Zampini } 514e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 515e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 51683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 51783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 51883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 51983b7ccabSStefano Zampini } 5206bfb1811SStefano Zampini PetscFunctionReturn(0); 5216bfb1811SStefano Zampini } 5226bfb1811SStefano Zampini 5236bfb1811SStefano Zampini #undef __FUNCT__ 52447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 52547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 52688ebb749SStefano Zampini { 52725084f0cSStefano Zampini PetscErrorCode ierr; 52825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 52988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 53088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 531d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 53225084f0cSStefano Zampini /* submatrices of local problem */ 53380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 53406656605SStefano Zampini /* submatrices of local coarse problem */ 53506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 53625084f0cSStefano Zampini /* working matrices */ 53706656605SStefano Zampini Mat C_CR; 53825084f0cSStefano Zampini /* additional working stuff */ 53906656605SStefano Zampini PC pc_R; 54006656605SStefano Zampini Mat F; 54106656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 54206656605SStefano Zampini 54325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 54406656605SStefano Zampini PetscScalar *work; 54506656605SStefano Zampini PetscInt *idx_V_B; 54606656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 54706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 548b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 54945a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 55088ebb749SStefano Zampini MatType impMatType; 55125084f0cSStefano Zampini /* some shortcuts to scalars */ 55206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 55388ebb749SStefano Zampini 55488ebb749SStefano Zampini PetscFunctionBegin; 555b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 55688ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 55788ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 558b371cd4fSStefano Zampini n_B = pcis->n_B; 559b371cd4fSStefano Zampini n_D = pcis->n - n_B; 56088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 56188ebb749SStefano Zampini 56288ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 56388ebb749SStefano Zampini impMatType = MATSEQDENSE; 56488ebb749SStefano Zampini 56588ebb749SStefano Zampini /* vertices in boundary numbering */ 566785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5670e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 56888ebb749SStefano Zampini if (i != n_vertices) { 56922d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 57088ebb749SStefano Zampini } 57188ebb749SStefano Zampini 57206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 57306656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 57406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 57506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 57606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 57706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 57806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 57906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 58006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 58106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 58206656605SStefano Zampini 58306656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 58406656605SStefano Zampini /* allocate workspace */ 58506656605SStefano Zampini n = 0; 58606656605SStefano Zampini if (n_constraints) { 58706656605SStefano Zampini n += n_R*n_constraints; 58806656605SStefano Zampini } 58906656605SStefano Zampini if (n_vertices) { 59006656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 59180677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 59206656605SStefano Zampini } 5933301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 59406656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 59506656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 59606656605SStefano Zampini } 59706656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 59806656605SStefano Zampini 59906656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 60006656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 60106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 60206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 60306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 60406656605SStefano Zampini if (isLU || isILU || isCHOL) { 60506656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 606d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 607d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 608d62866d3SStefano Zampini MatFactorType type; 609d62866d3SStefano Zampini 6106816873aSStefano Zampini F = reuse_mumps->F; 6116816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 612d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 61306656605SStefano Zampini } else { 61406656605SStefano Zampini F = NULL; 61506656605SStefano Zampini } 61606656605SStefano Zampini 61788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 61888ebb749SStefano Zampini if (n_constraints) { 61906656605SStefano Zampini Mat M1,M2,M3; 62080677318SStefano Zampini Mat auxmat; 62106656605SStefano Zampini IS is_aux; 62280677318SStefano Zampini PetscScalar *array,*array2; 62306656605SStefano Zampini 624f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 62580677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 62688ebb749SStefano Zampini 62725084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 62825084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6298ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 63080677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 63188ebb749SStefano Zampini 63280677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 63380677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 63406656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 63588ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 63606656605SStefano Zampini const PetscScalar *row_cmat_values; 63706656605SStefano Zampini const PetscInt *row_cmat_indices; 63806656605SStefano Zampini PetscInt size_of_constraint,j; 63988ebb749SStefano Zampini 64006656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64106656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 64206656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 64306656605SStefano Zampini } 64406656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 64506656605SStefano Zampini } 64680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 64706656605SStefano Zampini if (F) { 64806656605SStefano Zampini Mat B; 64906656605SStefano Zampini 65006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 65180677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 65206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 65306656605SStefano Zampini } else { 65480677318SStefano Zampini PetscScalar *marr; 65580677318SStefano Zampini 65680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 65806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 66006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 66106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 66306656605SStefano Zampini } 66480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 66506656605SStefano Zampini } 66680677318SStefano Zampini if (!pcbddc->switch_static) { 66780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 66880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 67180677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 67280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 67380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 67580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 67680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 67780677318SStefano Zampini } 67880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 67980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 68080677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68180677318SStefano Zampini } else { 68280677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 68380677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 68425084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 68580677318SStefano Zampini } 68680677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 68780677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 68880677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 68906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 69006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 69180677318SStefano Zampini if (isCHOL) { 69280677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 69380677318SStefano Zampini } else { 69425084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 69580677318SStefano Zampini } 69680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 69706656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 69825084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 69925084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 70025084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 70180677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 70280677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 70380677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 70406656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 70506656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 70688ebb749SStefano Zampini } 70788ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 70888ebb749SStefano Zampini if (n_vertices) { 70906656605SStefano Zampini IS is_aux; 7103a50541eSStefano Zampini 7116816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7126816873aSStefano Zampini IS tis; 7136816873aSStefano Zampini 7146816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7156816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7166816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7176816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7186816873aSStefano Zampini } else { 7193a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7206816873aSStefano Zampini } 7219577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7229577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72304708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 72425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 72588ebb749SStefano Zampini } 72688ebb749SStefano Zampini 72788ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 728f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 72906656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 73006656605SStefano Zampini if (pcbddc->coarse_phi_D) { 73106656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 73206656605SStefano Zampini } 733f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 73406656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 73506656605SStefano Zampini PetscScalar *marray; 73606656605SStefano Zampini 73706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 73806656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 739f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 740f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 741f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 742f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 743f4ddd8eeSStefano Zampini } 744f4ddd8eeSStefano Zampini } 74506656605SStefano Zampini 746f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 74706656605SStefano Zampini PetscScalar *marray; 74888ebb749SStefano Zampini 74906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7508eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 75106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 75288ebb749SStefano Zampini } 7533301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 75406656605SStefano Zampini n *= 2; 755f4ddd8eeSStefano Zampini } 75606656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 75706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 75806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 75906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 76106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76288ebb749SStefano Zampini } 7633301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 76406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7651b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 76706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 76888ebb749SStefano Zampini } 76988ebb749SStefano Zampini } else { 770c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 771c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7721b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 773c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 774c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 775c0553b1fSStefano Zampini } 77688ebb749SStefano Zampini } 77706656605SStefano Zampini } 77806656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 77906656605SStefano Zampini /* vertices */ 78006656605SStefano Zampini if (n_vertices) { 78116f15bc4SStefano Zampini 78204708bb6SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 78304708bb6SStefano Zampini 78416f15bc4SStefano Zampini if (n_R) { 78506656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 78606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 78716f15bc4SStefano Zampini PetscScalar *x,*y; 78804708bb6SStefano Zampini PetscBool isseqaij; 78906656605SStefano Zampini 79021eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 79106656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 79204708bb6SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 7936816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 79406656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 79506656605SStefano Zampini } else { 79606656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 79706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 79806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 79906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 80006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 80106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 80206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 80306656605SStefano Zampini } 80406656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 80506656605SStefano Zampini } 80680677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 80706656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 80806656605SStefano Zampini if (n_constraints) { 80906656605SStefano Zampini Mat B; 81080677318SStefano Zampini 811b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 81280677318SStefano Zampini for (i=0;i<n_vertices;i++) { 81380677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 81480677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 81580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 81780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 81880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 81980677318SStefano Zampini } 82080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82180677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 82280677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 82480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 82506656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 82606656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 82706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 82806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 82906656605SStefano Zampini } 83004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 83104708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 83204708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 83304708bb6SStefano Zampini } 83406656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 83580677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 83606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 83706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 83806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 83906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 84006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 84206656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84306656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 84416f15bc4SStefano Zampini } else { 84516f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 84616f15bc4SStefano Zampini } 84721eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 84806656605SStefano Zampini /* coarse basis functions */ 84906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85016f15bc4SStefano Zampini PetscScalar *y; 85116f15bc4SStefano Zampini 85206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 85306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 85506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 85706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 85806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 85906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86006656605SStefano Zampini 86106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 86206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 86806656605SStefano Zampini } 86906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87006656605SStefano Zampini } 87104708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 87204708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 87306656605SStefano Zampini } 87406656605SStefano Zampini 87506656605SStefano Zampini if (n_constraints) { 87606656605SStefano Zampini Mat B; 87706656605SStefano Zampini 87806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 87906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 88106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88206656605SStefano Zampini if (n_vertices) { 88380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 88480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 88580677318SStefano Zampini } else { 88680677318SStefano Zampini Mat S_VCt; 88780677318SStefano Zampini 88880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 88980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 89180677318SStefano Zampini } 89206656605SStefano Zampini } 89306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 89406656605SStefano Zampini /* coarse basis functions */ 89506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 89606656605SStefano Zampini PetscScalar *y; 89706656605SStefano Zampini 89806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 89906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 90106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 90506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 90606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 90806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91206656605SStefano Zampini } 91306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 91406656605SStefano Zampini } 91506656605SStefano Zampini } 91680677318SStefano Zampini if (n_constraints) { 91780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 91880677318SStefano Zampini } 91906656605SStefano Zampini 92006656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 9213301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 92206656605SStefano Zampini 92306656605SStefano Zampini if (n_constraints) { 92416f15bc4SStefano Zampini Mat S_CCT,B_C; 92506656605SStefano Zampini 92680677318SStefano Zampini /* this is a lazy thing */ 92780677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 92806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 92906656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 93006656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 93116f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 93206656605SStefano Zampini if (n_vertices) { 93316f15bc4SStefano Zampini Mat B_V,S_VCT; 93406656605SStefano Zampini 93506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 93606656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 93706656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 93806656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 93916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 94006656605SStefano Zampini } 94106656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 94280677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94380677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 94406656605SStefano Zampini } 94516f15bc4SStefano Zampini if (n_vertices && n_R) { 94606656605SStefano Zampini Mat A_VRT; 94780677318SStefano Zampini PetscScalar *marray; 94806656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 94906656605SStefano Zampini 95080677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 95180677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 95280677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95306656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 95480677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 95580677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 95616f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 95706656605SStefano Zampini } 95806656605SStefano Zampini 95906656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 96006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96206656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96306656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 96406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 96506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 96606656605SStefano Zampini } 96706656605SStefano Zampini } else { 96806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96906656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97406656605SStefano Zampini } 97506656605SStefano Zampini } 97606656605SStefano Zampini /* coarse basis functions */ 97706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97806656605SStefano Zampini PetscScalar *y; 97906656605SStefano Zampini 98006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 98106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 98506656605SStefano Zampini if (i<n_vertices) { 98606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 98706656605SStefano Zampini } 98806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 99006656605SStefano Zampini 99106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 99206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 99406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 99706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99806656605SStefano Zampini } 99906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 100006656605SStefano Zampini } 100106656605SStefano Zampini } 1002d62866d3SStefano Zampini /* free memory */ 100388ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 100406656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 100506656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 100606656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 100706656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1008d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1009d62866d3SStefano Zampini if (n_vertices) { 1010d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1011d62866d3SStefano Zampini } 1012d62866d3SStefano Zampini if (n_constraints) { 1013d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1014d62866d3SStefano Zampini } 101588ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 101688ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 101788ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 101825084f0cSStefano Zampini if (pcbddc->dbg_flag) { 101988ebb749SStefano Zampini Mat coarse_sub_mat; 102025084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 102188ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 102288ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102388ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10248bec7fa6SStefano Zampini Mat C_B,CPHI; 10258bec7fa6SStefano Zampini IS is_dummy; 10268bec7fa6SStefano Zampini Vec mones; 102788ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 102888ebb749SStefano Zampini PetscReal real_value; 102988ebb749SStefano Zampini 103088ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 103188ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 103288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 103488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 103588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1036c0553b1fSStefano Zampini if (unsymmetric_check) { 103788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 103988ebb749SStefano Zampini } 104088ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 104188ebb749SStefano Zampini 104225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 10433301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 104425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1045c0553b1fSStefano Zampini if (unsymmetric_check) { 104688ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 104788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 104888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 104988ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini } else { 105988ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106788ebb749SStefano Zampini } 106888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107188ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 107281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10738bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10740fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 107506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10768bec7fa6SStefano Zampini 10778bec7fa6SStefano Zampini /* check constraints */ 10788bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10798bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10808bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10818bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10838bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10848bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1085bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1086bdae7319SStefano Zampini if (unsymmetric_check) { 1087bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 1088bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 1089bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 1090bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 1091bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 1092bdae7319SStefano Zampini } 10938bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10948bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10958bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10968bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10978bec7fa6SStefano Zampini 109825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 109988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 110088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 110188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 110288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 110388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 110488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 110888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1109c0553b1fSStefano Zampini if (unsymmetric_check) { 111088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 111188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 111288ebb749SStefano Zampini } 111388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 111488ebb749SStefano Zampini } 11158629588bSStefano Zampini /* get back data */ 11168629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 111788ebb749SStefano Zampini PetscFunctionReturn(0); 111888ebb749SStefano Zampini } 111988ebb749SStefano Zampini 112088ebb749SStefano Zampini #undef __FUNCT__ 1121d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1122d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1123d65f70fdSStefano Zampini { 1124d65f70fdSStefano Zampini Mat *work_mat; 1125d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1126d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1127d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1128d65f70fdSStefano Zampini PetscErrorCode ierr; 1129d65f70fdSStefano Zampini 1130d65f70fdSStefano Zampini PetscFunctionBegin; 1131d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1134d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1135d65f70fdSStefano Zampini 1136d65f70fdSStefano Zampini if (!rsorted) { 1137d65f70fdSStefano Zampini const PetscInt *idxs; 1138d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1139d65f70fdSStefano Zampini 1140d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1142d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1143d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1144d65f70fdSStefano Zampini } 1145d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1147d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1148d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1149d65f70fdSStefano Zampini } 1150d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1152d65f70fdSStefano Zampini } else { 1153d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1154d65f70fdSStefano Zampini isrow_s = isrow; 1155d65f70fdSStefano Zampini } 1156d65f70fdSStefano Zampini 1157d65f70fdSStefano Zampini if (!csorted) { 1158d65f70fdSStefano Zampini if (isrow == iscol) { 1159d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1160d65f70fdSStefano Zampini iscol_s = isrow_s; 1161d65f70fdSStefano Zampini } else { 1162d65f70fdSStefano Zampini const PetscInt *idxs; 1163d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1164d65f70fdSStefano Zampini 1165d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1167d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1168d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1169d65f70fdSStefano Zampini } 1170d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1172d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1173d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1174d65f70fdSStefano Zampini } 1175d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1176d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1177d65f70fdSStefano Zampini } 1178d65f70fdSStefano Zampini } else { 1179d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1180d65f70fdSStefano Zampini iscol_s = iscol; 1181d65f70fdSStefano Zampini } 1182d65f70fdSStefano Zampini 1183d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1184d65f70fdSStefano Zampini 1185d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1186d65f70fdSStefano Zampini Mat new_mat; 1187d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1188d65f70fdSStefano Zampini 1189d65f70fdSStefano Zampini if (!rsorted) { 1190d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1191d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1192d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1193d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1194d65f70fdSStefano Zampini } 1195d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1197d65f70fdSStefano Zampini } else { 1198d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1199d65f70fdSStefano Zampini } 1200d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1201d65f70fdSStefano Zampini 1202d65f70fdSStefano Zampini if (!csorted) { 1203d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1204d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1205d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1206d65f70fdSStefano Zampini } else { 1207d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1208d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1209d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1210d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1211d65f70fdSStefano Zampini } 1212d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1214d65f70fdSStefano Zampini } 1215d65f70fdSStefano Zampini } else { 1216d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1217d65f70fdSStefano Zampini } 1218d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1219d65f70fdSStefano Zampini 1220d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1221d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1222d65f70fdSStefano Zampini work_mat[0] = new_mat; 1223d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1225d65f70fdSStefano Zampini } 1226d65f70fdSStefano Zampini 1227d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1228d65f70fdSStefano Zampini *B = work_mat[0]; 1229d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1230d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1231d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1232d65f70fdSStefano Zampini PetscFunctionReturn(0); 1233d65f70fdSStefano Zampini } 1234d65f70fdSStefano Zampini 1235d65f70fdSStefano Zampini #undef __FUNCT__ 12365e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12375e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1238aa0d41d4SStefano Zampini { 1239aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12405e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1241d65f70fdSStefano Zampini Mat new_mat; 12425e8657edSStefano Zampini IS is_local,is_global; 1243d65f70fdSStefano Zampini PetscInt local_size; 1244d65f70fdSStefano Zampini PetscBool isseqaij; 1245aa0d41d4SStefano Zampini PetscErrorCode ierr; 1246aa0d41d4SStefano Zampini 1247aa0d41d4SStefano Zampini PetscFunctionBegin; 1248aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12495e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12505e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1251906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1252906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1253d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1254906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1255906d46d4SStefano Zampini 1256906d46d4SStefano Zampini /* check */ 1257906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1258906d46d4SStefano Zampini Vec x,x_change; 1259906d46d4SStefano Zampini PetscReal error; 1260906d46d4SStefano Zampini 12615e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1262906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12635e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12645e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12655e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1266d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12675e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12685e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1269906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1270906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1271906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1272906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1273906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1274906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1275906d46d4SStefano Zampini } 1276906d46d4SStefano Zampini 127722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12789b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 127922d5777bSStefano Zampini if (isseqaij) { 1280d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1281aa0d41d4SStefano Zampini } else { 1282aa0d41d4SStefano Zampini Mat work_mat; 1283aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1284d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1285aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1286aa0d41d4SStefano Zampini } 12873301b35fSStefano Zampini if (matis->A->symmetric_set) { 12883301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 1289e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 12903301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 1291e496cd5dSStefano Zampini #endif 12923301b35fSStefano Zampini } 129345a1bb75SStefano Zampini /* 129445a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1295d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 129645a1bb75SStefano Zampini */ 1297d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1298aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1299aa0d41d4SStefano Zampini } 1300aa0d41d4SStefano Zampini 1301aa0d41d4SStefano Zampini #undef __FUNCT__ 1302a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13038ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1304a64d13efSStefano Zampini { 1305a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1306a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1307d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 130853892102SStefano Zampini PetscInt *idx_R_local=NULL; 13093a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13103a50541eSStefano Zampini PetscInt vbs,bs; 13116816873aSStefano Zampini PetscBT bitmask=NULL; 1312a64d13efSStefano Zampini PetscErrorCode ierr; 1313a64d13efSStefano Zampini 1314a64d13efSStefano Zampini PetscFunctionBegin; 1315b23d619eSStefano Zampini /* 1316b23d619eSStefano Zampini No need to setup local scatters if 1317b23d619eSStefano Zampini - primal space is unchanged 1318b23d619eSStefano Zampini AND 1319b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1320b23d619eSStefano Zampini AND 1321b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1322b23d619eSStefano Zampini */ 1323b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1324f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1325f4ddd8eeSStefano Zampini } 1326f4ddd8eeSStefano Zampini /* destroy old objects */ 1327f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1328f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1329f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1330a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1331b371cd4fSStefano Zampini n_B = pcis->n_B; 1332b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1333b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13346816873aSStefano Zampini 13356816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13366816873aSStefano Zampini 133753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13386816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 133953892102SStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13404641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13414641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13420e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13434641a718SStefano Zampini } 13443a50541eSStefano Zampini 1345a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13464641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13476816873aSStefano Zampini idx_R_local[n_R++] = i; 1348a64d13efSStefano Zampini } 1349a64d13efSStefano Zampini } 135053892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13516816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13526816873aSStefano Zampini 135353892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 135453892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13556816873aSStefano Zampini } 13563a50541eSStefano Zampini 13573a50541eSStefano Zampini /* Block code */ 13583a50541eSStefano Zampini vbs = 1; 13593a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13603a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13613a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13623a50541eSStefano Zampini PetscInt *vary; 1363d3df7717SStefano Zampini if (!sub_schurs->reuse_mumps) { 1364785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13653a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 1366d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1367d3df7717SStefano 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 */ 13680e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1369d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 13703a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13713a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13723a50541eSStefano Zampini break; 13733a50541eSStefano Zampini } 13743a50541eSStefano Zampini } 1375d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 1376d3df7717SStefano Zampini } else { 1377d3df7717SStefano Zampini /* Verify directly the R set */ 1378d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 1379d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 1380d3df7717SStefano Zampini for (j=1; j<bs; j++) { 1381d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 1382d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 1383d3df7717SStefano Zampini break; 1384d3df7717SStefano Zampini } 1385d3df7717SStefano Zampini } 1386d3df7717SStefano Zampini } 1387d3df7717SStefano Zampini } 13883a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13893a50541eSStefano Zampini vbs = bs; 13903a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13913a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13923a50541eSStefano Zampini } 13933a50541eSStefano Zampini } 13943a50541eSStefano Zampini } 13953a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 139653892102SStefano Zampini if (sub_schurs->reuse_mumps) { 139753892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 139853892102SStefano Zampini 139953892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 140053892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 140153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 140253892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 140353892102SStefano Zampini } else { 14043a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 140553892102SStefano Zampini } 1406a64d13efSStefano Zampini 1407a64d13efSStefano Zampini /* print some info if requested */ 1408a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1409a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1410a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 14110fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1412a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1413a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 14143a50541eSStefano 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); 1415a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1416a64d13efSStefano Zampini } 1417a64d13efSStefano Zampini 1418a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14196816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14206816873aSStefano Zampini IS is_aux1,is_aux2; 14216816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14226816873aSStefano Zampini 14233a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1424854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1425854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1426a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14274641a718SStefano Zampini for (i=0; i<n_D; i++) { 14284641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14294641a718SStefano Zampini } 1430a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1431a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14324641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14334641a718SStefano Zampini aux_array1[j++] = i; 1434a64d13efSStefano Zampini } 1435a64d13efSStefano Zampini } 1436a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1437a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1438a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14394641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14404641a718SStefano Zampini aux_array2[j++] = i; 1441a64d13efSStefano Zampini } 1442a64d13efSStefano Zampini } 1443a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1444a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1445a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1446a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1447a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1448a64d13efSStefano Zampini 14498eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1450785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1451a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14524641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14534641a718SStefano Zampini aux_array1[j++] = i; 1454a64d13efSStefano Zampini } 1455a64d13efSStefano Zampini } 1456a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1457a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1458a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1459a64d13efSStefano Zampini } 14604641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14613a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1462d62866d3SStefano Zampini } else { 146353892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14646816873aSStefano Zampini IS tis; 14656816873aSStefano Zampini PetscInt schur_size; 14666816873aSStefano Zampini 146753892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14686816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 146953892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14706816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14716816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14726816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14736816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14746816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1475d62866d3SStefano Zampini } 1476d62866d3SStefano Zampini } 1477a64d13efSStefano Zampini PetscFunctionReturn(0); 1478a64d13efSStefano Zampini } 1479a64d13efSStefano Zampini 1480304d26faSStefano Zampini 1481304d26faSStefano Zampini #undef __FUNCT__ 1482304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1483684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1484304d26faSStefano Zampini { 1485304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1486304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1487304d26faSStefano Zampini PC pc_temp; 1488304d26faSStefano Zampini Mat A_RR; 1489f4ddd8eeSStefano Zampini MatReuse reuse; 1490304d26faSStefano Zampini PetscScalar m_one = -1.0; 1491304d26faSStefano Zampini PetscReal value; 149204708bb6SStefano Zampini PetscInt n_D,n_R; 14939577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1494304d26faSStefano Zampini PetscErrorCode ierr; 1495e604994aSStefano Zampini /* prefixes stuff */ 1496312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1497e604994aSStefano Zampini size_t len; 1498304d26faSStefano Zampini 1499304d26faSStefano Zampini PetscFunctionBegin; 1500304d26faSStefano Zampini 1501e604994aSStefano Zampini /* compute prefixes */ 1502e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1503e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1504e604994aSStefano Zampini if (!pcbddc->current_level) { 1505e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1506e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1507e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1508e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1509e604994aSStefano Zampini } else { 1510e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1511312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1512e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1513e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1514312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1515312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 151634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 151734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1518e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1519e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1520e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1521e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1522e604994aSStefano Zampini } 1523e604994aSStefano Zampini 1524304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1525684f6988SStefano Zampini if (dirichlet) { 1526d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 15273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 15283301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 1529964fefecSStefano Zampini } 1530ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1531964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1532304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1533304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1534304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1535304d26faSStefano Zampini /* default */ 1536304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1537e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15389577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1539304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15409577ea80SStefano Zampini if (issbaij) { 15419577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15429577ea80SStefano Zampini } else { 1543304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15449577ea80SStefano Zampini } 1545304d26faSStefano Zampini /* Allow user's customization */ 1546304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1547304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1548304d26faSStefano Zampini } 1549d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1550d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1551d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1552d62866d3SStefano Zampini 1553d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1554d5574798SStefano Zampini } 1555304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1556304d26faSStefano Zampini if (!n_D) { 1557304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1558304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1559304d26faSStefano Zampini } 1560304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1561304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1562304d26faSStefano Zampini /* set ksp_D into pcis data */ 1563304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1564304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1565304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1566684f6988SStefano Zampini } 1567304d26faSStefano Zampini 1568304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1569684f6988SStefano Zampini A_RR = 0; 1570684f6988SStefano Zampini if (neumann) { 1571d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 157204708bb6SStefano Zampini PetscInt ibs,mbs; 157304708bb6SStefano Zampini PetscBool issbaij; 157404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1575f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15768ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1577f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1578f4ddd8eeSStefano Zampini PetscInt nn_R; 157981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1580f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1581f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1582f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1583f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1584f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1585f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1586f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1587727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1588f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1589f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1590f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1591f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1592f4ddd8eeSStefano Zampini } 1593f4ddd8eeSStefano Zampini } 1594f4ddd8eeSStefano Zampini /* last check */ 1595d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1596f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1597f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1598f4ddd8eeSStefano Zampini } 1599f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1600f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1601f4ddd8eeSStefano Zampini } 16026816873aSStefano Zampini /* extract A_RR */ 16036816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 16046816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 160504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 160604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 160704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 160804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 160904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16106816873aSStefano Zampini } else { 161104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 16126816873aSStefano Zampini } 161304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 161404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 161504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 161604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 161704708bb6SStefano Zampini } else { 161804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 161904708bb6SStefano Zampini } 162004708bb6SStefano Zampini } 162104708bb6SStefano Zampini if (!sub_schurs->reuse_mumps) { 162204708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 16233301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 16243301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 16256816873aSStefano Zampini } 16266816873aSStefano Zampini } else { 16276816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 16286816873aSStefano Zampini 16296816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16306816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16316816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16326816873aSStefano Zampini } 1633f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1634304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1635304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1636304d26faSStefano Zampini /* default */ 1637304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1638e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1639304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16409577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16419577ea80SStefano Zampini if (issbaij) { 16429577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16439577ea80SStefano Zampini } else { 1644304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16459577ea80SStefano Zampini } 1646304d26faSStefano Zampini /* Allow user's customization */ 1647304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1648304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1649304d26faSStefano Zampini } 1650d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1651304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1652304d26faSStefano Zampini if (!n_R) { 1653304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1654304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1655304d26faSStefano Zampini } 1656d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1657d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1658d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1659d62866d3SStefano Zampini 1660d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1661d62866d3SStefano Zampini } 1662304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1663304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1664684f6988SStefano Zampini } 16656816873aSStefano Zampini /* free Neumann problem's matrix */ 16666816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1667304d26faSStefano Zampini 1668304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16690fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1670684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1671684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1672684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1673684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1674684f6988SStefano Zampini } 1675684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16760fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16770fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16780fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16790fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16800fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1681304d26faSStefano Zampini /* need to be adapted? */ 1682b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1683b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1684b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1685304d26faSStefano Zampini /* print info */ 1686304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1687e604994aSStefano 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); 1688304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1689304d26faSStefano Zampini } 1690b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16918ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1692304d26faSStefano Zampini } 1693684f6988SStefano Zampini } 1694684f6988SStefano Zampini if (neumann) { /* Neumann */ 16956816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16960fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16970fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16980fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16990fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 17000fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1701304d26faSStefano Zampini /* need to be adapted? */ 1702b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1703b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1704304d26faSStefano Zampini /* print info */ 1705304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1706e604994aSStefano 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); 1707304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1708304d26faSStefano Zampini } 1709b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 17108ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1711304d26faSStefano Zampini } 17120fccc4e9SStefano Zampini } 1713684f6988SStefano Zampini } 1714304d26faSStefano Zampini PetscFunctionReturn(0); 1715304d26faSStefano Zampini } 1716304d26faSStefano Zampini 1717304d26faSStefano Zampini #undef __FUNCT__ 1718ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 171980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1720674ae819SStefano Zampini { 1721674ae819SStefano Zampini PetscErrorCode ierr; 1722674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1723be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1724674ae819SStefano Zampini 1725674ae819SStefano Zampini PetscFunctionBegin; 1726be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1728be83ff47SStefano Zampini } 172980677318SStefano Zampini if (!pcbddc->switch_static) { 173080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 173180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 173280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 173320c7b377SStefano Zampini } 1734be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 173580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 173720c7b377SStefano Zampini } else { 1738be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1739be83ff47SStefano Zampini 174053892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174153892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1742be83ff47SStefano Zampini } 1743be83ff47SStefano Zampini } else { 174480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 174880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 174980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 175080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 175180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 175320c7b377SStefano Zampini } 1754674ae819SStefano Zampini } 1755be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 175680677318SStefano Zampini if (applytranspose) { 175780677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 175880677318SStefano Zampini } else { 175980677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176080677318SStefano Zampini } 176153892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1762be83ff47SStefano Zampini } else { 1763be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1764be83ff47SStefano Zampini 1765be83ff47SStefano Zampini if (applytranspose) { 176653892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1767be83ff47SStefano Zampini } else { 176853892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1769be83ff47SStefano Zampini } 177053892102SStefano Zampini #endif 1771be83ff47SStefano Zampini } 177280677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 177380677318SStefano Zampini if (!pcbddc->switch_static) { 1774be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 177580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1777be83ff47SStefano Zampini } else { 1778be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1779be83ff47SStefano Zampini 178053892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178153892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1782be83ff47SStefano Zampini } 178380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 178480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 178580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 178680677318SStefano Zampini } 178780677318SStefano Zampini } else { 178880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 179380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 179480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 179580677318SStefano Zampini } 179680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 180080677318SStefano Zampini } 1801674ae819SStefano Zampini PetscFunctionReturn(0); 1802674ae819SStefano Zampini } 1803674ae819SStefano Zampini 1804dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1805674ae819SStefano Zampini #undef __FUNCT__ 1806674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1807dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1808674ae819SStefano Zampini { 1809674ae819SStefano Zampini PetscErrorCode ierr; 1810674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1811674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1812674ae819SStefano Zampini const PetscScalar zero = 0.0; 1813674ae819SStefano Zampini 1814674ae819SStefano Zampini PetscFunctionBegin; 1815dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1816dc359a40SStefano Zampini if (applytranspose) { 1817674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 18188eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1819dc359a40SStefano Zampini } else { 1820674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1821674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 182215aaf578SStefano Zampini } 182312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 182412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 182512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182712edc857SStefano Zampini 18289f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 182912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 183012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1831964fefecSStefano Zampini Vec rhs,sol; 1832964fefecSStefano Zampini 1833964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1834964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 183512edc857SStefano Zampini if (applytranspose) { 1836964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183712edc857SStefano Zampini } else { 1838964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 183912edc857SStefano Zampini } 184012edc857SStefano Zampini } 1841674ae819SStefano Zampini 1842674ae819SStefano Zampini /* Local solution on R nodes */ 184380677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 184480677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18459f00e9b4SStefano Zampini } 1846674ae819SStefano Zampini 18479f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18489f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 184912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1850674ae819SStefano Zampini 1851674ae819SStefano Zampini /* Sum contributions from two levels */ 1852dc359a40SStefano Zampini if (applytranspose) { 1853dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1854dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1855dc359a40SStefano Zampini } else { 1856674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18578eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1858dc359a40SStefano Zampini } 1859674ae819SStefano Zampini PetscFunctionReturn(0); 1860674ae819SStefano Zampini } 1861674ae819SStefano Zampini 1862674ae819SStefano Zampini #undef __FUNCT__ 1863674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 186412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1865674ae819SStefano Zampini { 1866674ae819SStefano Zampini PetscErrorCode ierr; 1867674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 186858da7f69SStefano Zampini PetscScalar *array; 186912edc857SStefano Zampini Vec from,to; 1870674ae819SStefano Zampini 1871674ae819SStefano Zampini PetscFunctionBegin; 187212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 187312edc857SStefano Zampini from = pcbddc->coarse_vec; 187412edc857SStefano Zampini to = pcbddc->vec1_P; 187512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 187612edc857SStefano Zampini Vec tvec; 187758da7f69SStefano Zampini 187858da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 187958da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 188012edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188158da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 188258da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 188358da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 188412edc857SStefano Zampini } 188512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 188612edc857SStefano Zampini from = pcbddc->vec1_P; 188712edc857SStefano Zampini to = pcbddc->coarse_vec; 188812edc857SStefano Zampini } 188912edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1890674ae819SStefano Zampini PetscFunctionReturn(0); 1891674ae819SStefano Zampini } 1892674ae819SStefano Zampini 1893674ae819SStefano Zampini #undef __FUNCT__ 1894674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 189512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1896674ae819SStefano Zampini { 1897674ae819SStefano Zampini PetscErrorCode ierr; 1898674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 189958da7f69SStefano Zampini PetscScalar *array; 190012edc857SStefano Zampini Vec from,to; 1901674ae819SStefano Zampini 1902674ae819SStefano Zampini PetscFunctionBegin; 190312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 190412edc857SStefano Zampini from = pcbddc->coarse_vec; 190512edc857SStefano Zampini to = pcbddc->vec1_P; 190612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 190712edc857SStefano Zampini from = pcbddc->vec1_P; 190812edc857SStefano Zampini to = pcbddc->coarse_vec; 190912edc857SStefano Zampini } 191012edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 191112edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 191212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 191312edc857SStefano Zampini Vec tvec; 191458da7f69SStefano Zampini 191512edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 191658da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 191758da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 191858da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 191958da7f69SStefano Zampini } 192058da7f69SStefano Zampini } else { 192158da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 192258da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 192312edc857SStefano Zampini } 192412edc857SStefano Zampini } 1925674ae819SStefano Zampini PetscFunctionReturn(0); 1926674ae819SStefano Zampini } 1927674ae819SStefano Zampini 1928984c4197SStefano Zampini /* uncomment for testing purposes */ 1929984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1930674ae819SStefano Zampini #undef __FUNCT__ 1931674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1932674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1933674ae819SStefano Zampini { 1934674ae819SStefano Zampini PetscErrorCode ierr; 1935674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1936674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1937674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1938984c4197SStefano Zampini /* one and zero */ 1939984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1940984c4197SStefano Zampini /* space to store constraints and their local indices */ 19419162d606SStefano Zampini PetscScalar *constraints_data; 19429162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19439162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19449162d606SStefano Zampini PetscInt *constraints_n; 1945984c4197SStefano Zampini /* iterators */ 1946b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1947984c4197SStefano Zampini /* BLAS integers */ 1948e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1949e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1950c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1951727cdba6SStefano Zampini /* reuse */ 19520e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19530e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1954984c4197SStefano Zampini /* change of basis */ 1955b3d85658SStefano Zampini PetscBool qr_needed; 19569162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1957984c4197SStefano Zampini /* auxiliary stuff */ 195864efe560SStefano Zampini PetscInt *nnz,*is_indices; 19598a0068c3SStefano Zampini PetscInt ncc; 1960984c4197SStefano Zampini /* some quantities */ 196145a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1962a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1963984c4197SStefano Zampini 1964674ae819SStefano Zampini PetscFunctionBegin; 19658e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19668e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19678e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1968088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1969088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19700e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19710e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19720e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19730e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19740e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1975088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1976cf5a6209SStefano Zampini 1977cf5a6209SStefano Zampini /* print some info */ 1978cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1979cf5a6209SStefano Zampini IS vertices; 1980cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1981cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1982cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1983cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1984cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1985cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1986cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1987fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1988fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1989cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1990cf5a6209SStefano Zampini } 1991cf5a6209SStefano Zampini 1992cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19939162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1994cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1995cf5a6209SStefano Zampini const Vec *nearnullvecs; 1996cf5a6209SStefano Zampini Vec *localnearnullsp; 1997cf5a6209SStefano Zampini PetscScalar *array; 1998cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1999cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 2000cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 2001b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 2002cf5a6209SStefano Zampini PetscScalar *work; 2003cf5a6209SStefano Zampini PetscReal *singular_vals; 2004cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2005cf5a6209SStefano Zampini PetscReal *rwork; 2006cf5a6209SStefano Zampini #endif 2007cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2008cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 2009cf5a6209SStefano Zampini #else 2010964fefecSStefano Zampini PetscBLASInt dummy_int=1; 2011964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 2012cf5a6209SStefano Zampini #endif 2013cf5a6209SStefano Zampini 2014674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 2015d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 2016d06fc5fdSStefano Zampini /* free unneeded index sets */ 2017d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 2018d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 2019674ae819SStefano Zampini } 2020d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 2021d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2022d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2023d06fc5fdSStefano Zampini } 2024d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2025d06fc5fdSStefano Zampini n_ISForEdges = 0; 2026d06fc5fdSStefano Zampini } 2027d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 2028d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2029d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2030d06fc5fdSStefano Zampini } 2031d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2032d06fc5fdSStefano Zampini n_ISForFaces = 0; 2033d06fc5fdSStefano Zampini } 203470022509SStefano Zampini 203570022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 203670022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 203770022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 203870022509SStefano Zampini if (pcbddc->NullSpace) { 203970022509SStefano Zampini PetscBool tbool[2],gbool[2]; 204070022509SStefano Zampini 204170022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2042b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2043d06fc5fdSStefano Zampini if (!ISForEdges) { 2044d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2045d06fc5fdSStefano Zampini } 2046b8ffe317SStefano Zampini } 2047d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2048d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2049d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2050d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2051d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 205298a51de6SStefano Zampini } 205370022509SStefano Zampini #endif 205408122e43SStefano Zampini 2055674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2056674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2057674ae819SStefano Zampini if (nearnullsp) { 2058674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2059f4ddd8eeSStefano Zampini /* remove any stored info */ 2060f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2061f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2062f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2063f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2064f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2065473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2066f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2067f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2068f4ddd8eeSStefano Zampini } 2069984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2070984c4197SStefano Zampini nnsp_size = 0; 2071674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2072674ae819SStefano Zampini } 2073984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2074984c4197SStefano Zampini max_constraints = nnsp_size; 2075984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2076984c4197SStefano Zampini 2077674ae819SStefano Zampini /* 2078674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20799162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20809162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20819162d606SStefano Zampini There can be multiple constraints per connected component 2082674ae819SStefano Zampini */ 2083674ae819SStefano Zampini n_vertices = 0; 2084674ae819SStefano Zampini if (ISForVertices) { 2085674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2086674ae819SStefano Zampini } 20879162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20889162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20899162d606SStefano Zampini 20909162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20919162d606SStefano Zampini total_counts *= max_constraints; 2092674ae819SStefano Zampini total_counts += n_vertices; 20934641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20949162d606SStefano Zampini 2095674ae819SStefano Zampini total_counts = 0; 2096674ae819SStefano Zampini max_size_of_constraint = 0; 2097674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20989162d606SStefano Zampini IS used_is; 2099674ae819SStefano Zampini if (i<n_ISForEdges) { 21009162d606SStefano Zampini used_is = ISForEdges[i]; 2101674ae819SStefano Zampini } else { 21029162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2103674ae819SStefano Zampini } 21049162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2105674ae819SStefano Zampini total_counts += j; 2106674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2107674ae819SStefano Zampini } 21089162d606SStefano 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); 21099162d606SStefano Zampini 2110984c4197SStefano Zampini /* get local part of global near null space vectors */ 2111785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2112984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2113984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2114984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2115984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2116984c4197SStefano Zampini } 2117674ae819SStefano Zampini 2118242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2119242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2120a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2121242a89d7SStefano Zampini 2122984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2123a773dcb8SStefano Zampini if (!skip_lapack) { 2124674ae819SStefano Zampini PetscScalar temp_work; 2125911cabfeSStefano Zampini 2126674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2127984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2128785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2129785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2130785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2131674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2132785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2133674ae819SStefano Zampini #endif 2134674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2135c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2136c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2137674ae819SStefano Zampini lwork = -1; 2138674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2139674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2140c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2141674ae819SStefano Zampini #else 2142c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2143674ae819SStefano Zampini #endif 2144674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2145984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2146674ae819SStefano Zampini #else /* on missing GESVD */ 2147674ae819SStefano Zampini /* SVD */ 2148674ae819SStefano Zampini PetscInt max_n,min_n; 2149674ae819SStefano Zampini max_n = max_size_of_constraint; 2150984c4197SStefano Zampini min_n = max_constraints; 2151984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2152674ae819SStefano Zampini min_n = max_size_of_constraint; 2153984c4197SStefano Zampini max_n = max_constraints; 2154674ae819SStefano Zampini } 2155785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2156674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2157785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2158674ae819SStefano Zampini #endif 2159674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2160674ae819SStefano Zampini lwork = -1; 2161e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2162e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2163b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2164674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2165674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21669162d606SStefano 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)); 2167674ae819SStefano Zampini #else 21689162d606SStefano 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)); 2169674ae819SStefano Zampini #endif 2170674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2171984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2172984c4197SStefano Zampini #endif /* on missing GESVD */ 2173674ae819SStefano Zampini /* Allocate optimal workspace */ 2174674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2175854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2176674ae819SStefano Zampini } 2177674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2178674ae819SStefano Zampini total_counts = 0; 21799162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21809162d606SStefano Zampini constraints_data_ptr[0] = 0; 2181674ae819SStefano Zampini /* vertices */ 21829162d606SStefano Zampini if (n_vertices) { 2183674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21849162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21859162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2186674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21879162d606SStefano Zampini constraints_n[total_counts] = 1; 21889162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21899162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21909162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2191674ae819SStefano Zampini total_counts++; 2192674ae819SStefano Zampini } 2193674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2194984c4197SStefano Zampini PetscBool used_vertex; 2195674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2196674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2197674ae819SStefano Zampini k = 0; 2198674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2199984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2200984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 22019162d606SStefano Zampini constraints_n[total_counts] = 1; 22029162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 22039162d606SStefano Zampini constraints_data[total_counts] = 1.0; 22049162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 22059162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2206674ae819SStefano Zampini total_counts++; 2207674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2208674ae819SStefano Zampini } 2209984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2210674ae819SStefano Zampini k++; 2211674ae819SStefano Zampini } 2212674ae819SStefano Zampini } 2213674ae819SStefano Zampini } 2214674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2215674ae819SStefano Zampini n_vertices = total_counts; 2216674ae819SStefano Zampini } 2217984c4197SStefano Zampini 2218674ae819SStefano Zampini /* edges and faces */ 22199162d606SStefano Zampini total_counts_cc = total_counts; 2220911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 22219162d606SStefano Zampini IS used_is; 22229162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 22239162d606SStefano Zampini 2224911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 22259162d606SStefano Zampini used_is = ISForEdges[ncc]; 2226984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2227674ae819SStefano Zampini } else { 22289162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2229984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2230674ae819SStefano Zampini } 2231674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22329162d606SStefano Zampini 22339162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22349162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2235984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2236984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2237674ae819SStefano Zampini if (nnsp_has_cnst) { 22385b08dc53SStefano Zampini PetscScalar quad_value; 22399162d606SStefano Zampini 22409162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22419162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22429162d606SStefano Zampini 2243a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2244674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2245a773dcb8SStefano Zampini } else { 2246a773dcb8SStefano Zampini quad_value = 1.0; 2247a773dcb8SStefano Zampini } 2248674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22499162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2250674ae819SStefano Zampini } 22519162d606SStefano Zampini temp_constraints++; 2252674ae819SStefano Zampini total_counts++; 2253674ae819SStefano Zampini } 2254674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2255984c4197SStefano Zampini PetscReal real_value; 22569162d606SStefano Zampini PetscScalar *ptr_to_data; 22579162d606SStefano Zampini 2258984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22599162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2260674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22619162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2262674ae819SStefano Zampini } 2263984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2264984c4197SStefano Zampini /* check if array is null on the connected component */ 2265e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22669162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22675b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2268674ae819SStefano Zampini temp_constraints++; 2269674ae819SStefano Zampini total_counts++; 22709162d606SStefano Zampini if (!idxs_copied) { 22719162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22729162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2273674ae819SStefano Zampini } 2274674ae819SStefano Zampini } 22759162d606SStefano Zampini } 22769162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 227745a1bb75SStefano Zampini valid_constraints = temp_constraints; 2278eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2279a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22809162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22819162d606SStefano Zampini 22829162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2283a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22849162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2285a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22869162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2287a773dcb8SStefano Zampini } else { /* perform SVD */ 2288984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22899162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2290674ae819SStefano Zampini 2291674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2292984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2293984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2294984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2295984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2296984c4197SStefano Zampini from that computed using LAPACKgesvd 2297984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2298984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2299984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2300674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2301e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2302984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2303674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2304674ae819SStefano Zampini for (k=0;k<j+1;k++) { 23059162d606SStefano 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)); 2306674ae819SStefano Zampini } 2307674ae819SStefano Zampini } 2308e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2309e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2310e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2311674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2312c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2313674ae819SStefano Zampini #else 2314c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2315674ae819SStefano Zampini #endif 2316674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2317984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2318984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2319674ae819SStefano Zampini j = 0; 2320984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2321674ae819SStefano Zampini total_counts = total_counts-j; 232245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2323e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2324c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2325c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2326c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2327c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2328c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2329c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2330674ae819SStefano Zampini if (j<temp_constraints) { 2331984c4197SStefano Zampini PetscInt ii; 2332984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2333674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23349162d606SStefano 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)); 2335674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2336984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2337674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23389162d606SStefano 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]; 2339674ae819SStefano Zampini } 2340674ae819SStefano Zampini } 2341674ae819SStefano Zampini } 2342674ae819SStefano Zampini #else /* on missing GESVD */ 2343e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2344e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2345b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2346674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2347674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23489162d606SStefano 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)); 2349674ae819SStefano Zampini #else 23509162d606SStefano 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)); 2351674ae819SStefano Zampini #endif 2352984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2353674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2354984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2355e310c8b4SStefano Zampini k = temp_constraints; 2356e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2357674ae819SStefano Zampini j = 0; 2358e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 235945a1bb75SStefano Zampini valid_constraints = k-j; 2360911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2361984c4197SStefano Zampini #endif /* on missing GESVD */ 2362674ae819SStefano Zampini } 2363a773dcb8SStefano Zampini } 23649162d606SStefano Zampini /* update pointers information */ 23659162d606SStefano Zampini if (valid_constraints) { 23669162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23679162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23689162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23699162d606SStefano Zampini /* set change_of_basis flag */ 237045a1bb75SStefano Zampini if (boolforchange) { 2371b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23729162d606SStefano Zampini } 2373b3d85658SStefano Zampini total_counts_cc++; 237445a1bb75SStefano Zampini } 2375674ae819SStefano Zampini } 2376984c4197SStefano Zampini /* free workspace */ 23778f1c130eSStefano Zampini if (!skip_lapack) { 2378984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2379984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2380984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2381984c4197SStefano Zampini #endif 2382984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2383984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2384984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2385984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2386984c4197SStefano Zampini #endif 2387984c4197SStefano Zampini } 2388984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2389984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2390984c4197SStefano Zampini } 2391984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2392cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2393cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2394cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2395cf5a6209SStefano Zampini } 2396cf5a6209SStefano Zampini if (n_ISForFaces) { 2397cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2398cf5a6209SStefano Zampini } 2399cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2400cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2401cf5a6209SStefano Zampini } 2402cf5a6209SStefano Zampini if (n_ISForEdges) { 2403cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2404cf5a6209SStefano Zampini } 2405cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 240608122e43SStefano Zampini } else { 240708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 240808122e43SStefano Zampini 240908122e43SStefano Zampini total_counts = 0; 241008122e43SStefano Zampini n_vertices = 0; 2411d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2412d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 241308122e43SStefano Zampini } 241408122e43SStefano Zampini max_constraints = 0; 24159162d606SStefano Zampini total_counts_cc = 0; 241608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 241708122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 24189162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 241908122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 242008122e43SStefano Zampini } 24219162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 24229162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 24239162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 24249162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 242574d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 24269162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 24279162d606SStefano Zampini total_counts_cc = 0; 24289162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24299162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24309162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 243108122e43SStefano Zampini } 243208122e43SStefano Zampini } 24339162d606SStefano Zampini #if 0 24349162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24359162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24369162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24379162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24389162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24399162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24409162d606SStefano Zampini } 24419162d606SStefano Zampini printf("\n"); 24429162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24439162d606SStefano Zampini } 24441b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24458bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24461b968477SStefano Zampini } 24471b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24488bec7fa6SStefano 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]); 24491b968477SStefano Zampini } 245008122e43SStefano Zampini #endif 245108122e43SStefano Zampini 24528bec7fa6SStefano Zampini max_size_of_constraint = 0; 24539162d606SStefano 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]); 24549162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 245508122e43SStefano Zampini /* Change of basis */ 2456b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 245708122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 245808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 245908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2460b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 246108122e43SStefano Zampini } 246208122e43SStefano Zampini } 246308122e43SStefano Zampini } 246408122e43SStefano Zampini } 2465b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24660e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 246708122e43SStefano Zampini 24689162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24699162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24709162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24719162d606SStefano 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); 247208122e43SStefano Zampini } 2473984c4197SStefano Zampini 2474674ae819SStefano Zampini /* Create constraint matrix */ 2475674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 247616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2477984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2478984c4197SStefano Zampini 2479984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 248074d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 248174d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 248274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2483984c4197SStefano Zampini total_primal_vertices=0; 2484b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24859162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24869162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2487984c4197SStefano Zampini if (size_of_constraint == 1) { 24889162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2489b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 249064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24919162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24929162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24939162d606SStefano Zampini } 2494b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 249574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 249674d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 249774d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 249874d5cdf7SStefano Zampini } 2499b3d85658SStefano Zampini } else { 2500b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2501984c4197SStefano Zampini } 2502984c4197SStefano Zampini } 2503b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2504b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 250570022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 250670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2507b3d85658SStefano Zampini 2508b3d85658SStefano 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); 25090e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 25100e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 251170022509SStefano Zampini 2512984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 251374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2514785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2515984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 251674d5cdf7SStefano Zampini 2517984c4197SStefano Zampini j = total_primal_vertices; 251874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2519b3d85658SStefano Zampini cum = total_primal_vertices; 25209162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25214641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2522b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2523b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2524b3d85658SStefano Zampini cum++; 25259162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 252674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 252774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 252874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 252974d5cdf7SStefano Zampini } 25309162d606SStefano Zampini j += constraints_n[i]; 2531674ae819SStefano Zampini } 2532674ae819SStefano Zampini } 2533674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2534674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2535088faed8SStefano Zampini 2536674ae819SStefano Zampini /* set values in constraint matrix */ 2537984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25380e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2539674ae819SStefano Zampini } 2540984c4197SStefano Zampini total_counts = total_primal_vertices; 25419162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25424641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25439162d606SStefano Zampini PetscInt *cols; 25449162d606SStefano Zampini 25459162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25469162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25479162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25489162d606SStefano Zampini PetscInt row = total_counts+k; 25499162d606SStefano Zampini PetscScalar *vals; 25509162d606SStefano Zampini 25519162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25529162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25539162d606SStefano Zampini } 25549162d606SStefano Zampini total_counts += constraints_n[i]; 2555674ae819SStefano Zampini } 2556674ae819SStefano Zampini } 2557674ae819SStefano Zampini /* assembling */ 2558674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2559674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2560088faed8SStefano Zampini 2561984c4197SStefano Zampini /* 256245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2563984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2564984c4197SStefano Zampini */ 2565674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2566674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2567026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2568984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2569984c4197SStefano Zampini /* working stuff for GEQRF */ 257081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2571984c4197SStefano Zampini PetscBLASInt lqr_work; 2572984c4197SStefano Zampini /* working stuff for UNGQR */ 2573984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2574984c4197SStefano Zampini PetscBLASInt lgqr_work; 2575984c4197SStefano Zampini /* working stuff for TRTRS */ 2576984c4197SStefano Zampini PetscScalar *trs_rhs; 25773f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2578984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2579984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2580984c4197SStefano Zampini PetscScalar *start_vals; 2581984c4197SStefano Zampini /* working stuff for values insertion */ 25824641a718SStefano Zampini PetscBT is_primal; 258364efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2584906d46d4SStefano Zampini /* matrix sizes */ 2585906d46d4SStefano Zampini PetscInt global_size,local_size; 2586906d46d4SStefano Zampini /* temporary change of basis */ 2587906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2588cf5a6209SStefano Zampini /* extra space for debugging */ 2589cf5a6209SStefano Zampini PetscScalar *dbg_work; 2590984c4197SStefano Zampini 2591906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2592906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 259316f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2594bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2595906d46d4SStefano Zampini /* nonzeros for local mat */ 2596bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2597bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25989162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2599a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 26009162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2601a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 26029162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2603a717540cSStefano Zampini } else { 26049162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 26059162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2606a717540cSStefano Zampini } 2607a717540cSStefano Zampini } 2608a717540cSStefano Zampini } 2609906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2610bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2611a717540cSStefano Zampini /* Set initial identity in the matrix */ 2612bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2613906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2614a717540cSStefano Zampini } 2615a717540cSStefano Zampini 2616a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2617a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2618a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2619a717540cSStefano Zampini } 2620a717540cSStefano Zampini 2621a717540cSStefano Zampini 2622a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2623a717540cSStefano Zampini /* 2624a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2625a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2626a717540cSStefano Zampini 2627a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2628a717540cSStefano Zampini 2629a6b551f4SStefano 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) 2630a6b551f4SStefano Zampini 2631a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2632a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2633a717540cSStefano Zampini | ... | 2634a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2635a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2636a717540cSStefano Zampini 2637a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2638a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2639a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2640a6b551f4SStefano Zampini 2641a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2642a717540cSStefano Zampini */ 2643a717540cSStefano Zampini if (qr_needed) { 2644984c4197SStefano Zampini /* space to store Q */ 2645854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2646984c4197SStefano Zampini /* first we issue queries for optimal work */ 26473f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26483f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26493f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2650984c4197SStefano Zampini lqr_work = -1; 26513f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2652984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2653984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2654785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2655984c4197SStefano Zampini lgqr_work = -1; 26563f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26573f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26583f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26593f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26603f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26613f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2662984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2663984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2664785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2665984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2666785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2667984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2668785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2669a717540cSStefano Zampini /* allocating workspace for check */ 2670a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2671cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2672a717540cSStefano Zampini } 2673a717540cSStefano Zampini } 2674984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26754641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2676473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26770e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 267839e2fb2aSStefano Zampini if (i != total_primal_vertices) { 267939e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26804641a718SStefano Zampini } 268139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 268239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 268339e2fb2aSStefano Zampini } 268439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2685984c4197SStefano Zampini 2686a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26879162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26889162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26894641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2690984c4197SStefano Zampini /* get constraint info */ 26919162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2692984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2693984c4197SStefano Zampini 2694984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26959162d606SStefano 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); 2696674ae819SStefano Zampini } 2697984c4197SStefano Zampini 2698fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2699a717540cSStefano Zampini 2700a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2701a717540cSStefano Zampini if (pcbddc->dbg_flag) { 27029162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2703a717540cSStefano Zampini } 2704984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 27059162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2706984c4197SStefano Zampini 2707984c4197SStefano Zampini /* compute QR decomposition of constraints */ 27083f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27093f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27103f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2711674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27123f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2713984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2714674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2715984c4197SStefano Zampini 2716984c4197SStefano Zampini /* explictly compute R^-T */ 2717984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2718984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 27193f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27203f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 27213f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27223f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2723984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27243f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2725984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2726984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2727984c4197SStefano Zampini 2728a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27313f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2733984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27343f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2735984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2736984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2737984c4197SStefano Zampini 2738984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2739984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2740984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27413f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27423f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2747984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27489162d606SStefano 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)); 2749984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27509162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2751984c4197SStefano Zampini 2752984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27539162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2754984c4197SStefano Zampini /* insert cols for primal dofs */ 2755984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2756984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27579162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2758906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2759984c4197SStefano Zampini } 2760984c4197SStefano Zampini /* insert cols for dual dofs */ 2761984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27629162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2763984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27649162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2765906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2766984c4197SStefano Zampini j++; 2767674ae819SStefano Zampini } 2768674ae819SStefano Zampini } 2769984c4197SStefano Zampini 2770984c4197SStefano Zampini /* check change of basis */ 2771984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2772984c4197SStefano Zampini PetscInt ii,jj; 2773984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2774c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2775c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2776c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2777c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2778c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2779c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2780984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2781cf5a6209SStefano 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)); 2782984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2783984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2784984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2785cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2786cf5a6209SStefano 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; 2787674ae819SStefano Zampini } 2788674ae819SStefano Zampini } 2789984c4197SStefano Zampini if (!valid_qr) { 279022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2791984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2792984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2793cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2794cf5a6209SStefano 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])); 2795674ae819SStefano Zampini } 2796cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2797cf5a6209SStefano 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])); 2798984c4197SStefano Zampini } 2799984c4197SStefano Zampini } 2800984c4197SStefano Zampini } 2801674ae819SStefano Zampini } else { 280222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2803674ae819SStefano Zampini } 2804674ae819SStefano Zampini } 2805a717540cSStefano Zampini } else { /* simple transformation block */ 2806a717540cSStefano Zampini PetscInt row,col; 2807a6b551f4SStefano Zampini PetscScalar val,norm; 2808a6b551f4SStefano Zampini 2809a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 28109162d606SStefano 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)); 2811a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 28129162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 28139162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2814bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 28159162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2816906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 28179162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2818a717540cSStefano Zampini } else { 2819a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 28209162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2821a717540cSStefano Zampini if (row != col) { 28229162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2823a717540cSStefano Zampini } else { 28249162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2825a717540cSStefano Zampini } 2826906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2827a717540cSStefano Zampini } 2828a717540cSStefano Zampini } 2829a717540cSStefano Zampini } 283098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 283122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2832a717540cSStefano Zampini } 2833674ae819SStefano Zampini } 2834984c4197SStefano Zampini } else { 2835984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28369162d606SStefano 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); 2837674ae819SStefano Zampini } 2838674ae819SStefano Zampini } 2839674ae819SStefano Zampini } 2840a717540cSStefano Zampini 2841a717540cSStefano Zampini /* free workspace */ 2842a717540cSStefano Zampini if (qr_needed) { 2843984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2844cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2845984c4197SStefano Zampini } 2846984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2847984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2848984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2849984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2850984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2851674ae819SStefano Zampini } 2852a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2853906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2854906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2855906d46d4SStefano Zampini 2856906d46d4SStefano Zampini /* assembling of global change of variable */ 2857bbb9e6c6SStefano Zampini { 2858bbb9e6c6SStefano Zampini Mat tmat; 285916f15bc4SStefano Zampini PetscInt bs; 286016f15bc4SStefano Zampini 2861906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2862906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2863bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2865bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2866bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 286716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 286816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2869906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2870bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2871bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2872bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2873bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2874bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2875bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2876bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2877bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2878bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2879906d46d4SStefano Zampini } 2880906d46d4SStefano Zampini /* check */ 2881906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2882906d46d4SStefano Zampini PetscReal error; 2883906d46d4SStefano Zampini Vec x,x_change; 2884906d46d4SStefano Zampini 2885906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2886906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2887906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2888906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2889bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2890bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2891bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2892bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2893bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2894906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2895906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2896906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2897906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2898bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2899906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2900906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2901906d46d4SStefano Zampini } 2902b96c3477SStefano Zampini 2903b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2904b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2905b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2906b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2907ac632422SStefano Zampini Mat S_new,tmat; 2908bbb9e6c6SStefano Zampini IS is_all_N; 2909bbb9e6c6SStefano Zampini 2910bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 29116816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2912bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2913ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2914b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2915ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2916ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2917ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2918ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2919ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2920b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2921ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2922ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2923ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2924ac632422SStefano Zampini } 2925b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2926b96c3477SStefano Zampini } 2927b96c3477SStefano Zampini } 2928906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2929906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2930b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2931b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2932b9b85e73SStefano Zampini } 2933906d46d4SStefano Zampini 2934906d46d4SStefano Zampini /* set up change of basis context */ 2935906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2936906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2937906d46d4SStefano Zampini 2938906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2939906d46d4SStefano Zampini PetscInt global_size,local_size; 2940906d46d4SStefano Zampini 2941906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2942906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2943906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2944906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2945906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2946906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2947906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2948906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2949906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2950906d46d4SStefano Zampini } else { 2951906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2952906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2953906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2954906d46d4SStefano Zampini } 2955906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2956906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2957906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2958906d46d4SStefano Zampini } else { 2959906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2960906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2961906d46d4SStefano Zampini } 2962906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2963906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2964906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2965906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2966b9b85e73SStefano Zampini } 2967a717540cSStefano Zampini 2968727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2969727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2970727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 2971aff50787SStefano 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); 29720e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29730e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 2974aff50787SStefano 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); 2975c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2976727cdba6SStefano Zampini } 29770e6343abSStefano Zampini } 29780e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2979727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2980727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2981727cdba6SStefano Zampini 2982a717540cSStefano Zampini /* flush dbg viewer */ 2983b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2984b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2985b8ffe317SStefano Zampini } 2986a717540cSStefano Zampini 2987e310c8b4SStefano Zampini /* free workspace */ 2988a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29894641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 299008122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29919162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29929162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 299308122e43SStefano Zampini } else { 29949162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29959162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29969162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 299708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 299808122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29999162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 30009162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 300108122e43SStefano Zampini } 3002674ae819SStefano Zampini PetscFunctionReturn(0); 3003674ae819SStefano Zampini } 3004674ae819SStefano Zampini 3005674ae819SStefano Zampini #undef __FUNCT__ 3006674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 3007674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 3008674ae819SStefano Zampini { 3009674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3010674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 3011674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 30127fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 3013674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 3014674ae819SStefano Zampini 3015674ae819SStefano Zampini PetscFunctionBegin; 30168e61c736SStefano Zampini /* Reset previously computed graph */ 30178e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 3018674ae819SStefano Zampini /* Init local Graph struct */ 30197fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 30207fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 3021674ae819SStefano Zampini 3022575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 3023575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 3024575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 3025575ad6abSStefano Zampini } 30269577ea80SStefano Zampini 3027674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 30284d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30294d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30304d379d7bSStefano Zampini PetscInt nvtxs; 3031e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3032674ae819SStefano Zampini 30334d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30342fffb893SStefano Zampini 30352fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30362fffb893SStefano Zampini if (flg_row) { 30374d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3038b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30392fffb893SStefano Zampini } 30402fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30419b28b941SStefano 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 */ 30424d379d7bSStefano Zampini IS is_dummy; 30434d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30444d379d7bSStefano Zampini PetscInt j,sum; 30454d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30464d379d7bSStefano Zampini const PetscInt *idxs; 30474d379d7bSStefano Zampini PCBDDCGraph graph; 30484d379d7bSStefano Zampini PetscBT is_on_boundary; 30494d379d7bSStefano Zampini 30504d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30514d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30524d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30534d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30547fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30554d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3056e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3057e496cd5dSStefano Zampini if (flg_row) { 30584d379d7bSStefano Zampini graph->xadj = xadj; 30594d379d7bSStefano Zampini graph->adjncy = adjncy; 3060e496cd5dSStefano Zampini } 30614d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30624d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3063e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30644d379d7bSStefano Zampini 30654d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30669b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30674d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30684d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30694d379d7bSStefano Zampini } 30704d379d7bSStefano Zampini } 30714d379d7bSStefano Zampini 3072e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30734d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30744d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30754d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30764d379d7bSStefano Zampini } 30774d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30784d379d7bSStefano Zampini 3079e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30804d379d7bSStefano Zampini sum = 0; 30814d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30824d379d7bSStefano Zampini PetscInt sizecc = 0; 30834d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30844d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30854d379d7bSStefano Zampini sizecc++; 30864d379d7bSStefano Zampini } 30874d379d7bSStefano Zampini } 30884d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30894d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30904d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30914d379d7bSStefano Zampini } 30924d379d7bSStefano Zampini } 30934d379d7bSStefano Zampini sum += sizecc*sizecc; 30944d379d7bSStefano Zampini } 30954d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30964d379d7bSStefano Zampini sum = 0; 3097e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30984d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30994d379d7bSStefano Zampini cxadj[i] = sum; 31004d379d7bSStefano Zampini sum += temp; 31014d379d7bSStefano Zampini } 3102e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 31034d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 31044d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 31054d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 31064d379d7bSStefano Zampini PetscInt k,sizecc = 0; 31074d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 31084d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 31094d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 31104d379d7bSStefano Zampini sizecc++; 31114d379d7bSStefano Zampini } 31124d379d7bSStefano Zampini } 31134d379d7bSStefano Zampini } 31144d379d7bSStefano Zampini } 31154d379d7bSStefano Zampini } 31169b28b941SStefano Zampini if (sum) { 3117e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 31184d379d7bSStefano Zampini } else { 31194d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 31204d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 31214d379d7bSStefano Zampini } 31224d379d7bSStefano Zampini graph->xadj = 0; 31234d379d7bSStefano Zampini graph->adjncy = 0; 31244d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 31254d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 31264d379d7bSStefano Zampini } 3127674ae819SStefano Zampini } 31289b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31299b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31309b28b941SStefano Zampini } 3131674ae819SStefano Zampini 313263602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3133674ae819SStefano Zampini vertex_size = 1; 313463602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 313563602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 313695ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 313763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3138a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 313963602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3140674ae819SStefano Zampini } 314163602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 314263602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 314363602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3144674ae819SStefano Zampini } 314563602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3146674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 314763602bcaSStefano Zampini } else { 314863602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 314963602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3150854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 315163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 315263602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 315363602bcaSStefano Zampini } 315463602bcaSStefano Zampini } 3155674ae819SStefano Zampini } 3156674ae819SStefano Zampini 3157674ae819SStefano Zampini /* Setup of Graph */ 3158785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3159a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3160785d1243SStefano Zampini } 3161785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3162a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3163785d1243SStefano Zampini } 3164302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3165674ae819SStefano Zampini 3166674ae819SStefano Zampini /* Graph's connected components analysis */ 3167674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3168674ae819SStefano Zampini 3169674ae819SStefano Zampini /* print some info to stdout */ 3170674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3171302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3172674ae819SStefano Zampini } 3173fb180af4SStefano Zampini 3174fb180af4SStefano Zampini /* mark topography has done */ 3175fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3176674ae819SStefano Zampini PetscFunctionReturn(0); 3177674ae819SStefano Zampini } 3178674ae819SStefano Zampini 3179dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 3180674ae819SStefano Zampini #undef __FUNCT__ 3181674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3182dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 3183674ae819SStefano Zampini { 3184dc456d91SStefano Zampini PetscSF sf; 3185dc456d91SStefano Zampini PetscLayout map; 3186dc456d91SStefano Zampini const PetscInt *idxs; 3187dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 3188dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 3189dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 3190dc456d91SStefano Zampini PetscMPIInt commsize; 3191674ae819SStefano Zampini PetscBool first_found; 3192674ae819SStefano Zampini PetscErrorCode ierr; 3193674ae819SStefano Zampini 3194674ae819SStefano Zampini PetscFunctionBegin; 3195dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 3196dc456d91SStefano Zampini if (subset_mult) { 3197dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 3198dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 3199dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 3200674ae819SStefano Zampini } 3201dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 3202dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 3203dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 3204dc456d91SStefano Zampini for (i=0;i<n;i++) { 3205dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 3206dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 3207674ae819SStefano Zampini } 3208dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 3209dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3210dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 3211dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 3212dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 3213dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 3214dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 3215dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 3216dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 3217dc456d91SStefano Zampini 3218dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 3219dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 3220dc456d91SStefano Zampini if (subset_mult) { 3221dc456d91SStefano Zampini const PetscInt* idxs_mult; 3222dc456d91SStefano Zampini 3223dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3224dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 3225dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3226674ae819SStefano Zampini } else { 3227dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 3228674ae819SStefano Zampini } 3229dc456d91SStefano Zampini /* local size of new subset */ 3230dc456d91SStefano Zampini n_n = 0; 3231dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 3232dc456d91SStefano Zampini 3233dc456d91SStefano Zampini /* global indexes in layout */ 3234dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 3235dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 3236dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 3237dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 3238dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 3239dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 3240dc456d91SStefano Zampini 3241dc456d91SStefano Zampini /* reduce from leaves to roots */ 3242dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 3243dc456d91SStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPIU_MAX);CHKERRQ(ierr); 3244dc456d91SStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPIU_MAX);CHKERRQ(ierr); 3245dc456d91SStefano Zampini 3246dc456d91SStefano Zampini /* count indexes in local part of layout */ 3247674ae819SStefano Zampini nlocals = 0; 3248674ae819SStefano Zampini first_index = -1; 3249674ae819SStefano Zampini first_found = PETSC_FALSE; 3250dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 3251dc456d91SStefano Zampini if (!first_found && root_data[i]) { 3252674ae819SStefano Zampini first_found = PETSC_TRUE; 3253674ae819SStefano Zampini first_index = i; 3254674ae819SStefano Zampini } 3255dc456d91SStefano Zampini nlocals += root_data[i]; 3256674ae819SStefano Zampini } 3257dc456d91SStefano Zampini 3258dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 3259dc456d91SStefano Zampini start = 0; 3260dc456d91SStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPIU_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3261dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 3262dc456d91SStefano Zampini *N_n = start + nlocals; 3263dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 3264dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 3265674ae819SStefano Zampini } 3266dc456d91SStefano Zampini /* adapt roof data with cumulative */ 3267674ae819SStefano Zampini if (first_found) { 3268dc456d91SStefano Zampini PetscInt old_index; 3269dc456d91SStefano Zampini 3270dc456d91SStefano Zampini root_data[first_index] += start; 3271674ae819SStefano Zampini old_index = first_index; 3272dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 3273dc456d91SStefano Zampini if (root_data[i]) { 3274dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 3275674ae819SStefano Zampini old_index = i; 3276674ae819SStefano Zampini } 3277674ae819SStefano Zampini } 3278674ae819SStefano Zampini } 3279dc456d91SStefano Zampini 3280dc456d91SStefano Zampini /* from roots to leaves */ 3281dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3282dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 3283dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 3284dc456d91SStefano Zampini 3285dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 3286dc456d91SStefano Zampini if (subset_mult) { 3287dc456d91SStefano Zampini const PetscInt* idxs_mult; 3288dc456d91SStefano Zampini PetscInt cum; 3289dc456d91SStefano Zampini 3290dc456d91SStefano Zampini cum = 0; 3291dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3292dc456d91SStefano Zampini for (i=0;i<n;i++) { 3293dc456d91SStefano Zampini PetscInt j; 3294dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 3295674ae819SStefano Zampini } 3296dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 3297674ae819SStefano Zampini } else { 3298dc456d91SStefano Zampini for (i=0;i<n;i++) { 3299dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 3300674ae819SStefano Zampini } 3301674ae819SStefano Zampini } 3302dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 3303dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 3304674ae819SStefano Zampini PetscFunctionReturn(0); 3305674ae819SStefano Zampini } 33069a7d3425SStefano Zampini 33079a7d3425SStefano Zampini #undef __FUNCT__ 33089a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 33099a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 33109a7d3425SStefano Zampini { 33119a7d3425SStefano Zampini PetscInt i,j; 33129a7d3425SStefano Zampini PetscScalar *alphas; 33139a7d3425SStefano Zampini PetscErrorCode ierr; 33149a7d3425SStefano Zampini 33159a7d3425SStefano Zampini PetscFunctionBegin; 33169a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3317785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 33189a7d3425SStefano Zampini for (i=0;i<n;i++) { 33199a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 33209a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 33219a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 33229a7d3425SStefano Zampini } 33239a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 33249a7d3425SStefano Zampini PetscFunctionReturn(0); 33259a7d3425SStefano Zampini } 33269a7d3425SStefano Zampini 3327e7931f94SStefano Zampini #undef __FUNCT__ 332870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 3329b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 3330e7931f94SStefano Zampini { 3331e7931f94SStefano Zampini Mat subdomain_adj; 3332e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3333e7931f94SStefano Zampini MatPartitioning partitioner; 3334e7931f94SStefano Zampini Mat_IS *matis; 3335e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3336e7931f94SStefano Zampini PetscInt prank; 3337e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3338e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3339e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33403837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3341e7931f94SStefano Zampini PetscErrorCode ierr; 33422b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3343e7931f94SStefano Zampini PetscSubcomm subcomm; 3344a57a6d2fSStefano Zampini 3345e7931f94SStefano Zampini PetscFunctionBegin; 33462b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33472b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33482b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3349e7931f94SStefano Zampini 3350e7931f94SStefano Zampini /* Get info on mapping */ 3351e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3352e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3353e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3354e7931f94SStefano Zampini 3355e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3356785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3357e7931f94SStefano Zampini xadj[0] = 0; 3358e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3359785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3360785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3361e7931f94SStefano Zampini 33622b510759SStefano Zampini if (threshold) { 3363d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33642b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3365d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3366d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3367d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3368d023bfaeSStefano Zampini xadj_count++; 3369e7931f94SStefano Zampini } 3370e7931f94SStefano Zampini } 3371d023bfaeSStefano Zampini xadj[1] = xadj_count; 3372c8587f34SStefano Zampini } else { 3373e7931f94SStefano Zampini if (xadj[1]) { 3374e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3375e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3376c8587f34SStefano Zampini } 3377e7931f94SStefano Zampini } 3378e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3379e7931f94SStefano Zampini if (use_square) { 3380e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3381e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3382e7931f94SStefano Zampini } 3383e7931f94SStefano Zampini } 3384e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3385e7931f94SStefano Zampini 33863837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3387e7931f94SStefano Zampini 3388e7931f94SStefano Zampini /* 3389e7931f94SStefano Zampini Restrict work on active processes only. 3390e7931f94SStefano Zampini */ 3391e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3392e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3393e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33942b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3395d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3396e7931f94SStefano Zampini if (color) { 3397e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3398e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3399e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3400c8587f34SStefano Zampini } else { 3401b0c7d250SStefano Zampini PetscInt rstart=0,rend=0; 3402b0c7d250SStefano Zampini PetscBool aggregate; 3403b0c7d250SStefano Zampini 3404306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3405785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3406e7931f94SStefano Zampini prank = rank; 3407306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 34088002ef2cSStefano Zampini /* 3409e7931f94SStefano Zampini for (i=0;i<size;i++) { 3410e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3411c8587f34SStefano Zampini } 34128002ef2cSStefano Zampini */ 3413e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3414e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3415c8587f34SStefano Zampini } 3416e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3417b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 3418b0c7d250SStefano Zampini if (aggregate) { 3419b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 3420b0c7d250SStefano Zampini PetscMPIInt nrank; 3421b0c7d250SStefano Zampini PetscScalar *vals; 3422b0c7d250SStefano Zampini 3423b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 3424b0c7d250SStefano Zampini lrows = 0; 3425b0c7d250SStefano Zampini if (nrank<redprocs) { 3426b0c7d250SStefano Zampini lrows = size/redprocs; 3427b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 3428b0c7d250SStefano Zampini } 3429b0c7d250SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,100,NULL,100,NULL,&subdomain_adj);CHKERRQ(ierr); 3430b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 3431b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3432b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 3433b0c7d250SStefano Zampini row = nrank; 3434b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 3435b0c7d250SStefano Zampini cols = adjncy; 3436b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 3437b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 3438b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 3439b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3440b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3441b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 3442b0c7d250SStefano Zampini } else { 3443306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 3444b0c7d250SStefano Zampini } 344522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3446e7931f94SStefano Zampini 3447e7931f94SStefano Zampini /* Partition */ 3448306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3449e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3450e7931f94SStefano Zampini if (use_vwgt) { 34513837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3452e7931f94SStefano Zampini v_wgt[0] = local_size; 3453e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3454c8587f34SStefano Zampini } 345528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 345628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3457e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3458e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 345922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3460e7931f94SStefano Zampini 3461b0c7d250SStefano Zampini /* TODO: renumber new_ranks to avoid "holes" in recv processors */ 3462e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3463b0c7d250SStefano Zampini if (!redprocs) { 3464b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 346528143c3dSStefano Zampini } else { 3466b0c7d250SStefano Zampini PetscInt idxs[1]; 3467b0c7d250SStefano Zampini PetscMPIInt tag; 3468b0c7d250SStefano Zampini MPI_Request *reqs; 3469b0c7d250SStefano Zampini 3470b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 3471b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 3472b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 3473b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 3474b0c7d250SStefano Zampini } 3475b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 3476b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3477b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 3478b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 347928143c3dSStefano Zampini } 3480e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3481e7931f94SStefano Zampini /* clean up */ 3482e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3483e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3484e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3485e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3486e7931f94SStefano Zampini } 3487e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3488e7931f94SStefano Zampini 3489e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3490e7931f94SStefano Zampini i = 1; 3491e7931f94SStefano Zampini if (color) i=0; 3492e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3493e7931f94SStefano Zampini /* get back IS */ 3494e7931f94SStefano Zampini *is_sends = ranks_send_to; 3495e7931f94SStefano Zampini PetscFunctionReturn(0); 3496e7931f94SStefano Zampini } 3497e7931f94SStefano Zampini 3498e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3499e7931f94SStefano Zampini 3500e7931f94SStefano Zampini #undef __FUNCT__ 3501e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 350253a05cb3SStefano 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[]) 3503e7931f94SStefano Zampini { 350470cf5478SStefano Zampini Mat local_mat; 3505e7931f94SStefano Zampini Mat_IS *matis; 3506e7931f94SStefano Zampini IS is_sends_internal; 35079d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 350828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 35099d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3510e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3511e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3512e7931f94SStefano Zampini const PetscInt* is_indices; 3513e7931f94SStefano Zampini MatType new_local_type; 3514e7931f94SStefano Zampini /* buffers */ 3515e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 351628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 35179d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3518e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3519e7931f94SStefano Zampini /* MPI */ 352028143c3dSStefano Zampini MPI_Comm comm,comm_n; 352128143c3dSStefano Zampini PetscSubcomm subcomm; 3522e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 352328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 352428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 352528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 352628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 352728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3528e7931f94SStefano Zampini PetscErrorCode ierr; 3529e7931f94SStefano Zampini 3530e7931f94SStefano Zampini PetscFunctionBegin; 353128143c3dSStefano Zampini /* TODO: add missing checks */ 353228143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 353328143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 353428143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 353528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3536e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 353728143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3538e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3539e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3540e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3541e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3542e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 354328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 354470cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 354570cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 354628143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 354770cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 354870cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 354970cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 355070cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 355170cf5478SStefano Zampini } 3552e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3553e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3554e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3555e7931f94SStefano Zampini if (!is_sends) { 355628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 3557b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 3558c8587f34SStefano Zampini } else { 3559e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3560e7931f94SStefano Zampini is_sends_internal = is_sends; 3561c8587f34SStefano Zampini } 3562e7931f94SStefano Zampini 3563e7931f94SStefano Zampini /* get pointer of MATIS data */ 3564e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3565e7931f94SStefano Zampini 3566e7931f94SStefano Zampini /* get comm */ 3567a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3568e7931f94SStefano Zampini 3569e7931f94SStefano Zampini /* compute number of sends */ 3570e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3571e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3572e7931f94SStefano Zampini 3573e7931f94SStefano Zampini /* compute number of receives */ 3574e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3575785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3576e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3577e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3578e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3579e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3580e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3581e7931f94SStefano Zampini 358228143c3dSStefano Zampini /* restrict comm if requested */ 358328143c3dSStefano Zampini subcomm = 0; 358428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 358528143c3dSStefano Zampini if (restrict_comm) { 3586779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3587779c1cceSStefano Zampini 358828143c3dSStefano Zampini color = 0; 358953a05cb3SStefano Zampini if (restrict_full) { 359053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 359153a05cb3SStefano Zampini } else { 359253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 359353a05cb3SStefano Zampini } 359428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 359528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 359628143c3dSStefano Zampini /* check if reuse has been requested */ 359728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 359828143c3dSStefano Zampini if (*mat_n) { 359928143c3dSStefano Zampini PetscMPIInt subcommsize2; 360028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 360128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 360228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 360328143c3dSStefano Zampini } else { 360428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 360528143c3dSStefano Zampini } 360628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3607779c1cceSStefano Zampini PetscMPIInt rank; 3608779c1cceSStefano Zampini 3609779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 361028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 361128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 361228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3613306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 361428143c3dSStefano Zampini } 361528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 361628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 361728143c3dSStefano Zampini } else { 361828143c3dSStefano Zampini comm_n = comm; 361928143c3dSStefano Zampini } 362028143c3dSStefano Zampini 3621e7931f94SStefano Zampini /* prepare send/receive buffers */ 3622785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3623e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3624785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3625e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 362628143c3dSStefano Zampini if (nis) { 3627854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 362828143c3dSStefano Zampini } 3629e7931f94SStefano Zampini 363028143c3dSStefano Zampini /* Get data from local matrices */ 3631e7931f94SStefano Zampini if (!isdense) { 3632a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3633e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3634e7931f94SStefano Zampini /* 3635e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3636e7931f94SStefano Zampini send_buffer_idxs should contain: 3637e7931f94SStefano Zampini - MatType_PRIVATE type 3638e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3639e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3640e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3641e7931f94SStefano Zampini */ 3642e7931f94SStefano Zampini } else { 3643e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3644e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3645854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3646e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3647e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3648e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3649e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3650e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3651e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3652e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3653e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3654e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3655c8587f34SStefano Zampini } 3656c8587f34SStefano Zampini } 3657e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 365828143c3dSStefano Zampini /* additional is (if any) */ 365928143c3dSStefano Zampini if (nis) { 366028143c3dSStefano Zampini PetscMPIInt psum; 366128143c3dSStefano Zampini PetscInt j; 366228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 366328143c3dSStefano Zampini PetscInt plen; 366428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 366528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 366628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 366728143c3dSStefano Zampini } 3668854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 366928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 367028143c3dSStefano Zampini PetscInt plen; 367128143c3dSStefano Zampini const PetscInt *is_array_idxs; 367228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 367328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 367428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 367528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 367628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 367728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 367828143c3dSStefano Zampini } 367928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 368028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 368128143c3dSStefano Zampini } 368228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 368328143c3dSStefano Zampini } 368428143c3dSStefano Zampini 3685e7931f94SStefano Zampini buf_size_idxs = 0; 3686e7931f94SStefano Zampini buf_size_vals = 0; 368728143c3dSStefano Zampini buf_size_idxs_is = 0; 3688e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3689e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3690e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 369128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3692e7931f94SStefano Zampini } 3693785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3694785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 369595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3696e7931f94SStefano Zampini 3697e7931f94SStefano Zampini /* get new tags for clean communications */ 3698e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3699e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 370028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3701e7931f94SStefano Zampini 3702e7931f94SStefano Zampini /* allocate for requests */ 3703785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3704785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 370595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3706785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3707785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 370895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3709e7931f94SStefano Zampini 3710e7931f94SStefano Zampini /* communications */ 3711e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3712e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 371328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3714e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3715e7931f94SStefano Zampini source_dest = onodes[i]; 3716e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3717e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3718e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3719e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 372028143c3dSStefano Zampini if (nis) { 372128143c3dSStefano 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); 372228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 372328143c3dSStefano Zampini } 3724e7931f94SStefano Zampini } 3725e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3726e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3727e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3728e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 372928143c3dSStefano Zampini if (nis) { 373028143c3dSStefano 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); 373128143c3dSStefano Zampini } 3732e7931f94SStefano Zampini } 3733e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3734e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3735e7931f94SStefano Zampini 3736e7931f94SStefano Zampini /* assemble new l2g map */ 3737e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3738e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37399d30be91SStefano Zampini new_local_rows = 0; 3740e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37419d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3742e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3743e7931f94SStefano Zampini } 37449d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3745e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 37469d30be91SStefano Zampini new_local_rows = 0; 3747e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 37489d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 37499d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3750e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3751e7931f94SStefano Zampini } 37529d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 37539d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3754e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3755e7931f94SStefano Zampini 3756e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3757e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3758e7931f94SStefano 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) */ 3759e7931f94SStefano Zampini if (n_recvs) { 376028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3761e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3762e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3763e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3764e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3765e7931f94SStefano Zampini break; 3766e7931f94SStefano Zampini } 3767e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3768e7931f94SStefano Zampini } 3769e7931f94SStefano Zampini switch (new_local_type_private) { 377028143c3dSStefano Zampini case MATDENSE_PRIVATE: 377128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3772e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3773e7931f94SStefano Zampini bs = 1; 377428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 377528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 377628143c3dSStefano Zampini bs = 1; 377728143c3dSStefano Zampini } 3778e7931f94SStefano Zampini break; 3779e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3780e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3781e7931f94SStefano Zampini bs = 1; 3782e7931f94SStefano Zampini break; 3783e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3784e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3785e7931f94SStefano Zampini break; 3786e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3787e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3788e7931f94SStefano Zampini break; 3789e7931f94SStefano Zampini default: 37909d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3791e7931f94SStefano Zampini break; 3792e7931f94SStefano Zampini } 379328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 379428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 379528143c3dSStefano Zampini bs = 1; 3796e7931f94SStefano Zampini } 3797e7931f94SStefano Zampini 379870cf5478SStefano Zampini /* create MATIS object if needed */ 379970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3800e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 380128143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 380270cf5478SStefano Zampini } else { 380370cf5478SStefano Zampini /* it also destroys the local matrices */ 380470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 380570cf5478SStefano Zampini } 380670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3807e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 38089d30be91SStefano Zampini 38099d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 38109d30be91SStefano Zampini 38119d30be91SStefano Zampini /* Global to local map of received indices */ 38129d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 38139d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 38149d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 38159d30be91SStefano Zampini 38169d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 38179d30be91SStefano Zampini buf_size_idxs = 0; 38189d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38199d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 38209d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 38219d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 38229d30be91SStefano Zampini } 38239d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 38249d30be91SStefano Zampini 38259d30be91SStefano Zampini /* set preallocation */ 38269d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 38279d30be91SStefano Zampini if (!newisdense) { 38289d30be91SStefano Zampini PetscInt *new_local_nnz=0; 38299d30be91SStefano Zampini 38309d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 38319d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 38329d30be91SStefano Zampini if (n_recvs) { 38339d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 38349d30be91SStefano Zampini } 38359d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 38369d30be91SStefano Zampini PetscInt j; 38379d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 38389d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 38399d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 38409d30be91SStefano Zampini } 38419d30be91SStefano Zampini } else { 38429d30be91SStefano Zampini /* TODO */ 38439d30be91SStefano Zampini } 38449d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 38459d30be91SStefano Zampini } 38469d30be91SStefano Zampini if (new_local_nnz) { 38479d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 38489d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 38499d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 38509d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38519d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 38529d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 38539d30be91SStefano Zampini } else { 38549d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38559d30be91SStefano Zampini } 38569d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 38579d30be91SStefano Zampini } else { 38589d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 38599d30be91SStefano Zampini } 3860e7931f94SStefano Zampini 3861e7931f94SStefano Zampini /* set values */ 3862e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 38639d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3864e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3865e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3866e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 38679d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3868e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3869e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3870e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 387128143c3dSStefano Zampini } else { 387228143c3dSStefano Zampini /* TODO */ 3873e7931f94SStefano Zampini } 3874e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3875e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3876e7931f94SStefano Zampini } 3877e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3878e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 387970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 388070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38819d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38829d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3883e7931f94SStefano Zampini 3884dfd14d43SStefano Zampini #if 0 388528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3886e7931f94SStefano Zampini Vec lvec,rvec; 3887e7931f94SStefano Zampini PetscReal infty_error; 3888e7931f94SStefano Zampini 38892a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3890e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3891e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3892e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 389370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3894e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3895e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3896e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3897e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3898e7931f94SStefano Zampini } 389928143c3dSStefano Zampini #endif 3900e7931f94SStefano Zampini 390128143c3dSStefano Zampini /* assemble new additional is (if any) */ 390228143c3dSStefano Zampini if (nis) { 390328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 390428143c3dSStefano Zampini 390528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3906854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 390728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 390828143c3dSStefano Zampini psum = 0; 390928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 391028143c3dSStefano Zampini for (j=0;j<nis;j++) { 391128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 391228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 391328143c3dSStefano Zampini psum += plen; 391428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 391528143c3dSStefano Zampini } 391628143c3dSStefano Zampini } 3917854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3918854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 391928143c3dSStefano Zampini for (i=1;i<nis;i++) { 392028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 392128143c3dSStefano Zampini } 392228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 392328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 392428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 392528143c3dSStefano Zampini for (j=0;j<nis;j++) { 392628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 392728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 392828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 392928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 393028143c3dSStefano Zampini } 393128143c3dSStefano Zampini } 393228143c3dSStefano Zampini for (i=0;i<nis;i++) { 393328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 393428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 393528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 393628143c3dSStefano Zampini } 393728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 393828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 393928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 394028143c3dSStefano Zampini } 3941e7931f94SStefano Zampini /* free workspace */ 394228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3943e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3944e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3945e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3946e7931f94SStefano Zampini if (isdense) { 3947e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3948e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3949e7931f94SStefano Zampini } else { 3950e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3951e7931f94SStefano Zampini } 395228143c3dSStefano Zampini if (nis) { 395328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 395428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 395528143c3dSStefano Zampini } 3956e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3957e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 395828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3959e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3960e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 396128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3962e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3963e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3964e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3965e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3966e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 396728143c3dSStefano Zampini if (nis) { 396828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 396928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 397028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 397128143c3dSStefano Zampini } 397228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 397328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 397428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 397528143c3dSStefano Zampini for (i=0;i<nis;i++) { 397628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 397728143c3dSStefano Zampini } 397853a05cb3SStefano Zampini *mat_n = NULL; 397928143c3dSStefano Zampini } 3980e7931f94SStefano Zampini PetscFunctionReturn(0); 3981e7931f94SStefano Zampini } 3982a57a6d2fSStefano Zampini 398312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3984af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 398512edc857SStefano Zampini 3986c8587f34SStefano Zampini #undef __FUNCT__ 3987c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3988c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3989c8587f34SStefano Zampini { 3990c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3991c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 399220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39939881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 399420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39956e683305SStefano Zampini IS coarse_is,*isarray; 39966e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39976e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3998f9eb5b7dSStefano Zampini PC pc_temp; 3999c8587f34SStefano Zampini PCType coarse_pc_type; 4000c8587f34SStefano Zampini KSPType coarse_ksp_type; 4001f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 40024f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 40036e683305SStefano Zampini Mat t_coarse_mat_is; 40046e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 40056e683305SStefano Zampini PetscMPIInt all_procs; 400674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 400768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 400822bc73bbSStefano Zampini PetscScalar *array; 40099881197aSStefano Zampini PetscErrorCode ierr; 4010fdc09c96SStefano Zampini 4011c8587f34SStefano Zampini PetscFunctionBegin; 4012c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 401368457ee5SStefano 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 */ 4014fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 40155a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 4016fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 4017f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 4018f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 4019f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 4020fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 4021fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 4022727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 4023fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4024fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 4025fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4026f4ddd8eeSStefano Zampini } 4027fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 4028fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 4029f4ddd8eeSStefano Zampini } 403070cf5478SStefano Zampini /* reset any subassembling information */ 403170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 40326e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 40336e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 4034fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 4035f4ddd8eeSStefano Zampini } 4036c8587f34SStefano Zampini 40376e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 40382b510759SStefano Zampini im_active = !!(pcis->n); 40392b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 40406e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 40416e683305SStefano Zampini void_procs = all_procs-active_procs; 40426e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 404374e2c79eSStefano Zampini redist = PETSC_FALSE; 404422bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 40456e683305SStefano Zampini csin_ml = PETSC_TRUE; 40466e683305SStefano Zampini ncoarse_ml = void_procs; 4047779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 4048779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 404918a45a71SStefano Zampini csin_ds = PETSC_TRUE; 405018a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 405118a45a71SStefano Zampini redist = PETSC_TRUE; 405218a45a71SStefano Zampini } else { 40536e683305SStefano Zampini csin_ds = PETSC_TRUE; 4054779c1cceSStefano Zampini ncoarse_ds = active_procs; 4055779c1cceSStefano Zampini redist = PETSC_TRUE; 405618a45a71SStefano Zampini } 40576e683305SStefano Zampini } else { 40586e683305SStefano Zampini csin_ml = PETSC_FALSE; 40596e683305SStefano Zampini ncoarse_ml = all_procs; 40606e683305SStefano Zampini if (void_procs) { 40616e683305SStefano Zampini csin_ds = PETSC_TRUE; 40626e683305SStefano Zampini ncoarse_ds = void_procs; 40636e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 40646e683305SStefano Zampini } else { 4065779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 406674e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 406774e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 406874e2c79eSStefano Zampini redist = PETSC_TRUE; 406974e2c79eSStefano Zampini } else { 40706e683305SStefano Zampini csin_ds = PETSC_FALSE; 40716e683305SStefano Zampini ncoarse_ds = all_procs; 40726e683305SStefano Zampini } 40736e683305SStefano Zampini } 407474e2c79eSStefano Zampini } 40756e683305SStefano Zampini 40766e683305SStefano Zampini /* 40776e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40786e683305SStefano Zampini - we have not exceeded the number of levels requested 40796e683305SStefano Zampini - we can actually subassemble the active processes 40806e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40816e683305SStefano Zampini */ 40826e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40836e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40846e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40856e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40866e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4087f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40882b510759SStefano Zampini } else { 4089f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4090c8587f34SStefano Zampini } 4091c8587f34SStefano Zampini } 40926e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40936e683305SStefano Zampini if (multilevel_allowed) { 40946e683305SStefano Zampini ncoarse = ncoarse_ml; 40956e683305SStefano Zampini csin = csin_ml; 409658da7f69SStefano Zampini redist = PETSC_FALSE; 40976e683305SStefano Zampini } else { 40986e683305SStefano Zampini ncoarse = ncoarse_ds; 40996e683305SStefano Zampini csin = csin_ds; 41006e683305SStefano Zampini } 4101e7931f94SStefano Zampini 4102abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4103abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4104abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4105abbbba34SStefano Zampini 4106abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 410722bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 410822bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 410922bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 411022bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4111b9b85e73SStefano Zampini #if 0 4112b9b85e73SStefano Zampini { 4113b9b85e73SStefano Zampini PetscViewer viewer; 4114b9b85e73SStefano Zampini char filename[256]; 4115b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4116b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4117b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4118b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4119b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4120b9b85e73SStefano Zampini } 4121b9b85e73SStefano Zampini #endif 41226e683305SStefano 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); 41236e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 41246e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 41256e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4126abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4127abbbba34SStefano Zampini 41286e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 41296e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 41306e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 41316e683305SStefano Zampini const PetscInt *idxs; 41326e683305SStefano Zampini ISLocalToGlobalMapping tmap; 41336e683305SStefano Zampini 41346e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 41350be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 41366e683305SStefano Zampini /* allocate space for temporary storage */ 4137854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4138854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 41396e683305SStefano Zampini /* allocate for IS array */ 41406e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 41416e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 41426e683305SStefano Zampini nis = nisdofs + nisneu; 4143854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 41446e683305SStefano Zampini /* dofs splitting */ 41456e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 41466e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 41476e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 41486e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41496e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41506e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 41516e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41526e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 41536e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 41546e683305SStefano Zampini } 41556e683305SStefano Zampini /* neumann boundaries */ 41566e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 41576e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 41586e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 41596e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41606e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 41616e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 41626e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 41636e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 41646e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 41656e683305SStefano Zampini } 41666e683305SStefano Zampini /* free memory */ 41676e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 41686e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 41696e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 41706e683305SStefano Zampini } else { 41716e683305SStefano Zampini nis = 0; 41726e683305SStefano Zampini nisdofs = 0; 41736e683305SStefano Zampini nisneu = 0; 41746e683305SStefano Zampini isarray = NULL; 41756e683305SStefano Zampini } 41766e683305SStefano Zampini /* destroy no longer needed map */ 41776e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41786e683305SStefano Zampini 41796e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41806e683305SStefano Zampini coarse_mat_is = NULL; 41816e683305SStefano Zampini if (csin) { 41826e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 418374e2c79eSStefano Zampini if (redist) { 418474e2c79eSStefano Zampini PetscMPIInt rank; 4185779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 418674e2c79eSStefano Zampini 418774e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 418858da7f69SStefano Zampini spc = active_procs/ncoarse; 418958da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4190779c1cceSStefano Zampini if (im_active) { 4191779c1cceSStefano Zampini destsize = 1; 419274e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 419374e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 419474e2c79eSStefano Zampini } else { 419574e2c79eSStefano Zampini dest[0] = rank/(spc+1); 419674e2c79eSStefano Zampini } 419774e2c79eSStefano Zampini } else { 4198779c1cceSStefano Zampini destsize = 0; 41996e683305SStefano Zampini } 4200779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4201779c1cceSStefano Zampini } else if (csin_type_simple) { 42026e683305SStefano Zampini PetscMPIInt rank; 42036e683305SStefano Zampini PetscInt issize,isidx; 4204779c1cceSStefano Zampini 42056e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 42066e683305SStefano Zampini if (im_active) { 42076e683305SStefano Zampini issize = 1; 42086e683305SStefano Zampini isidx = (PetscInt)rank; 42096e683305SStefano Zampini } else { 42106e683305SStefano Zampini issize = 0; 42116e683305SStefano Zampini isidx = -1; 42126e683305SStefano Zampini } 42136e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4214779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 4215b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 42166e683305SStefano Zampini } 4217779c1cceSStefano Zampini 4218779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4219779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4220779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4221779c1cceSStefano Zampini PetscInt *coarse_candidates; 4222779c1cceSStefano Zampini const PetscInt* tisindices; 4223779c1cceSStefano Zampini 4224779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4225779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4226779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4227779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4228779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4229779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4230779c1cceSStefano Zampini } 4231779c1cceSStefano Zampini } 4232779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4233779c1cceSStefano Zampini 4234779c1cceSStefano Zampini 42356e683305SStefano Zampini if (pcbddc->dbg_flag) { 42366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 42386e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4240779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 42416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 42426e683305SStefano Zampini } 42436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 42446e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42456e683305SStefano Zampini } 42466e683305SStefano Zampini /* shift the pattern on coarse candidates */ 42476e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 42486e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4249854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 42506e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 42516e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 42526e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 42536e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 42546e683305SStefano Zampini } 42556e683305SStefano Zampini if (pcbddc->dbg_flag) { 42566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 42586e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 42596e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42606e683305SStefano Zampini } 4261779c1cceSStefano Zampini } 42626e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 426353a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 426453a05cb3SStefano 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); 426553a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 426653a05cb3SStefano 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); 426753a05cb3SStefano Zampini } 42686e683305SStefano Zampini } else { 42696e683305SStefano Zampini if (pcbddc->dbg_flag) { 42706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 42726e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42736e683305SStefano Zampini } 42746e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42756e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42766e683305SStefano Zampini } 42776e683305SStefano Zampini 42786e683305SStefano Zampini /* create local to global scatters for coarse problem */ 427968457ee5SStefano Zampini if (compute_vecs) { 42806e683305SStefano Zampini PetscInt lrows; 42816e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42826e683305SStefano Zampini if (coarse_mat_is) { 42836e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42846e683305SStefano Zampini } else { 42856e683305SStefano Zampini lrows = 0; 42866e683305SStefano Zampini } 42876e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42886e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42896e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42906e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42916e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42926e683305SStefano Zampini } 42936e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42946e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4295c8587f34SStefano Zampini 4296f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4297f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4298f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4299f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4300f9eb5b7dSStefano Zampini } else { 4301f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4302f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4303c8587f34SStefano Zampini } 4304c8587f34SStefano Zampini 43056e683305SStefano Zampini /* print some info if requested */ 43066e683305SStefano Zampini if (pcbddc->dbg_flag) { 43076e683305SStefano Zampini if (!multilevel_allowed) { 43086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43096e683305SStefano Zampini if (multilevel_requested) { 43106e683305SStefano 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); 43116e683305SStefano Zampini } else if (pcbddc->max_levels) { 43126e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 43136e683305SStefano Zampini } 43146e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 43156e683305SStefano Zampini } 43166e683305SStefano Zampini } 43176e683305SStefano Zampini 4318f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 43196e683305SStefano Zampini if (coarse_mat_is) { 43206e683305SStefano Zampini MatReuse coarse_mat_reuse; 43216a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 43226e683305SStefano Zampini if (pcbddc->dbg_flag) { 43236e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 43246e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43256e683305SStefano Zampini } 4326f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4327312be037SStefano Zampini char prefix[256],str_level[16]; 4328e604994aSStefano Zampini size_t len; 43296e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4330c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4331f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 43325f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4333c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 43346e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4335c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4336c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4337e604994aSStefano Zampini /* prefix */ 4338e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4339e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4340e604994aSStefano Zampini if (!pcbddc->current_level) { 4341e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4342e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4343c8587f34SStefano Zampini } else { 4344e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4345312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4346312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 434734d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4348312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4349e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4350e604994aSStefano Zampini } 4351e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 43523e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43533e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 43543e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 43553e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4356f9eb5b7dSStefano Zampini /* allow user customization */ 4357f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 43583e3c6dadSStefano Zampini } 43593e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 43603e3c6dadSStefano Zampini if (nisdofs) { 43613e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 43623e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 43633e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 43643e3c6dadSStefano Zampini } 43653e3c6dadSStefano Zampini } 43663e3c6dadSStefano Zampini if (nisneu) { 43673e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 43683e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4369312be037SStefano Zampini } 4370f9eb5b7dSStefano Zampini 4371f9eb5b7dSStefano Zampini /* get some info after set from options */ 4372f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4373f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4374f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43754f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43766e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4377f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4378f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4379f9eb5b7dSStefano Zampini } 4380*39f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 43814f3a063dSStefano Zampini if (isredundant) { 43824f3a063dSStefano Zampini KSP inner_ksp; 43834f3a063dSStefano Zampini PC inner_pc; 43844f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43854f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43864f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43874f3a063dSStefano Zampini } 4388f9eb5b7dSStefano Zampini 4389f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4390fa7f1dd8SStefano Zampini if (coarse_reuse) { 439181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4392fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43936e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4394fa7f1dd8SStefano Zampini } else { 43956e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4396fa7f1dd8SStefano Zampini } 4397c8587f34SStefano Zampini if (isbddc || isnn) { 439822bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 439970cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 4400b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 440122b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 44026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 44036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 44046e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 44056e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 440622b6e8a2SStefano Zampini } 440770cf5478SStefano Zampini } 440853a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 440970cf5478SStefano Zampini } else { 441022bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 441122bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 441222bc73bbSStefano Zampini } 441322bc73bbSStefano Zampini } else { 44142e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4415c8587f34SStefano Zampini } 4416c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4417c8587f34SStefano Zampini 44183301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 44195a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 44203301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 44213301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 44223301b35fSStefano Zampini } 44233301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 44243301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 44253301b35fSStefano Zampini } 44263301b35fSStefano Zampini if (pc->pmat->spd_set) { 44273301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 44283301b35fSStefano Zampini } 44296e683305SStefano Zampini /* set operators */ 44305f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 44316e683305SStefano Zampini if (pcbddc->dbg_flag) { 44326e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 44336e683305SStefano Zampini } 44346e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 44356e683305SStefano Zampini coarse_mat = 0; 44366e683305SStefano Zampini } 44376e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4438b9b85e73SStefano Zampini #if 0 4439b9b85e73SStefano Zampini { 4440b9b85e73SStefano Zampini PetscViewer viewer; 4441b9b85e73SStefano Zampini char filename[256]; 4442b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4443b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4444b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4445b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4446b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4447b9b85e73SStefano Zampini } 4448b9b85e73SStefano Zampini #endif 4449c8587f34SStefano Zampini 4450c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4451c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4452c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 445398a51de6SStefano Zampini } 445498a51de6SStefano Zampini 445598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 445698a51de6SStefano Zampini Vec crhs,csol; 445798a51de6SStefano Zampini PetscBool ispreonly; 445804708bb6SStefano Zampini 445998a51de6SStefano Zampini if (CoarseNullSpace) { 4460c8587f34SStefano Zampini if (isbddc) { 4461c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4462c8587f34SStefano Zampini } else { 4463c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4464c8587f34SStefano Zampini } 4465c8587f34SStefano Zampini } 4466f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4467f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4468f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4469f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 44706e683305SStefano Zampini /* hack */ 4471f347579bSStefano Zampini if (!csol) { 44722a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4473f9eb5b7dSStefano Zampini } 4474f347579bSStefano Zampini if (!crhs) { 44752a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4476f347579bSStefano Zampini } 4477cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4478cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 44796e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4480c8587f34SStefano Zampini KSP check_ksp; 44812b510759SStefano Zampini KSPType check_ksp_type; 4482c8587f34SStefano Zampini PC check_pc; 44836e683305SStefano Zampini Vec check_vec,coarse_vec; 44846a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44852b510759SStefano Zampini PetscInt its; 44866e683305SStefano Zampini PetscBool compute_eigs; 44876e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44886e683305SStefano Zampini PetscInt neigs; 44898e185a42SStefano Zampini const char *prefix; 4490c8587f34SStefano Zampini 44912b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44926e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 449323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4494f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44952b510759SStefano Zampini if (ispreonly) { 44962b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44976e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44982b510759SStefano Zampini } else { 4499cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 45006e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4501c8587f34SStefano Zampini } 4502c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 45036e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 45046e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 45056e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4506a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4507a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4508a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4509a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4510c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4511c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4512c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4513c8587f34SStefano Zampini /* create random vec */ 45146e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 45156e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4516c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4517c8587f34SStefano Zampini if (CoarseNullSpace) { 4518c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4519c8587f34SStefano Zampini } 45206e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4521c8587f34SStefano Zampini /* solve coarse problem */ 45226e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4523c8587f34SStefano Zampini if (CoarseNullSpace) { 45246e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4525c8587f34SStefano Zampini } 4526cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 45276e683305SStefano Zampini if (compute_eigs) { 4528854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4529854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 45306e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 45316e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 45326e683305SStefano Zampini lambda_min = eigs_r[0]; 45336e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 45346e683305SStefano Zampini if (lambda_max>lambda_min) { 4535cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4536cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4537cbcc2c2aSStefano Zampini } 4538c8587f34SStefano Zampini } 4539c8587f34SStefano Zampini } 4540cbcc2c2aSStefano Zampini 4541c8587f34SStefano Zampini /* check coarse problem residual error */ 45426e683305SStefano Zampini if (pcbddc->dbg_flag) { 45436e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 45446e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45456e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4546c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 45476e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 45486e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4549c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4550779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 45516e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 45526e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 45536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 45546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 45556e683305SStefano Zampini if (compute_eigs) { 45566e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4557deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4558c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 45596e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 45606e683305SStefano 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); 45616e683305SStefano Zampini for (i=0;i<neigs;i++) { 45626e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4563c8587f34SStefano Zampini } 45646e683305SStefano Zampini } 45656e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 45666e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 45676e683305SStefano Zampini } 4568c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 45696e683305SStefano Zampini if (compute_eigs) { 45706e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 45716e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4572c8587f34SStefano Zampini } 45736e683305SStefano Zampini } 45746e683305SStefano Zampini } 4575cbcc2c2aSStefano Zampini /* print additional info */ 4576cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 45776e683305SStefano Zampini /* waits until all processes reaches this point */ 45786e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4579cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4580cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4581cbcc2c2aSStefano Zampini } 4582cbcc2c2aSStefano Zampini 45832b510759SStefano Zampini /* free memory */ 4584c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4585fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4586c8587f34SStefano Zampini PetscFunctionReturn(0); 4587c8587f34SStefano Zampini } 4588674ae819SStefano Zampini 4589f34684f1SStefano Zampini #undef __FUNCT__ 4590f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4591f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4592f34684f1SStefano Zampini { 4593f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4594f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4595f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 4596dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 4597dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 459873be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 4599dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 4600f34684f1SStefano Zampini PetscErrorCode ierr; 4601f34684f1SStefano Zampini 4602f34684f1SStefano Zampini PetscFunctionBegin; 4603f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 46040e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 46050e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4606727cdba6SStefano Zampini } 4607dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 4608dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,subset_n,&subset);CHKERRQ(ierr); 4609dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4610dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 4611dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 4612dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 4613dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 4614dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 46150e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 46160e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 46170e6343abSStefano Zampini } 4618dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 4619dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4620dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 4621dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 4622dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 4623f34684f1SStefano Zampini 4624f34684f1SStefano Zampini /* check numbering */ 4625f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4626f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4627dc456d91SStefano Zampini PetscInt i; 4628b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4629f34684f1SStefano Zampini 4630f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4631f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4632f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 46330fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4634f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4635f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4636727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4637f34684f1SStefano Zampini } 4638f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4639f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4640f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4641f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4642f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4643f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4644f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4645f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4646f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4647f34684f1SStefano Zampini if (array[i] == 1.0) { 4648b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4649f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4650f34684f1SStefano Zampini } 4651f34684f1SStefano Zampini } 4652b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4653f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4654f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4655f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4656f34684f1SStefano Zampini } 4657f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4658f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4659f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4660f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4661f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4662f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4663b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4664ca8b9ea9SStefano Zampini PetscInt *gidxs; 4665ca8b9ea9SStefano Zampini 4666ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4667ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4668f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4669f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4670f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4671f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 46724bc2dc4bSStefano 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); 4673f34684f1SStefano Zampini } 4674f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4675ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4676f34684f1SStefano Zampini } 4677f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4678302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4679f34684f1SStefano Zampini } 46808bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4681f34684f1SStefano Zampini /* get back data */ 4682f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4683f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4684674ae819SStefano Zampini PetscFunctionReturn(0); 4685674ae819SStefano Zampini } 4686674ae819SStefano Zampini 4687e456f2a8SStefano Zampini #undef __FUNCT__ 4688e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4689a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4690e456f2a8SStefano Zampini { 4691e456f2a8SStefano Zampini IS localis_t; 4692a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4693e456f2a8SStefano Zampini PetscScalar *vals; 4694e456f2a8SStefano Zampini PetscErrorCode ierr; 4695e456f2a8SStefano Zampini 4696e456f2a8SStefano Zampini PetscFunctionBegin; 4697a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4698e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4699854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4700e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4701e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4702a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4703a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 47041035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4705a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 47061035eff8SStefano Zampini } 4707a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4708e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4709e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4710a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4711a7dc3881SStefano Zampini /* now compute set in local ordering */ 4712a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4713a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4714a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4715a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4716a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4717ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4718e456f2a8SStefano Zampini lsize++; 4719e456f2a8SStefano Zampini } 4720e456f2a8SStefano Zampini } 4721854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4722a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4723ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4724e456f2a8SStefano Zampini idxs[lsize++] = i; 4725e456f2a8SStefano Zampini } 4726e456f2a8SStefano Zampini } 4727a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4728a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4729e456f2a8SStefano Zampini *localis = localis_t; 4730e456f2a8SStefano Zampini PetscFunctionReturn(0); 4731e456f2a8SStefano Zampini } 4732906d46d4SStefano Zampini 4733906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4734906d46d4SStefano Zampini #undef __FUNCT__ 4735906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4736906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4737906d46d4SStefano Zampini { 4738906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4739906d46d4SStefano Zampini PetscErrorCode ierr; 4740906d46d4SStefano Zampini 4741906d46d4SStefano Zampini PetscFunctionBegin; 4742906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4743906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4744906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4745906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4746906d46d4SStefano Zampini PetscFunctionReturn(0); 4747906d46d4SStefano Zampini } 4748906d46d4SStefano Zampini 4749906d46d4SStefano Zampini #undef __FUNCT__ 4750906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4751906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4752906d46d4SStefano Zampini { 4753906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4754906d46d4SStefano Zampini PetscErrorCode ierr; 4755906d46d4SStefano Zampini 4756906d46d4SStefano Zampini PetscFunctionBegin; 4757906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4758906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4759906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4760906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4761906d46d4SStefano Zampini PetscFunctionReturn(0); 4762906d46d4SStefano Zampini } 4763b96c3477SStefano Zampini 4764b96c3477SStefano Zampini #undef __FUNCT__ 4765b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 476608122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4767b96c3477SStefano Zampini { 4768a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4769b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4770b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4771a64f4aa4SStefano Zampini Mat S_j; 4772b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4773b96c3477SStefano Zampini PetscBool free_used_adj; 4774b96c3477SStefano Zampini PetscErrorCode ierr; 4775b96c3477SStefano Zampini 4776b96c3477SStefano Zampini PetscFunctionBegin; 4777b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4778b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 477908122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4780b96c3477SStefano Zampini used_xadj = NULL; 4781b96c3477SStefano Zampini used_adjncy = NULL; 4782b96c3477SStefano Zampini } else { 478308122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 478408122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 478508122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 478608122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4787b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4788b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4789b96c3477SStefano Zampini } else { 47902fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4791b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4792b96c3477SStefano Zampini PetscInt nvtxs; 4793b96c3477SStefano Zampini 47942fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47952fffb893SStefano Zampini if (flg_row) { 4796b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4797b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4798b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4799b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 48002fffb893SStefano Zampini } else { 48012fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 48022fffb893SStefano Zampini used_xadj = NULL; 48032fffb893SStefano Zampini used_adjncy = NULL; 48042fffb893SStefano Zampini } 48052fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4806b96c3477SStefano Zampini } 4807b96c3477SStefano Zampini } 4808d5574798SStefano Zampini 4809d5574798SStefano Zampini /* setup sub_schurs data */ 4810a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4811a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4812a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4813a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 4814862806e4SStefano 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); 4815a64f4aa4SStefano Zampini } else { 48166816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 481704708bb6SStefano Zampini PetscBool isseqaij; 48185feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 48195feab87aSStefano Zampini PetscInt n_vertices; 48205feab87aSStefano Zampini 48215feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 48222034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 48235feab87aSStefano Zampini } 482404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 482504708bb6SStefano Zampini if (!isseqaij) { 482604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 482704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 482804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 482904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483004708bb6SStefano Zampini } else { 483104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 483204708bb6SStefano Zampini } 483304708bb6SStefano Zampini } 4834862806e4SStefano 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); 4835a64f4aa4SStefano Zampini } 4836a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4837b96c3477SStefano Zampini 4838b96c3477SStefano Zampini /* free adjacency */ 4839b96c3477SStefano Zampini if (free_used_adj) { 4840b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4841b96c3477SStefano Zampini } 4842b96c3477SStefano Zampini PetscFunctionReturn(0); 4843b96c3477SStefano Zampini } 4844b96c3477SStefano Zampini 4845b96c3477SStefano Zampini #undef __FUNCT__ 4846b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 484708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4848b96c3477SStefano Zampini { 4849b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4850b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4851b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4852b96c3477SStefano Zampini PCBDDCGraph graph; 4853b96c3477SStefano Zampini PetscErrorCode ierr; 4854b96c3477SStefano Zampini 4855b96c3477SStefano Zampini PetscFunctionBegin; 4856b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 485708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 48583301b35fSStefano Zampini IS verticesIS,verticescomm; 48593301b35fSStefano Zampini PetscInt vsize,*idxs; 4860b96c3477SStefano Zampini 4861b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 48623301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 48633301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48643301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 48653301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 48663301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4867b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 48687fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 48693301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 48703301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 4871b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4872b96c3477SStefano Zampini /* 4873b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4874b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4875b96c3477SStefano Zampini } 4876b96c3477SStefano Zampini */ 4877b96c3477SStefano Zampini } else { 4878b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4879b96c3477SStefano Zampini } 4880b96c3477SStefano Zampini 4881b96c3477SStefano Zampini /* sub_schurs init */ 4882a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4883a64f4aa4SStefano Zampini 4884b96c3477SStefano Zampini /* free graph struct */ 488508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4886b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4887b96c3477SStefano Zampini } 4888b96c3477SStefano Zampini PetscFunctionReturn(0); 4889b96c3477SStefano Zampini } 4890