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 if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 5108122e43SStefano Zampini PetscInt subset_size; 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 5708122e43SStefano Zampini /* min/max and threshold */ 5808122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 59f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 6008122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 61f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 62f6f667cfSStefano Zampini if (nmin) { 63f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 64f6f667cfSStefano Zampini } 6508122e43SStefano Zampini 6608122e43SStefano Zampini /* allocate lapack workspace */ 6708122e43SStefano Zampini cum = cum2 = 0; 6808122e43SStefano Zampini maxneigs = 0; 6908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 7108122e43SStefano Zampini PetscInt n,subset_size; 72f6f667cfSStefano Zampini 7308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 759162d606SStefano Zampini cum += subset_size; 769162d606SStefano Zampini cum2 += subset_size*n; 7708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 7808122e43SStefano Zampini } 7908122e43SStefano Zampini } 8008122e43SStefano Zampini if (mss) { 819ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 8208122e43SStefano Zampini PetscBLASInt B_itype = 1; 8308122e43SStefano Zampini PetscBLASInt B_N = mss; 844c6709b3SStefano Zampini PetscReal zero = 0.0; 854c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 8608122e43SStefano Zampini 8708122e43SStefano Zampini B_lwork = -1; 8808122e43SStefano Zampini S = NULL; 8908122e43SStefano Zampini St = NULL; 90a58a30b4SStefano Zampini eigs = NULL; 91a58a30b4SStefano Zampini eigv = NULL; 92a58a30b4SStefano Zampini B_iwork = NULL; 93a58a30b4SStefano Zampini B_ifail = NULL; 948bec7fa6SStefano Zampini thresh = 1.0; 9508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 9608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 9708122e43SStefano 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)); 9808122e43SStefano Zampini #else 9908122e43SStefano 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)); 10008122e43SStefano Zampini #endif 10108122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 10208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 10308122e43SStefano Zampini } else { 10408122e43SStefano Zampini /* TODO */ 10508122e43SStefano Zampini } 10608122e43SStefano Zampini } else { 10708122e43SStefano Zampini lwork = 0; 10808122e43SStefano Zampini } 10908122e43SStefano Zampini 11008122e43SStefano Zampini nv = 0; 111d62866d3SStefano 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) */ 112d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 11308122e43SStefano Zampini } 1144c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 115f6f667cfSStefano Zampini if (allocated_S_St) { 116f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 117f6f667cfSStefano Zampini } 118f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 11908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 12008122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 12108122e43SStefano Zampini #endif 1229162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 1239162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 1249162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 12508122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 1269162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 12708122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 12808122e43SStefano Zampini 12908122e43SStefano Zampini maxneigs = 0; 13008122e43SStefano Zampini cum = cum2 = cumarray = 0; 1319162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 1329162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 133d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 13408122e43SStefano Zampini const PetscInt *idxs; 13508122e43SStefano Zampini 136d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 13708122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 13808122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 13908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 14008122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 1419162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 1429162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 14308122e43SStefano Zampini } 14408122e43SStefano Zampini cum2 = cum; 145d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 14608122e43SStefano Zampini } 14708122e43SStefano Zampini 14808122e43SStefano Zampini if (mss) { /* multilevel */ 14908122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 15008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 15108122e43SStefano Zampini } 15208122e43SStefano Zampini 15308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15408122e43SStefano Zampini PetscInt j,subset_size; 15508122e43SStefano Zampini 15608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 15708122e43SStefano Zampini if (PetscBTLookup(sub_schurs->computed_Stilda_subs,i)) { 15808122e43SStefano Zampini const PetscInt *idxs; 159f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 160f6f667cfSStefano Zampini PetscInt eigs_start = 0; 16108122e43SStefano Zampini PetscBLASInt B_N; 16208122e43SStefano Zampini 163f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 164f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 1659ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 16608122e43SStefano Zampini PetscInt j; 16708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 16808122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 16908122e43SStefano Zampini } 17008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 17108122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 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 /* Threshold: this is an heuristic for edges */ 183f6f667cfSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 184f6f667cfSStefano Zampini thresh = pcbddc->mat_graph->count[idxs[0]]*pcbddc->adaptive_threshold; 185f6f667cfSStefano Zampini 1869ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 18708122e43SStefano Zampini PetscBLASInt B_itype = 1; 188f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 1894c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 1909552c7c7SStefano Zampini PetscInt nmin_s; 19108122e43SStefano Zampini 192f6f667cfSStefano Zampini /* ask for eigenvalues larger than thresh */ 193fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1948bec7fa6SStefano 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]]); 195fd14bc51SStefano Zampini } 19608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 19708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 198f6f667cfSStefano 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)); 19908122e43SStefano Zampini #else 200f6f667cfSStefano 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)); 20108122e43SStefano Zampini #endif 20208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 20308122e43SStefano Zampini if (B_ierr) { 20408122e43SStefano Zampini if (B_ierr < 0 ) { 20508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 20608122e43SStefano Zampini } else if (B_ierr <= B_N) { 20708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 20808122e43SStefano Zampini } else { 2099552c7c7SStefano 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); 21008122e43SStefano Zampini } 21108122e43SStefano Zampini } 21208122e43SStefano Zampini 21308122e43SStefano Zampini if (B_neigs > nmax) { 214fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 215fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 216fd14bc51SStefano Zampini } 217f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 21808122e43SStefano Zampini B_neigs = nmax; 21908122e43SStefano Zampini } 22008122e43SStefano Zampini 2219552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 2229552c7c7SStefano Zampini if (B_neigs < nmin_s) { 22308122e43SStefano Zampini PetscBLASInt B_neigs2; 22408122e43SStefano Zampini 225f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 226f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 227fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 228fd14bc51SStefano 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); 229fd14bc51SStefano Zampini } 2309ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 23108122e43SStefano Zampini PetscInt j; 23208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23408122e43SStefano Zampini } 23508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 23608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 23708122e43SStefano Zampini } 23808122e43SStefano Zampini } else { 23908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 24108122e43SStefano Zampini } 24208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 24308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 244f6f667cfSStefano 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)); 24508122e43SStefano Zampini #else 246f6f667cfSStefano 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)); 24708122e43SStefano Zampini #endif 24808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 24908122e43SStefano Zampini B_neigs += B_neigs2; 25008122e43SStefano Zampini } 25108122e43SStefano Zampini if (B_ierr) { 25208122e43SStefano Zampini if (B_ierr < 0 ) { 25308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 25408122e43SStefano Zampini } else if (B_ierr <= B_N) { 25508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 25608122e43SStefano Zampini } else { 2579552c7c7SStefano 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); 25808122e43SStefano Zampini } 25908122e43SStefano Zampini } 260fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 261fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs); 26208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 26308122e43SStefano Zampini if (eigs[j] == 0.0) { 264fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n"); 26508122e43SStefano Zampini } else { 266f6f667cfSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]); 267fd14bc51SStefano Zampini } 26808122e43SStefano Zampini } 26908122e43SStefano Zampini } 27008122e43SStefano Zampini } else { 27108122e43SStefano Zampini /* TODO */ 27208122e43SStefano Zampini } 27308122e43SStefano Zampini 2748bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 2758bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 2769162d606SStefano Zampini if (B_neigs) { 2779162d606SStefano 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); 278fd14bc51SStefano Zampini 279fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 2809552c7c7SStefano Zampini PetscInt ii; 2819552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 282fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N); 2839552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 2849162d606SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 2859552c7c7SStefano Zampini } 2869552c7c7SStefano Zampini } 287fd14bc51SStefano Zampini } 28808122e43SStefano Zampini #if 0 2899162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 29008122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 29108122e43SStefano Zampini PetscScalar norm; 29208122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 2939162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 2949162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 29508122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 29608122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 29708122e43SStefano Zampini } else { 29808122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 29908122e43SStefano Zampini } 3009162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 301b1b3d7a2SStefano Zampini } 302b1b3d7a2SStefano Zampini #endif 3039162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 3049162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 3059162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 3069162d606SStefano Zampini cum++; 30708122e43SStefano Zampini } 30808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 30908122e43SStefano Zampini } 31008122e43SStefano Zampini /* shift for next computation */ 31108122e43SStefano Zampini cumarray += subset_size*subset_size; 31208122e43SStefano Zampini } 313fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 314fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 315fd14bc51SStefano Zampini } 31608122e43SStefano Zampini 31708122e43SStefano Zampini if (mss) { 31808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 31908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 320f6f667cfSStefano Zampini /* destroy matrices (junk) */ 321f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 322f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 32308122e43SStefano Zampini } 324f6f667cfSStefano Zampini if (allocated_S_St) { 325f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 326f6f667cfSStefano Zampini } 327f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 32808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 32908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 33008122e43SStefano Zampini #endif 33108122e43SStefano Zampini if (pcbddc->dbg_flag) { 3321b968477SStefano Zampini PetscInt maxneigs_r; 33308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3349b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 33508122e43SStefano Zampini } 33608122e43SStefano Zampini PetscFunctionReturn(0); 33708122e43SStefano Zampini } 338b1b3d7a2SStefano Zampini 339674ae819SStefano Zampini #undef __FUNCT__ 340c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 341c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 342c8587f34SStefano Zampini { 343c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3448629588bSStefano Zampini PetscScalar *coarse_submat_vals; 345c8587f34SStefano Zampini PetscErrorCode ierr; 346c8587f34SStefano Zampini 347c8587f34SStefano Zampini PetscFunctionBegin; 348f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 3495e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 350c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 351c8587f34SStefano Zampini 352684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 3530fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 354684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 355c8587f34SStefano Zampini 356c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 357b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 358c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 359c8587f34SStefano Zampini } 360c8587f34SStefano Zampini 3618629588bSStefano Zampini /* 3628629588bSStefano Zampini Setup local correction and local part of coarse basis. 3638629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 3648629588bSStefano Zampini */ 36547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 3668629588bSStefano Zampini 3678629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 3688629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 3698629588bSStefano Zampini 3708629588bSStefano Zampini /* free */ 3718629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 372c8587f34SStefano Zampini PetscFunctionReturn(0); 373c8587f34SStefano Zampini } 374c8587f34SStefano Zampini 375c8587f34SStefano Zampini #undef __FUNCT__ 376674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 377674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 378674ae819SStefano Zampini { 379674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 380674ae819SStefano Zampini PetscErrorCode ierr; 381674ae819SStefano Zampini 382674ae819SStefano Zampini PetscFunctionBegin; 383674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 384674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 385674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 386674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 387785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 388674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 389f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 390f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 391785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 39263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 39363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 394674ae819SStefano Zampini PetscFunctionReturn(0); 395674ae819SStefano Zampini } 396674ae819SStefano Zampini 397674ae819SStefano Zampini #undef __FUNCT__ 398674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 399674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 400674ae819SStefano Zampini { 401674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 402674ae819SStefano Zampini PetscErrorCode ierr; 403674ae819SStefano Zampini 404674ae819SStefano Zampini PetscFunctionBegin; 405b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 406674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 407674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 408674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 409b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 410674ae819SStefano Zampini PetscFunctionReturn(0); 411674ae819SStefano Zampini } 412674ae819SStefano Zampini 413674ae819SStefano Zampini #undef __FUNCT__ 414674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 415674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 416674ae819SStefano Zampini { 417674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 41806656605SStefano Zampini PetscScalar *array; 419674ae819SStefano Zampini PetscErrorCode ierr; 420674ae819SStefano Zampini 421674ae819SStefano Zampini PetscFunctionBegin; 422674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42358da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 42406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 42506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 42658da7f69SStefano Zampini } 427674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 428674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 42915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 43015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 431674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 432674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 433674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 43406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 435674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 436674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4378ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 438674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 439674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 440674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 441f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 442f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 443f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 444f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 445727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 4460e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 447f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 44870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 4496e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 450674ae819SStefano Zampini PetscFunctionReturn(0); 451674ae819SStefano Zampini } 452674ae819SStefano Zampini 453674ae819SStefano Zampini #undef __FUNCT__ 454f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 455f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 4566bfb1811SStefano Zampini { 4576bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4586bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4596bfb1811SStefano Zampini VecType impVecType; 460e9189074SStefano Zampini PetscInt n_constraints,n_R,old_size; 4616bfb1811SStefano Zampini PetscErrorCode ierr; 4626bfb1811SStefano Zampini 4636bfb1811SStefano Zampini PetscFunctionBegin; 464f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 465f4ddd8eeSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 466f4ddd8eeSStefano Zampini } 467e7b262bdSStefano Zampini /* get sizes */ 468b371cd4fSStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->n_vertices; 469b371cd4fSStefano Zampini n_R = pcis->n-pcbddc->n_vertices; 4706bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 471e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 472e7b262bdSStefano Zampini /* R nodes */ 473e7b262bdSStefano Zampini old_size = -1; 474e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 475e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 476e7b262bdSStefano Zampini } 477e7b262bdSStefano Zampini if (n_R != old_size) { 478e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 479e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 4806bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 4816bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 4826bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 4836bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 484e7b262bdSStefano Zampini } 485e7b262bdSStefano Zampini /* local primal dofs */ 486e7b262bdSStefano Zampini old_size = -1; 487e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 488e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 489e7b262bdSStefano Zampini } 490e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 491e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 49283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 493e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 4946bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 495e7b262bdSStefano Zampini } 496e7b262bdSStefano Zampini /* local explicit constraints */ 497e7b262bdSStefano Zampini old_size = -1; 498e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 499e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 500e7b262bdSStefano Zampini } 501e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 502e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 50383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 50483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 50583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 50683b7ccabSStefano Zampini } 5076bfb1811SStefano Zampini PetscFunctionReturn(0); 5086bfb1811SStefano Zampini } 5096bfb1811SStefano Zampini 5106bfb1811SStefano Zampini #undef __FUNCT__ 51147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 51247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 51388ebb749SStefano Zampini { 51425084f0cSStefano Zampini PetscErrorCode ierr; 51525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 51688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 51788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 518d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 51925084f0cSStefano Zampini /* submatrices of local problem */ 52080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 52106656605SStefano Zampini /* submatrices of local coarse problem */ 52206656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 52325084f0cSStefano Zampini /* working matrices */ 52406656605SStefano Zampini Mat C_CR; 52525084f0cSStefano Zampini /* additional working stuff */ 52606656605SStefano Zampini PC pc_R; 52706656605SStefano Zampini Mat F; 52806656605SStefano Zampini PetscBool isLU,isCHOL,isILU; 52906656605SStefano Zampini 53025084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 53106656605SStefano Zampini PetscScalar *work; 53206656605SStefano Zampini PetscInt *idx_V_B; 53306656605SStefano Zampini PetscInt n,n_vertices,n_constraints; 53406656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 535b9d89cd5SStefano Zampini PetscBool unsymmetric_check; 53645a1bb75SStefano Zampini /* matrix type (vector type propagated downstream from vec1_C and local matrix type) */ 53788ebb749SStefano Zampini MatType impMatType; 53825084f0cSStefano Zampini /* some shortcuts to scalars */ 53906656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 54088ebb749SStefano Zampini 54188ebb749SStefano Zampini PetscFunctionBegin; 542b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 54388ebb749SStefano Zampini n_constraints = pcbddc->local_primal_size-n_vertices; 54488ebb749SStefano Zampini /* Set Non-overlapping dimensions */ 545b371cd4fSStefano Zampini n_B = pcis->n_B; 546b371cd4fSStefano Zampini n_D = pcis->n - n_B; 54788ebb749SStefano Zampini n_R = pcis->n - n_vertices; 54888ebb749SStefano Zampini 54988ebb749SStefano Zampini /* Set types for local objects needed by BDDC precondtioner */ 55088ebb749SStefano Zampini impMatType = MATSEQDENSE; 55188ebb749SStefano Zampini 55288ebb749SStefano Zampini /* vertices in boundary numbering */ 553785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 5540e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 55588ebb749SStefano Zampini if (i != n_vertices) { 55622d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 55788ebb749SStefano Zampini } 55888ebb749SStefano Zampini 55906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 56006656605SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 56106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 56206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 56306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 56406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 56506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 56606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 56706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 56806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 56906656605SStefano Zampini 57006656605SStefano Zampini unsymmetric_check = PETSC_FALSE; 57106656605SStefano Zampini /* allocate workspace */ 57206656605SStefano Zampini n = 0; 57306656605SStefano Zampini if (n_constraints) { 57406656605SStefano Zampini n += n_R*n_constraints; 57506656605SStefano Zampini } 57606656605SStefano Zampini if (n_vertices) { 57706656605SStefano Zampini n = PetscMax(2*n_R*n_vertices,n); 57880677318SStefano Zampini n = PetscMax((n_R+n_B)*n_vertices,n); 57906656605SStefano Zampini } 58006656605SStefano Zampini if (!pcbddc->issym) { 58106656605SStefano Zampini n = PetscMax(2*n_R*pcbddc->local_primal_size,n); 58206656605SStefano Zampini unsymmetric_check = PETSC_TRUE; 58306656605SStefano Zampini } 58406656605SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 58506656605SStefano Zampini 58606656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 58706656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 58806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 58906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 59006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 59106656605SStefano Zampini if (isLU || isILU || isCHOL) { 59206656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 593d62866d3SStefano Zampini } else if (sub_schurs->reuse_mumps) { 594d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 595d62866d3SStefano Zampini MatFactorType type; 596d62866d3SStefano Zampini 5976816873aSStefano Zampini F = reuse_mumps->F; 5986816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 599d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 60006656605SStefano Zampini } else { 60106656605SStefano Zampini F = NULL; 60206656605SStefano Zampini } 60306656605SStefano Zampini 60488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 60588ebb749SStefano Zampini if (n_constraints) { 60606656605SStefano Zampini Mat M1,M2,M3; 60780677318SStefano Zampini Mat auxmat; 60806656605SStefano Zampini IS is_aux; 60980677318SStefano Zampini PetscScalar *array,*array2; 61006656605SStefano Zampini 611f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 61280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 61388ebb749SStefano Zampini 61425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 61525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 6168ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 61780677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 61888ebb749SStefano Zampini 61980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 62080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 62106656605SStefano Zampini ierr = PetscMemzero(work,n_R*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 62288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 62306656605SStefano Zampini const PetscScalar *row_cmat_values; 62406656605SStefano Zampini const PetscInt *row_cmat_indices; 62506656605SStefano Zampini PetscInt size_of_constraint,j; 62688ebb749SStefano Zampini 62706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 62806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 62906656605SStefano Zampini work[row_cmat_indices[j]+i*n_R] = -row_cmat_values[j]; 63006656605SStefano Zampini } 63106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 63206656605SStefano Zampini } 63380677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 63406656605SStefano Zampini if (F) { 63506656605SStefano Zampini Mat B; 63606656605SStefano Zampini 63706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 63880677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 63906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 64006656605SStefano Zampini } else { 64180677318SStefano Zampini PetscScalar *marr; 64280677318SStefano Zampini 64380677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 64406656605SStefano Zampini for (i=0;i<n_constraints;i++) { 64506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 64680677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*n_R);CHKERRQ(ierr); 64706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 64806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 64906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 65006656605SStefano Zampini } 65180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 65206656605SStefano Zampini } 65380677318SStefano Zampini if (!pcbddc->switch_static) { 65480677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 65580677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 65680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 65780677318SStefano Zampini for (i=0;i<n_constraints;i++) { 65880677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*n_R);CHKERRQ(ierr); 65980677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 66080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 66280677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 66380677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 66480677318SStefano Zampini } 66580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 66680677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 66780677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 66880677318SStefano Zampini } else { 66980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 67080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 67125084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 67280677318SStefano Zampini } 67380677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 67480677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 67580677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 67606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 67706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 67880677318SStefano Zampini if (isCHOL) { 67980677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 68080677318SStefano Zampini } else { 68125084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 68280677318SStefano Zampini } 68380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 68406656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 68525084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 68625084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 68725084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 68880677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 68980677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 69080677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 69106656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 69206656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 69388ebb749SStefano Zampini } 69488ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 69588ebb749SStefano Zampini if (n_vertices) { 69606656605SStefano Zampini Mat newmat; 69706656605SStefano Zampini IS is_aux; 6983a50541eSStefano Zampini PetscInt ibs,mbs; 6993a50541eSStefano Zampini PetscBool issbaij; 7003a50541eSStefano Zampini 7016816873aSStefano Zampini if (sub_schurs->reuse_mumps) { /* is_R_local is not sorted, ISComplement doesn't like it */ 7026816873aSStefano Zampini IS tis; 7036816873aSStefano Zampini 7046816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 7056816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 7066816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 7076816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 7086816873aSStefano Zampini } else { 7093a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 7106816873aSStefano Zampini } 7119577ea80SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 7129577ea80SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 7133a50541eSStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ */ 7149577ea80SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7159577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7169577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7179577ea80SStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 71880677318SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7199577ea80SStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7209577ea80SStefano Zampini } else { 7213a50541eSStefano Zampini /* this is safe */ 7229577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 7233a50541eSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 7243a50541eSStefano Zampini if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 725b3d85658SStefano Zampini ierr = MatConvert(A_VV,MATSEQBAIJ,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 726b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7273a50541eSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 7283a50541eSStefano Zampini ierr = MatGetSubMatrix(newmat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 72906656605SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7303a50541eSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 7313a50541eSStefano Zampini } else { 732b3d85658SStefano Zampini ierr = MatConvert(A_VV,impMatType,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 7339577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 7349577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 7359577ea80SStefano Zampini } 73680677318SStefano Zampini /* need to convert to SeqAIJ for MatMatMult with dense matrix */ 73780677318SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 7383a50541eSStefano Zampini } 73925084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 74088ebb749SStefano Zampini } 74188ebb749SStefano Zampini 74288ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 743f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 74406656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 74506656605SStefano Zampini if (pcbddc->coarse_phi_D) { 74606656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 74706656605SStefano Zampini } 748f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 74906656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 75006656605SStefano Zampini PetscScalar *marray; 75106656605SStefano Zampini 75206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 75306656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 754f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 755f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 756f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 757f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 758f4ddd8eeSStefano Zampini } 759f4ddd8eeSStefano Zampini } 76006656605SStefano Zampini 761f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 76206656605SStefano Zampini PetscScalar *marray; 76388ebb749SStefano Zampini 76406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 7658eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 76606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 76788ebb749SStefano Zampini } 768b9d89cd5SStefano Zampini if (!pcbddc->issym) { 76906656605SStefano Zampini n *= 2; 770f4ddd8eeSStefano Zampini } 77106656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 77206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 77306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 77406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 77506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 77606656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 77788ebb749SStefano Zampini } 77806656605SStefano Zampini if (!pcbddc->issym) { 77906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 7801b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 78106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 78206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 78388ebb749SStefano Zampini } 78488ebb749SStefano Zampini } else { 785c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 786c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 7871b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 788c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 789c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 790c0553b1fSStefano Zampini } 79188ebb749SStefano Zampini } 79206656605SStefano Zampini } 79306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 79406656605SStefano Zampini /* vertices */ 79506656605SStefano Zampini if (n_vertices) { 79616f15bc4SStefano Zampini 79716f15bc4SStefano Zampini if (n_R) { 79806656605SStefano Zampini Mat A_RRmA_RV,S_VVt; /* S_VVt with LDA=N */ 79906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 80016f15bc4SStefano Zampini PetscScalar *x,*y; 80106656605SStefano Zampini 80206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 80321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 80406656605SStefano Zampini ierr = MatConvert(A_RV,impMatType,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 8056816873aSStefano Zampini if (F) { /* TODO could be optimized for symmetric problems */ 80606656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 80706656605SStefano Zampini } else { 80806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 80906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 81006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*n_R);CHKERRQ(ierr); 81106656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 81206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 81306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 81406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 81506656605SStefano Zampini } 81606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 81706656605SStefano Zampini } 81880677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 81906656605SStefano Zampini /* S_VV and S_CV are the subdomain contribution to coarse matrix. WARNING -> column major ordering */ 82006656605SStefano Zampini if (n_constraints) { 82106656605SStefano Zampini Mat B; 82280677318SStefano Zampini 823b3d85658SStefano Zampini ierr = PetscMemzero(work+n_R*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 82480677318SStefano Zampini for (i=0;i<n_vertices;i++) { 82580677318SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 82680677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+n_R*n_vertices+i*n_B);CHKERRQ(ierr); 82780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 82980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 83080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 83180677318SStefano Zampini } 83280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 83380677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 83480677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 83506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work+n_R*n_vertices,&B);CHKERRQ(ierr); 83680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 83706656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 83806656605SStefano Zampini ierr = PetscBLASIntCast(n_R*n_vertices,&B_N);CHKERRQ(ierr); 83906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+n_R*n_vertices,&B_one,work,&B_one)); 84006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 84106656605SStefano Zampini } 84206656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 84380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 84406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 84506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 84606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 84706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 84806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 84906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 85006656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85106656605SStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 85216f15bc4SStefano Zampini } else { 85316f15bc4SStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 85416f15bc4SStefano Zampini } 85521eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 85606656605SStefano Zampini /* coarse basis functions */ 85706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 85816f15bc4SStefano Zampini PetscScalar *y; 85916f15bc4SStefano Zampini 86006656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 86106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 86306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 86506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 86606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 86706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 86806656605SStefano Zampini 86906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 87006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 87206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 87406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 87506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 87606656605SStefano Zampini } 87706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 87806656605SStefano Zampini } 87906656605SStefano Zampini } 88006656605SStefano Zampini 88106656605SStefano Zampini if (n_constraints) { 88206656605SStefano Zampini Mat B; 88306656605SStefano Zampini 88406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 88506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 88706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 88806656605SStefano Zampini if (n_vertices) { 88980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 89080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 89180677318SStefano Zampini } else { 89280677318SStefano Zampini Mat S_VCt; 89380677318SStefano Zampini 89480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 89580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 89680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 89780677318SStefano Zampini } 89806656605SStefano Zampini } 89906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 90006656605SStefano Zampini /* coarse basis functions */ 90106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 90206656605SStefano Zampini PetscScalar *y; 90306656605SStefano Zampini 90406656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*i);CHKERRQ(ierr); 90506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 90606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 90706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 90906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 91006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 91106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 91206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 91406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 91606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 91706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 91806656605SStefano Zampini } 91906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 92006656605SStefano Zampini } 92106656605SStefano Zampini } 92280677318SStefano Zampini if (n_constraints) { 92380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 92480677318SStefano Zampini } 92506656605SStefano Zampini 92606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 92706656605SStefano Zampini if (!pcbddc->issym) { 92806656605SStefano Zampini 92906656605SStefano Zampini if (n_constraints) { 93016f15bc4SStefano Zampini Mat S_CCT,B_C; 93106656605SStefano Zampini 93280677318SStefano Zampini /* this is a lazy thing */ 93380677318SStefano Zampini ierr = MatConvert(C_CR,impMatType,MAT_REUSE_MATRIX,&C_CR);CHKERRQ(ierr); 93406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work+n_vertices*n_R,&B_C);CHKERRQ(ierr); 93506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 93606656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_CCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 93716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 93806656605SStefano Zampini if (n_vertices) { 93916f15bc4SStefano Zampini Mat B_V,S_VCT; 94006656605SStefano Zampini 94106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B_V);CHKERRQ(ierr); 94206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 94306656605SStefano Zampini ierr = MatTransposeMatMult(C_CR,S_VCT,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 94406656605SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 94516f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 94606656605SStefano Zampini } 94706656605SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 94880677318SStefano Zampini } else { /* if there are no constraints, reset work */ 94980677318SStefano Zampini ierr = PetscMemzero(work,n_R*pcbddc->local_primal_size*sizeof(PetscScalar));CHKERRQ(ierr); 95006656605SStefano Zampini } 95116f15bc4SStefano Zampini if (n_vertices && n_R) { 95206656605SStefano Zampini Mat A_VRT; 95380677318SStefano Zampini PetscScalar *marray; 95406656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 95506656605SStefano Zampini 95680677318SStefano Zampini ierr = MatTranspose(A_VR,MAT_INITIAL_MATRIX,&A_VRT);CHKERRQ(ierr); 95780677318SStefano Zampini ierr = MatConvert(A_VRT,impMatType,MAT_REUSE_MATRIX,&A_VRT);CHKERRQ(ierr); 95880677318SStefano Zampini ierr = MatDenseGetArray(A_VRT,&marray);CHKERRQ(ierr); 95906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_R,&B_N);CHKERRQ(ierr); 96080677318SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&m_one,marray,&B_one,work,&B_one)); 96180677318SStefano Zampini ierr = MatDenseRestoreArray(A_VRT,&marray);CHKERRQ(ierr); 96216f15bc4SStefano Zampini ierr = MatDestroy(&A_VRT);CHKERRQ(ierr); 96306656605SStefano Zampini } 96406656605SStefano Zampini 96506656605SStefano Zampini if (F) { /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 96606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 96706656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 96806656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 96906656605SStefano Zampini ierr = MatSolveTranspose(F,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 97206656605SStefano Zampini } 97306656605SStefano Zampini } else { 97406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 97506656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 97606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+(i+pcbddc->local_primal_size)*n_R);CHKERRQ(ierr); 97706656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 97806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 97906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 98006656605SStefano Zampini } 98106656605SStefano Zampini } 98206656605SStefano Zampini /* coarse basis functions */ 98306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 98406656605SStefano Zampini PetscScalar *y; 98506656605SStefano Zampini 98606656605SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+n_R*(i+pcbddc->local_primal_size));CHKERRQ(ierr); 98706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 98806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 98906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 99106656605SStefano Zampini if (i<n_vertices) { 99206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 99306656605SStefano Zampini } 99406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 99506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 99606656605SStefano Zampini 99706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 99806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 99906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 100006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 100206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 100306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 100406656605SStefano Zampini } 100506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 100606656605SStefano Zampini } 100706656605SStefano Zampini } 1008d62866d3SStefano Zampini /* free memory */ 100988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 101006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 101106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 101206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 101306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 1014d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1015d62866d3SStefano Zampini if (n_vertices) { 1016d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 1017d62866d3SStefano Zampini } 1018d62866d3SStefano Zampini if (n_constraints) { 1019d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 1020d62866d3SStefano Zampini } 102188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 102288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 102425084f0cSStefano Zampini if (pcbddc->dbg_flag) { 102588ebb749SStefano Zampini Mat coarse_sub_mat; 102625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 102788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 102888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 102988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 10308bec7fa6SStefano Zampini Mat C_B,CPHI; 10318bec7fa6SStefano Zampini IS is_dummy; 10328bec7fa6SStefano Zampini Vec mones; 103388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 103488ebb749SStefano Zampini PetscReal real_value; 103588ebb749SStefano Zampini 103688ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 103788ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 103888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 103988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 104088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 104188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 1042c0553b1fSStefano Zampini if (unsymmetric_check) { 104388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 104488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 104588ebb749SStefano Zampini } 104688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 104788ebb749SStefano Zampini 104825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 104906656605SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation\n");CHKERRQ(ierr); 105025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1051c0553b1fSStefano Zampini if (unsymmetric_check) { 105288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 105488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 105788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 105888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 105988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 106388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 106488ebb749SStefano Zampini } else { 106588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 106688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 106788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 106888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 106988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 107188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 107288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 107388ebb749SStefano Zampini } 107488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 107788ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 107881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 10798bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10800fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 108106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10828bec7fa6SStefano Zampini 10838bec7fa6SStefano Zampini /* check constraints */ 10848bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 10858bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 10868bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 10878bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 10888bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 10898bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 10908bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 10918bec7fa6SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 10928bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 10938bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 10948bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 10958bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 10968bec7fa6SStefano Zampini 109725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 109888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 109988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 110088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 110188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 110288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 110388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 110488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 110588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 110688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 110788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 1108c0553b1fSStefano Zampini if (unsymmetric_check) { 110988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 111088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 111188ebb749SStefano Zampini } 111288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 111388ebb749SStefano Zampini } 11148629588bSStefano Zampini /* get back data */ 11158629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 111688ebb749SStefano Zampini PetscFunctionReturn(0); 111788ebb749SStefano Zampini } 111888ebb749SStefano Zampini 111988ebb749SStefano Zampini #undef __FUNCT__ 1120d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 1121d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 1122d65f70fdSStefano Zampini { 1123d65f70fdSStefano Zampini Mat *work_mat; 1124d65f70fdSStefano Zampini IS isrow_s,iscol_s; 1125d65f70fdSStefano Zampini PetscBool rsorted,csorted; 1126d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 1127d65f70fdSStefano Zampini PetscErrorCode ierr; 1128d65f70fdSStefano Zampini 1129d65f70fdSStefano Zampini PetscFunctionBegin; 1130d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 1131d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 1132d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 1133d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1134d65f70fdSStefano Zampini 1135d65f70fdSStefano Zampini if (!rsorted) { 1136d65f70fdSStefano Zampini const PetscInt *idxs; 1137d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1138d65f70fdSStefano Zampini 1139d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 1140d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 1141d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1142d65f70fdSStefano Zampini idxs_perm_r[i] = i; 1143d65f70fdSStefano Zampini } 1144d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 1145d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 1146d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1147d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 1148d65f70fdSStefano Zampini } 1149d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 1150d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 1151d65f70fdSStefano Zampini } else { 1152d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1153d65f70fdSStefano Zampini isrow_s = isrow; 1154d65f70fdSStefano Zampini } 1155d65f70fdSStefano Zampini 1156d65f70fdSStefano Zampini if (!csorted) { 1157d65f70fdSStefano Zampini if (isrow == iscol) { 1158d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 1159d65f70fdSStefano Zampini iscol_s = isrow_s; 1160d65f70fdSStefano Zampini } else { 1161d65f70fdSStefano Zampini const PetscInt *idxs; 1162d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 1163d65f70fdSStefano Zampini 1164d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 1165d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 1166d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1167d65f70fdSStefano Zampini idxs_perm_c[i] = i; 1168d65f70fdSStefano Zampini } 1169d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 1170d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 1171d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1172d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 1173d65f70fdSStefano Zampini } 1174d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 1175d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 1176d65f70fdSStefano Zampini } 1177d65f70fdSStefano Zampini } else { 1178d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1179d65f70fdSStefano Zampini iscol_s = iscol; 1180d65f70fdSStefano Zampini } 1181d65f70fdSStefano Zampini 1182d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1183d65f70fdSStefano Zampini 1184d65f70fdSStefano Zampini if (!rsorted || !csorted) { 1185d65f70fdSStefano Zampini Mat new_mat; 1186d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 1187d65f70fdSStefano Zampini 1188d65f70fdSStefano Zampini if (!rsorted) { 1189d65f70fdSStefano Zampini PetscInt *idxs_r,i; 1190d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 1191d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 1192d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 1193d65f70fdSStefano Zampini } 1194d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 1195d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 1196d65f70fdSStefano Zampini } else { 1197d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 1198d65f70fdSStefano Zampini } 1199d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 1200d65f70fdSStefano Zampini 1201d65f70fdSStefano Zampini if (!csorted) { 1202d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 1203d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 1204d65f70fdSStefano Zampini is_perm_c = is_perm_r; 1205d65f70fdSStefano Zampini } else { 1206d65f70fdSStefano Zampini PetscInt *idxs_c,i; 1207d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 1208d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 1209d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 1210d65f70fdSStefano Zampini } 1211d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 1212d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 1213d65f70fdSStefano Zampini } 1214d65f70fdSStefano Zampini } else { 1215d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 1216d65f70fdSStefano Zampini } 1217d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 1218d65f70fdSStefano Zampini 1219d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 1220d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 1221d65f70fdSStefano Zampini work_mat[0] = new_mat; 1222d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 1223d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 1224d65f70fdSStefano Zampini } 1225d65f70fdSStefano Zampini 1226d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 1227d65f70fdSStefano Zampini *B = work_mat[0]; 1228d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 1229d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 1230d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 1231d65f70fdSStefano Zampini PetscFunctionReturn(0); 1232d65f70fdSStefano Zampini } 1233d65f70fdSStefano Zampini 1234d65f70fdSStefano Zampini #undef __FUNCT__ 12355e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 12365e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 1237aa0d41d4SStefano Zampini { 1238aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 12395e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1240d65f70fdSStefano Zampini Mat new_mat; 12415e8657edSStefano Zampini IS is_local,is_global; 1242d65f70fdSStefano Zampini PetscInt local_size; 1243d65f70fdSStefano Zampini PetscBool isseqaij; 1244aa0d41d4SStefano Zampini PetscErrorCode ierr; 1245aa0d41d4SStefano Zampini 1246aa0d41d4SStefano Zampini PetscFunctionBegin; 1247aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 12485e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 12495e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 1250906d46d4SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(matis->mapping,is_local,&is_global);CHKERRQ(ierr); 1251906d46d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 1252d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 1253906d46d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 1254906d46d4SStefano Zampini 1255906d46d4SStefano Zampini /* check */ 1256906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 1257906d46d4SStefano Zampini Vec x,x_change; 1258906d46d4SStefano Zampini PetscReal error; 1259906d46d4SStefano Zampini 12605e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 1261906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 12625e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 12635e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 12645e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1265d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 12665e8657edSStefano Zampini ierr = VecScatterBegin(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 12675e8657edSStefano Zampini ierr = VecScatterEnd(matis->ctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1268906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 1269906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 1270906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1271906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 1272906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 1273906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 1274906d46d4SStefano Zampini } 1275906d46d4SStefano Zampini 127622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 12779b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 127822d5777bSStefano Zampini if (isseqaij) { 1279d65f70fdSStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1280aa0d41d4SStefano Zampini } else { 1281aa0d41d4SStefano Zampini Mat work_mat; 1282aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 1283d65f70fdSStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 1284aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1285aa0d41d4SStefano Zampini } 1286d65f70fdSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 1287e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 1288e496cd5dSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,pcbddc->issym);CHKERRQ(ierr); 1289e496cd5dSStefano Zampini #endif 129045a1bb75SStefano Zampini /* 129145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 1292d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 129345a1bb75SStefano Zampini */ 1294d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 1295aa0d41d4SStefano Zampini PetscFunctionReturn(0); 1296aa0d41d4SStefano Zampini } 1297aa0d41d4SStefano Zampini 1298aa0d41d4SStefano Zampini #undef __FUNCT__ 1299a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 13008ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 1301a64d13efSStefano Zampini { 1302a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1303a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1304d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1305*53892102SStefano Zampini PetscInt *idx_R_local=NULL; 13063a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 13073a50541eSStefano Zampini PetscInt vbs,bs; 13086816873aSStefano Zampini PetscBT bitmask=NULL; 1309a64d13efSStefano Zampini PetscErrorCode ierr; 1310a64d13efSStefano Zampini 1311a64d13efSStefano Zampini PetscFunctionBegin; 1312b23d619eSStefano Zampini /* 1313b23d619eSStefano Zampini No need to setup local scatters if 1314b23d619eSStefano Zampini - primal space is unchanged 1315b23d619eSStefano Zampini AND 1316b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 1317b23d619eSStefano Zampini AND 1318b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 1319b23d619eSStefano Zampini */ 1320b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 1321f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 1322f4ddd8eeSStefano Zampini } 1323f4ddd8eeSStefano Zampini /* destroy old objects */ 1324f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1325f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1326f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1327a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 1328b371cd4fSStefano Zampini n_B = pcis->n_B; 1329b371cd4fSStefano Zampini n_D = pcis->n - n_B; 1330b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 13316816873aSStefano Zampini 13326816873aSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 13336816873aSStefano Zampini 1334*53892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 13356816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 1336*53892102SStefano Zampini ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 13374641a718SStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 13384641a718SStefano Zampini for (i=0;i<n_vertices;i++) { 13390e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 13404641a718SStefano Zampini } 13413a50541eSStefano Zampini 1342a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 13434641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 13446816873aSStefano Zampini idx_R_local[n_R++] = i; 1345a64d13efSStefano Zampini } 1346a64d13efSStefano Zampini } 1347*53892102SStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing MUMPS Schur solver */ 13486816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 13496816873aSStefano Zampini 1350*53892102SStefano Zampini ierr = ISGetIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1351*53892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_R,&n_R);CHKERRQ(ierr); 13526816873aSStefano Zampini } 13533a50541eSStefano Zampini 13543a50541eSStefano Zampini /* Block code */ 13553a50541eSStefano Zampini vbs = 1; 13563a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 13573a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 13583a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 13593a50541eSStefano Zampini PetscInt *vary; 13603a50541eSStefano Zampini /* Verify if the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 1361785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 13623a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 13630e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 1364964fefecSStefano Zampini for (i=0; i<n_vertices/bs; i++) { 13653a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 13663a50541eSStefano Zampini is_blocked = PETSC_FALSE; 13673a50541eSStefano Zampini break; 13683a50541eSStefano Zampini } 13693a50541eSStefano Zampini } 13703a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 13713a50541eSStefano Zampini vbs = bs; 13723a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 13733a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 13743a50541eSStefano Zampini } 13753a50541eSStefano Zampini } 13763a50541eSStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 13773a50541eSStefano Zampini } 13783a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 1379*53892102SStefano Zampini if (sub_schurs->reuse_mumps) { 1380*53892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1381*53892102SStefano Zampini 1382*53892102SStefano Zampini ierr = ISRestoreIndices(reuse_mumps->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1383*53892102SStefano Zampini ierr = ISDestroy(&reuse_mumps->is_R);CHKERRQ(ierr); 1384*53892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 1385*53892102SStefano Zampini reuse_mumps->is_R = pcbddc->is_R_local; 1386*53892102SStefano Zampini } else { 13873a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 1388*53892102SStefano Zampini } 1389a64d13efSStefano Zampini 1390a64d13efSStefano Zampini /* print some info if requested */ 1391a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 1392a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1393a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 13940fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1395a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 1396a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 13973a50541eSStefano 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); 1398a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1399a64d13efSStefano Zampini } 1400a64d13efSStefano Zampini 1401a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 14026816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 14036816873aSStefano Zampini IS is_aux1,is_aux2; 14046816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 14056816873aSStefano Zampini 14063a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1407854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 1408854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 1409a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 14104641a718SStefano Zampini for (i=0; i<n_D; i++) { 14114641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 14124641a718SStefano Zampini } 1413a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1414a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14154641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 14164641a718SStefano Zampini aux_array1[j++] = i; 1417a64d13efSStefano Zampini } 1418a64d13efSStefano Zampini } 1419a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1420a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1421a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 14224641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 14234641a718SStefano Zampini aux_array2[j++] = i; 1424a64d13efSStefano Zampini } 1425a64d13efSStefano Zampini } 1426a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 1427a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 1428a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 1429a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1430a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 1431a64d13efSStefano Zampini 14328eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1433785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 1434a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 14354641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 14364641a718SStefano Zampini aux_array1[j++] = i; 1437a64d13efSStefano Zampini } 1438a64d13efSStefano Zampini } 1439a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 1440a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 1441a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 1442a64d13efSStefano Zampini } 14434641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 14443a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 1445d62866d3SStefano Zampini } else { 1446*53892102SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 14476816873aSStefano Zampini IS tis; 14486816873aSStefano Zampini PetscInt schur_size; 14496816873aSStefano Zampini 1450*53892102SStefano Zampini ierr = ISGetLocalSize(reuse_mumps->is_B,&schur_size);CHKERRQ(ierr); 14516816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 1452*53892102SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_mumps->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 14536816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 14546816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 14556816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 14566816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 14576816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 1458d62866d3SStefano Zampini } 1459d62866d3SStefano Zampini } 1460a64d13efSStefano Zampini PetscFunctionReturn(0); 1461a64d13efSStefano Zampini } 1462a64d13efSStefano Zampini 1463304d26faSStefano Zampini 1464304d26faSStefano Zampini #undef __FUNCT__ 1465304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 1466684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 1467304d26faSStefano Zampini { 1468304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1469304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 1470304d26faSStefano Zampini PC pc_temp; 1471304d26faSStefano Zampini Mat A_RR; 1472f4ddd8eeSStefano Zampini MatReuse reuse; 1473304d26faSStefano Zampini PetscScalar m_one = -1.0; 1474304d26faSStefano Zampini PetscReal value; 1475af732b37SStefano Zampini PetscInt n_D,n_R,ibs,mbs; 14769577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 1477304d26faSStefano Zampini PetscErrorCode ierr; 1478e604994aSStefano Zampini /* prefixes stuff */ 1479312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 1480e604994aSStefano Zampini size_t len; 1481304d26faSStefano Zampini 1482304d26faSStefano Zampini PetscFunctionBegin; 1483304d26faSStefano Zampini 1484e604994aSStefano Zampini /* compute prefixes */ 1485e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 1486e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 1487e604994aSStefano Zampini if (!pcbddc->current_level) { 1488e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1489e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 1490e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1491e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1492e604994aSStefano Zampini } else { 1493e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 1494312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 1495e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 1496e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 1497312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 1498312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 149934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 150034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 1501e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 1502e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 1503e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 1504e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 1505e604994aSStefano Zampini } 1506e604994aSStefano Zampini 1507304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 1508684f6988SStefano Zampini if (dirichlet) { 1509d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1510964fefecSStefano Zampini if (pcbddc->issym) { 1511964fefecSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 1512964fefecSStefano Zampini } 1513ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 1514964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 1515304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 1516304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 1517304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 1518304d26faSStefano Zampini /* default */ 1519304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 1520e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 15219577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 1522304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 15239577ea80SStefano Zampini if (issbaij) { 15249577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 15259577ea80SStefano Zampini } else { 1526304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 15279577ea80SStefano Zampini } 1528304d26faSStefano Zampini /* Allow user's customization */ 1529304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 1530304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1531304d26faSStefano Zampini } 1532d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 1533d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1534d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1535d62866d3SStefano Zampini 1536d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_mumps->interior_solver);CHKERRQ(ierr); 1537d5574798SStefano Zampini } 1538304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1539304d26faSStefano Zampini if (!n_D) { 1540304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 1541304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1542304d26faSStefano Zampini } 1543304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 1544304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 1545304d26faSStefano Zampini /* set ksp_D into pcis data */ 1546304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 1547304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 1548304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 1549684f6988SStefano Zampini } 1550304d26faSStefano Zampini 1551304d26faSStefano Zampini /* NEUMANN PROBLEM */ 1552684f6988SStefano Zampini A_RR = 0; 1553684f6988SStefano Zampini if (neumann) { 1554d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1555f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 15568ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 1557f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 1558f4ddd8eeSStefano Zampini PetscInt nn_R; 155981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 1560f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 1561f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 1562f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 1563f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 1564f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1565f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1566f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 1567727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 1568f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1569f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1570f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 1571f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 1572f4ddd8eeSStefano Zampini } 1573f4ddd8eeSStefano Zampini } 1574f4ddd8eeSStefano Zampini /* last check */ 1575d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 1576f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1577f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1578f4ddd8eeSStefano Zampini } 1579f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 1580f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 1581f4ddd8eeSStefano Zampini } 15826816873aSStefano Zampini /* extract A_RR */ 15836816873aSStefano Zampini if (!sub_schurs->reuse_mumps) { 15846816873aSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 15856816873aSStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 15866816873aSStefano Zampini if (ibs != mbs) { 15876816873aSStefano Zampini Mat newmat; 15886816873aSStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr); 15896816873aSStefano Zampini ierr = MatGetSubMatrix(newmat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 15906816873aSStefano Zampini ierr = MatDestroy(&newmat);CHKERRQ(ierr); 15916816873aSStefano Zampini } else { 15926816873aSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 15936816873aSStefano Zampini } 15946816873aSStefano Zampini if (pcbddc->issym) { 15956816873aSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 15966816873aSStefano Zampini } 15976816873aSStefano Zampini } else { 15986816873aSStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 15996816873aSStefano Zampini 16006816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 16016816873aSStefano Zampini ierr = PCGetOperators(reuse_mumps->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 16026816873aSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 16036816873aSStefano Zampini } 1604f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 1605304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 1606304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 1607304d26faSStefano Zampini /* default */ 1608304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 1609e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 1610304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 16119577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 16129577ea80SStefano Zampini if (issbaij) { 16139577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 16149577ea80SStefano Zampini } else { 1615304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 16169577ea80SStefano Zampini } 1617304d26faSStefano Zampini /* Allow user's customization */ 1618304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 1619304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 1620304d26faSStefano Zampini } 1621d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 1622304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 1623304d26faSStefano Zampini if (!n_R) { 1624304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 1625304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 1626304d26faSStefano Zampini } 1627d62866d3SStefano Zampini /* Reuse MUMPS solver if it is present */ 1628d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1629d62866d3SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1630d62866d3SStefano Zampini 1631d62866d3SStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_mumps->correction_solver);CHKERRQ(ierr); 1632d62866d3SStefano Zampini } 1633304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 1634304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 1635684f6988SStefano Zampini } 16366816873aSStefano Zampini /* free Neumann problem's matrix */ 16376816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 1638304d26faSStefano Zampini 1639304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 16400fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 1641684f6988SStefano Zampini if (pcbddc->dbg_flag) { 1642684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1643684f6988SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1644684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 1645684f6988SStefano Zampini } 1646684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 16470fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 16480fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 16490fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 16500fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 16510fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 1652304d26faSStefano Zampini /* need to be adapted? */ 1653b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1654b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1655b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 1656304d26faSStefano Zampini /* print info */ 1657304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1658e604994aSStefano 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); 1659304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1660304d26faSStefano Zampini } 1661b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 16628ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcis->is_I_local);CHKERRQ(ierr); 1663304d26faSStefano Zampini } 1664684f6988SStefano Zampini } 1665684f6988SStefano Zampini if (neumann) { /* Neumann */ 16666816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 16670fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 16680fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 16690fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 16700fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 16710fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 1672304d26faSStefano Zampini /* need to be adapted? */ 1673b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 1674b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 1675304d26faSStefano Zampini /* print info */ 1676304d26faSStefano Zampini if (pcbddc->dbg_flag) { 1677e604994aSStefano 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); 1678304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1679304d26faSStefano Zampini } 1680b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 16818ce42a96SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,pcbddc->is_R_local);CHKERRQ(ierr); 1682304d26faSStefano Zampini } 16830fccc4e9SStefano Zampini } 1684684f6988SStefano Zampini } 1685304d26faSStefano Zampini PetscFunctionReturn(0); 1686304d26faSStefano Zampini } 1687304d26faSStefano Zampini 1688304d26faSStefano Zampini #undef __FUNCT__ 1689ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 169080677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 1691674ae819SStefano Zampini { 1692674ae819SStefano Zampini PetscErrorCode ierr; 1693674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1694be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 1695674ae819SStefano Zampini 1696674ae819SStefano Zampini PetscFunctionBegin; 1697be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 169880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 1699be83ff47SStefano Zampini } 170080677318SStefano Zampini if (!pcbddc->switch_static) { 170180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 170280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 170380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 170420c7b377SStefano Zampini } 1705be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 170680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 170780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 170820c7b377SStefano Zampini } else { 1709be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1710be83ff47SStefano Zampini 1711*53892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1712*53892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,inout_B,reuse_mumps->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1713be83ff47SStefano Zampini } 1714be83ff47SStefano Zampini } else { 171580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 171980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 172080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 172180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 172280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 172420c7b377SStefano Zampini } 1725674ae819SStefano Zampini } 1726be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 172780677318SStefano Zampini if (applytranspose) { 172880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 172980677318SStefano Zampini } else { 173080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 173180677318SStefano Zampini } 1732*53892102SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1733be83ff47SStefano Zampini } else { 1734be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1735be83ff47SStefano Zampini 1736be83ff47SStefano Zampini if (applytranspose) { 1737*53892102SStefano Zampini ierr = MatMumpsSolveSchurComplementTranspose(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1738be83ff47SStefano Zampini } else { 1739*53892102SStefano Zampini ierr = MatMumpsSolveSchurComplement(reuse_mumps->F,reuse_mumps->rhs_B,reuse_mumps->sol_B);CHKERRQ(ierr); 1740be83ff47SStefano Zampini } 1741*53892102SStefano Zampini #endif 1742be83ff47SStefano Zampini } 174380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 174480677318SStefano Zampini if (!pcbddc->switch_static) { 1745be83ff47SStefano Zampini if (!sub_schurs->reuse_mumps) { 174680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 174780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1748be83ff47SStefano Zampini } else { 1749be83ff47SStefano Zampini PCBDDCReuseMumps reuse_mumps = sub_schurs->reuse_mumps; 1750be83ff47SStefano Zampini 1751*53892102SStefano Zampini ierr = VecScatterBegin(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1752*53892102SStefano Zampini ierr = VecScatterEnd(reuse_mumps->correction_scatter_B,reuse_mumps->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1753be83ff47SStefano Zampini } 175480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 175580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 175680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 175780677318SStefano Zampini } 175880677318SStefano Zampini } else { 175980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 176480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 176580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 176680677318SStefano Zampini } 176780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 177180677318SStefano Zampini } 1772674ae819SStefano Zampini PetscFunctionReturn(0); 1773674ae819SStefano Zampini } 1774674ae819SStefano Zampini 1775dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 1776674ae819SStefano Zampini #undef __FUNCT__ 1777674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 1778dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 1779674ae819SStefano Zampini { 1780674ae819SStefano Zampini PetscErrorCode ierr; 1781674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 1782674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 1783674ae819SStefano Zampini const PetscScalar zero = 0.0; 1784674ae819SStefano Zampini 1785674ae819SStefano Zampini PetscFunctionBegin; 1786dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 1787dc359a40SStefano Zampini if (applytranspose) { 1788674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 17898eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 1790dc359a40SStefano Zampini } else { 1791674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 1792674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 179315aaf578SStefano Zampini } 179412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 179512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 179612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179812edc857SStefano Zampini 17999f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 180012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 180112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 1802964fefecSStefano Zampini Vec rhs,sol; 1803964fefecSStefano Zampini 1804964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 1805964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 180612edc857SStefano Zampini if (applytranspose) { 1807964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 180812edc857SStefano Zampini } else { 1809964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 181012edc857SStefano Zampini } 181112edc857SStefano Zampini } 1812674ae819SStefano Zampini 1813674ae819SStefano Zampini /* Local solution on R nodes */ 181480677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 181580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 18169f00e9b4SStefano Zampini } 1817674ae819SStefano Zampini 18189f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 18199f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 182012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1821674ae819SStefano Zampini 1822674ae819SStefano Zampini /* Sum contributions from two levels */ 1823dc359a40SStefano Zampini if (applytranspose) { 1824dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 1825dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1826dc359a40SStefano Zampini } else { 1827674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 18288eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 1829dc359a40SStefano Zampini } 1830674ae819SStefano Zampini PetscFunctionReturn(0); 1831674ae819SStefano Zampini } 1832674ae819SStefano Zampini 1833674ae819SStefano Zampini #undef __FUNCT__ 1834674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 183512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 1836674ae819SStefano Zampini { 1837674ae819SStefano Zampini PetscErrorCode ierr; 1838674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 183958da7f69SStefano Zampini PetscScalar *array; 184012edc857SStefano Zampini Vec from,to; 1841674ae819SStefano Zampini 1842674ae819SStefano Zampini PetscFunctionBegin; 184312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 184412edc857SStefano Zampini from = pcbddc->coarse_vec; 184512edc857SStefano Zampini to = pcbddc->vec1_P; 184612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 184712edc857SStefano Zampini Vec tvec; 184858da7f69SStefano Zampini 184958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 185058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 185112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 185258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 185358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 185458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 185512edc857SStefano Zampini } 185612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 185712edc857SStefano Zampini from = pcbddc->vec1_P; 185812edc857SStefano Zampini to = pcbddc->coarse_vec; 185912edc857SStefano Zampini } 186012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 1861674ae819SStefano Zampini PetscFunctionReturn(0); 1862674ae819SStefano Zampini } 1863674ae819SStefano Zampini 1864674ae819SStefano Zampini #undef __FUNCT__ 1865674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 186612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 1867674ae819SStefano Zampini { 1868674ae819SStefano Zampini PetscErrorCode ierr; 1869674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 187058da7f69SStefano Zampini PetscScalar *array; 187112edc857SStefano Zampini Vec from,to; 1872674ae819SStefano Zampini 1873674ae819SStefano Zampini PetscFunctionBegin; 187412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 187512edc857SStefano Zampini from = pcbddc->coarse_vec; 187612edc857SStefano Zampini to = pcbddc->vec1_P; 187712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 187812edc857SStefano Zampini from = pcbddc->vec1_P; 187912edc857SStefano Zampini to = pcbddc->coarse_vec; 188012edc857SStefano Zampini } 188112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 188212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 188312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 188412edc857SStefano Zampini Vec tvec; 188558da7f69SStefano Zampini 188612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 188758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 188858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 188958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 189058da7f69SStefano Zampini } 189158da7f69SStefano Zampini } else { 189258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 189358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 189412edc857SStefano Zampini } 189512edc857SStefano Zampini } 1896674ae819SStefano Zampini PetscFunctionReturn(0); 1897674ae819SStefano Zampini } 1898674ae819SStefano Zampini 1899984c4197SStefano Zampini /* uncomment for testing purposes */ 1900984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 1901674ae819SStefano Zampini #undef __FUNCT__ 1902674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 1903674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 1904674ae819SStefano Zampini { 1905674ae819SStefano Zampini PetscErrorCode ierr; 1906674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 1907674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1908674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 1909984c4197SStefano Zampini /* one and zero */ 1910984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 1911984c4197SStefano Zampini /* space to store constraints and their local indices */ 19129162d606SStefano Zampini PetscScalar *constraints_data; 19139162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 19149162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 19159162d606SStefano Zampini PetscInt *constraints_n; 1916984c4197SStefano Zampini /* iterators */ 1917b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 1918984c4197SStefano Zampini /* BLAS integers */ 1919e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 1920e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 1921c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 1922727cdba6SStefano Zampini /* reuse */ 19230e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 19240e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 1925984c4197SStefano Zampini /* change of basis */ 1926b3d85658SStefano Zampini PetscBool qr_needed; 19279162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 1928984c4197SStefano Zampini /* auxiliary stuff */ 192964efe560SStefano Zampini PetscInt *nnz,*is_indices; 19308a0068c3SStefano Zampini PetscInt ncc; 1931984c4197SStefano Zampini /* some quantities */ 193245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 1933a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 1934984c4197SStefano Zampini 1935674ae819SStefano Zampini PetscFunctionBegin; 19368e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 19378e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 19388e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1939088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 1940088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 19410e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 19420e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 19430e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19440e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 19450e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1946088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 1947cf5a6209SStefano Zampini 1948cf5a6209SStefano Zampini /* print some info */ 1949cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 1950cf5a6209SStefano Zampini IS vertices; 1951cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 1952cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 1953cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 1954cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 1955cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 1956cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 1957cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 1958fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 1959fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 1960cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1961cf5a6209SStefano Zampini } 1962cf5a6209SStefano Zampini 1963cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 19649162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 1965cf5a6209SStefano Zampini MatNullSpace nearnullsp; 1966cf5a6209SStefano Zampini const Vec *nearnullvecs; 1967cf5a6209SStefano Zampini Vec *localnearnullsp; 1968cf5a6209SStefano Zampini PetscScalar *array; 1969cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 1970cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 1971cf5a6209SStefano Zampini /* LAPACK working arrays for SVD or POD */ 1972b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 1973cf5a6209SStefano Zampini PetscScalar *work; 1974cf5a6209SStefano Zampini PetscReal *singular_vals; 1975cf5a6209SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1976cf5a6209SStefano Zampini PetscReal *rwork; 1977cf5a6209SStefano Zampini #endif 1978cf5a6209SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 1979cf5a6209SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 1980cf5a6209SStefano Zampini #else 1981964fefecSStefano Zampini PetscBLASInt dummy_int=1; 1982964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 1983cf5a6209SStefano Zampini #endif 1984cf5a6209SStefano Zampini 1985674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 1986d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 1987d06fc5fdSStefano Zampini /* free unneeded index sets */ 1988d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 1989d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 1990674ae819SStefano Zampini } 1991d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 1992d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 1993d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 1994d06fc5fdSStefano Zampini } 1995d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 1996d06fc5fdSStefano Zampini n_ISForEdges = 0; 1997d06fc5fdSStefano Zampini } 1998d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 1999d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2000d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2001d06fc5fdSStefano Zampini } 2002d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2003d06fc5fdSStefano Zampini n_ISForFaces = 0; 2004d06fc5fdSStefano Zampini } 200570022509SStefano Zampini 200670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 200770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 200870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 200970022509SStefano Zampini if (pcbddc->NullSpace) { 201070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 201170022509SStefano Zampini 201270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 2013b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 2014d06fc5fdSStefano Zampini if (!ISForEdges) { 2015d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 2016d06fc5fdSStefano Zampini } 2017b8ffe317SStefano Zampini } 2018d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 2019d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 2020d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2021d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 2022d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 202398a51de6SStefano Zampini } 202470022509SStefano Zampini #endif 202508122e43SStefano Zampini 2026674ae819SStefano Zampini /* check if near null space is attached to global mat */ 2027674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 2028674ae819SStefano Zampini if (nearnullsp) { 2029674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 2030f4ddd8eeSStefano Zampini /* remove any stored info */ 2031f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 2032f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2033f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 2034f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 2035f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 2036473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 2037f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 2038f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 2039f4ddd8eeSStefano Zampini } 2040984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 2041984c4197SStefano Zampini nnsp_size = 0; 2042674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 2043674ae819SStefano Zampini } 2044984c4197SStefano Zampini /* get max number of constraints on a single cc */ 2045984c4197SStefano Zampini max_constraints = nnsp_size; 2046984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 2047984c4197SStefano Zampini 2048674ae819SStefano Zampini /* 2049674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 20509162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 20519162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 20529162d606SStefano Zampini There can be multiple constraints per connected component 2053674ae819SStefano Zampini */ 2054674ae819SStefano Zampini n_vertices = 0; 2055674ae819SStefano Zampini if (ISForVertices) { 2056674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 2057674ae819SStefano Zampini } 20589162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 20599162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 20609162d606SStefano Zampini 20619162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 20629162d606SStefano Zampini total_counts *= max_constraints; 2063674ae819SStefano Zampini total_counts += n_vertices; 20644641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 20659162d606SStefano Zampini 2066674ae819SStefano Zampini total_counts = 0; 2067674ae819SStefano Zampini max_size_of_constraint = 0; 2068674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 20699162d606SStefano Zampini IS used_is; 2070674ae819SStefano Zampini if (i<n_ISForEdges) { 20719162d606SStefano Zampini used_is = ISForEdges[i]; 2072674ae819SStefano Zampini } else { 20739162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 2074674ae819SStefano Zampini } 20759162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 2076674ae819SStefano Zampini total_counts += j; 2077674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 2078674ae819SStefano Zampini } 20799162d606SStefano 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); 20809162d606SStefano Zampini 2081984c4197SStefano Zampini /* get local part of global near null space vectors */ 2082785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 2083984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2084984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 2085984c4197SStefano Zampini ierr = VecScatterBegin(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2086984c4197SStefano Zampini ierr = VecScatterEnd(matis->ctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2087984c4197SStefano Zampini } 2088674ae819SStefano Zampini 2089242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 2090242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 2091a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 2092242a89d7SStefano Zampini 2093984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 2094a773dcb8SStefano Zampini if (!skip_lapack) { 2095674ae819SStefano Zampini PetscScalar temp_work; 2096911cabfeSStefano Zampini 2097674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2098984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 2099785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 2100785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 2101785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 2102674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2103785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 2104674ae819SStefano Zampini #endif 2105674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2106c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 2107c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 2108674ae819SStefano Zampini lwork = -1; 2109674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2110674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2111c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 2112674ae819SStefano Zampini #else 2113c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 2114674ae819SStefano Zampini #endif 2115674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2116984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 2117674ae819SStefano Zampini #else /* on missing GESVD */ 2118674ae819SStefano Zampini /* SVD */ 2119674ae819SStefano Zampini PetscInt max_n,min_n; 2120674ae819SStefano Zampini max_n = max_size_of_constraint; 2121984c4197SStefano Zampini min_n = max_constraints; 2122984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 2123674ae819SStefano Zampini min_n = max_size_of_constraint; 2124984c4197SStefano Zampini max_n = max_constraints; 2125674ae819SStefano Zampini } 2126785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 2127674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2128785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 2129674ae819SStefano Zampini #endif 2130674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 2131674ae819SStefano Zampini lwork = -1; 2132e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 2133e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 2134b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 2135674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2136674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 21379162d606SStefano 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)); 2138674ae819SStefano Zampini #else 21399162d606SStefano 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)); 2140674ae819SStefano Zampini #endif 2141674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2142984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 2143984c4197SStefano Zampini #endif /* on missing GESVD */ 2144674ae819SStefano Zampini /* Allocate optimal workspace */ 2145674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 2146854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 2147674ae819SStefano Zampini } 2148674ae819SStefano Zampini /* Now we can loop on constraining sets */ 2149674ae819SStefano Zampini total_counts = 0; 21509162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 21519162d606SStefano Zampini constraints_data_ptr[0] = 0; 2152674ae819SStefano Zampini /* vertices */ 21539162d606SStefano Zampini if (n_vertices) { 2154674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 21559162d606SStefano Zampini if (nnsp_has_cnst) { /* it considers all possible vertices */ 21569162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 2157674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 21589162d606SStefano Zampini constraints_n[total_counts] = 1; 21599162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21609162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21619162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2162674ae819SStefano Zampini total_counts++; 2163674ae819SStefano Zampini } 2164674ae819SStefano Zampini } else { /* consider vertices for which exist at least a localnearnullsp which is not null there */ 2165984c4197SStefano Zampini PetscBool used_vertex; 2166674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 2167674ae819SStefano Zampini used_vertex = PETSC_FALSE; 2168674ae819SStefano Zampini k = 0; 2169674ae819SStefano Zampini while (!used_vertex && k<nnsp_size) { 2170984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2171984c4197SStefano Zampini if (PetscAbsScalar(array[is_indices[i]])>0.0) { 21729162d606SStefano Zampini constraints_n[total_counts] = 1; 21739162d606SStefano Zampini constraints_idxs[total_counts] = is_indices[i]; 21749162d606SStefano Zampini constraints_data[total_counts] = 1.0; 21759162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 21769162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 2177674ae819SStefano Zampini total_counts++; 2178674ae819SStefano Zampini used_vertex = PETSC_TRUE; 2179674ae819SStefano Zampini } 2180984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2181674ae819SStefano Zampini k++; 2182674ae819SStefano Zampini } 2183674ae819SStefano Zampini } 2184674ae819SStefano Zampini } 2185674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2186674ae819SStefano Zampini n_vertices = total_counts; 2187674ae819SStefano Zampini } 2188984c4197SStefano Zampini 2189674ae819SStefano Zampini /* edges and faces */ 21909162d606SStefano Zampini total_counts_cc = total_counts; 2191911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 21929162d606SStefano Zampini IS used_is; 21939162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 21949162d606SStefano Zampini 2195911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 21969162d606SStefano Zampini used_is = ISForEdges[ncc]; 2197984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 2198674ae819SStefano Zampini } else { 21999162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 2200984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 2201674ae819SStefano Zampini } 2202674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 22039162d606SStefano Zampini 22049162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 22059162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2206984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 2207984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 2208674ae819SStefano Zampini if (nnsp_has_cnst) { 22095b08dc53SStefano Zampini PetscScalar quad_value; 22109162d606SStefano Zampini 22119162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22129162d606SStefano Zampini idxs_copied = PETSC_TRUE; 22139162d606SStefano Zampini 2214a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 2215674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 2216a773dcb8SStefano Zampini } else { 2217a773dcb8SStefano Zampini quad_value = 1.0; 2218a773dcb8SStefano Zampini } 2219674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22209162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 2221674ae819SStefano Zampini } 22229162d606SStefano Zampini temp_constraints++; 2223674ae819SStefano Zampini total_counts++; 2224674ae819SStefano Zampini } 2225674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 2226984c4197SStefano Zampini PetscReal real_value; 22279162d606SStefano Zampini PetscScalar *ptr_to_data; 22289162d606SStefano Zampini 2229984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 22309162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 2231674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 22329162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 2233674ae819SStefano Zampini } 2234984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 2235984c4197SStefano Zampini /* check if array is null on the connected component */ 2236e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22379162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 22385b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 2239674ae819SStefano Zampini temp_constraints++; 2240674ae819SStefano Zampini total_counts++; 22419162d606SStefano Zampini if (!idxs_copied) { 22429162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 22439162d606SStefano Zampini idxs_copied = PETSC_TRUE; 2244674ae819SStefano Zampini } 2245674ae819SStefano Zampini } 22469162d606SStefano Zampini } 22479162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 224845a1bb75SStefano Zampini valid_constraints = temp_constraints; 2249eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 2250a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 22519162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 22529162d606SStefano Zampini 22539162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2254a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 22559162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 2256a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 22579162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 2258a773dcb8SStefano Zampini } else { /* perform SVD */ 2259984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 22609162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 2261674ae819SStefano Zampini 2262674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2263984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 2264984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 2265984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 2266984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 2267984c4197SStefano Zampini from that computed using LAPACKgesvd 2268984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 2269984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 2270984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 2271674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 2272e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2273984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2274674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 2275674ae819SStefano Zampini for (k=0;k<j+1;k++) { 22769162d606SStefano 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)); 2277674ae819SStefano Zampini } 2278674ae819SStefano Zampini } 2279e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 2280e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2281e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 2282674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 2283c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 2284674ae819SStefano Zampini #else 2285c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 2286674ae819SStefano Zampini #endif 2287674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2288984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 2289984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 2290674ae819SStefano Zampini j = 0; 2291984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 2292674ae819SStefano Zampini total_counts = total_counts-j; 229345a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 2294e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 2295c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2296c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2297c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 2298c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2299c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 2300c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2301674ae819SStefano Zampini if (j<temp_constraints) { 2302984c4197SStefano Zampini PetscInt ii; 2303984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 2304674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 23059162d606SStefano 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)); 2306674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2307984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 2308674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 23099162d606SStefano 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]; 2310674ae819SStefano Zampini } 2311674ae819SStefano Zampini } 2312674ae819SStefano Zampini } 2313674ae819SStefano Zampini #else /* on missing GESVD */ 2314e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 2315e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 2316b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2317674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2318674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 23199162d606SStefano 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)); 2320674ae819SStefano Zampini #else 23219162d606SStefano 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)); 2322674ae819SStefano Zampini #endif 2323984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 2324674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2325984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 2326e310c8b4SStefano Zampini k = temp_constraints; 2327e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 2328674ae819SStefano Zampini j = 0; 2329e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 233045a1bb75SStefano Zampini valid_constraints = k-j; 2331911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 2332984c4197SStefano Zampini #endif /* on missing GESVD */ 2333674ae819SStefano Zampini } 2334a773dcb8SStefano Zampini } 23359162d606SStefano Zampini /* update pointers information */ 23369162d606SStefano Zampini if (valid_constraints) { 23379162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 23389162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 23399162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 23409162d606SStefano Zampini /* set change_of_basis flag */ 234145a1bb75SStefano Zampini if (boolforchange) { 2342b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 23439162d606SStefano Zampini } 2344b3d85658SStefano Zampini total_counts_cc++; 234545a1bb75SStefano Zampini } 2346674ae819SStefano Zampini } 2347984c4197SStefano Zampini /* free workspace */ 23488f1c130eSStefano Zampini if (!skip_lapack) { 2349984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2350984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 2351984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 2352984c4197SStefano Zampini #endif 2353984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 2354984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 2355984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 2356984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 2357984c4197SStefano Zampini #endif 2358984c4197SStefano Zampini } 2359984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 2360984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 2361984c4197SStefano Zampini } 2362984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 2363cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 2364cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 2365cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 2366cf5a6209SStefano Zampini } 2367cf5a6209SStefano Zampini if (n_ISForFaces) { 2368cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 2369cf5a6209SStefano Zampini } 2370cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 2371cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 2372cf5a6209SStefano Zampini } 2373cf5a6209SStefano Zampini if (n_ISForEdges) { 2374cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 2375cf5a6209SStefano Zampini } 2376cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 237708122e43SStefano Zampini } else { 237808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 237908122e43SStefano Zampini 238008122e43SStefano Zampini total_counts = 0; 238108122e43SStefano Zampini n_vertices = 0; 2382d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 2383d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 238408122e43SStefano Zampini } 238508122e43SStefano Zampini max_constraints = 0; 23869162d606SStefano Zampini total_counts_cc = 0; 238708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 238808122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 23899162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 239008122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 239108122e43SStefano Zampini } 23929162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 23939162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 23949162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 23959162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 239674d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 23979162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 23989162d606SStefano Zampini total_counts_cc = 0; 23999162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 24009162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 24019162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 240208122e43SStefano Zampini } 240308122e43SStefano Zampini } 24049162d606SStefano Zampini #if 0 24059162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 24069162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24079162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 24089162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 24099162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 24109162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 24119162d606SStefano Zampini } 24129162d606SStefano Zampini printf("\n"); 24139162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 24149162d606SStefano Zampini } 24151b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 24168bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 24171b968477SStefano Zampini } 24181b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 24198bec7fa6SStefano 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]); 24201b968477SStefano Zampini } 242108122e43SStefano Zampini #endif 242208122e43SStefano Zampini 24238bec7fa6SStefano Zampini max_size_of_constraint = 0; 24249162d606SStefano 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]); 24259162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 242608122e43SStefano Zampini /* Change of basis */ 2427b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 242808122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 242908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 243008122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 2431b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 243208122e43SStefano Zampini } 243308122e43SStefano Zampini } 243408122e43SStefano Zampini } 243508122e43SStefano Zampini } 2436b371cd4fSStefano Zampini pcbddc->local_primal_size = total_counts; 24370e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 243808122e43SStefano Zampini 24399162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 24409162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 24419162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 24429162d606SStefano 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); 244308122e43SStefano Zampini } 2444984c4197SStefano Zampini 2445674ae819SStefano Zampini /* Create constraint matrix */ 2446674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 244716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 2448984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 2449984c4197SStefano Zampini 2450984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 245174d5cdf7SStefano Zampini /* determine if a QR strategy is needed for change of basis */ 245274d5cdf7SStefano Zampini qr_needed = PETSC_FALSE; 245374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 2454984c4197SStefano Zampini total_primal_vertices=0; 2455b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 24569162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 24579162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2458984c4197SStefano Zampini if (size_of_constraint == 1) { 24599162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 2460b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 246164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 24629162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 24639162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 24649162d606SStefano Zampini } 2465b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 246674d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 246774d5cdf7SStefano Zampini PetscBTSet(qr_needed_idx,i); 246874d5cdf7SStefano Zampini qr_needed = PETSC_TRUE; 246974d5cdf7SStefano Zampini } 2470b3d85658SStefano Zampini } else { 2471b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 2472984c4197SStefano Zampini } 2473984c4197SStefano Zampini } 2474b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 2475b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 247670022509SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 247770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 2478b3d85658SStefano Zampini 2479b3d85658SStefano 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); 24800e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 24810e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 248270022509SStefano Zampini 2483984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 248474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 2485785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 2486984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 248774d5cdf7SStefano Zampini 2488984c4197SStefano Zampini j = total_primal_vertices; 248974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 2490b3d85658SStefano Zampini cum = total_primal_vertices; 24919162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 24924641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 2493b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 2494b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 2495b3d85658SStefano Zampini cum++; 24969162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 249774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 249874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 249974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 250074d5cdf7SStefano Zampini } 25019162d606SStefano Zampini j += constraints_n[i]; 2502674ae819SStefano Zampini } 2503674ae819SStefano Zampini } 2504674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 2505674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2506088faed8SStefano Zampini 2507674ae819SStefano Zampini /* set values in constraint matrix */ 2508984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 25090e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 2510674ae819SStefano Zampini } 2511984c4197SStefano Zampini total_counts = total_primal_vertices; 25129162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 25134641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 25149162d606SStefano Zampini PetscInt *cols; 25159162d606SStefano Zampini 25169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 25179162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 25189162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 25199162d606SStefano Zampini PetscInt row = total_counts+k; 25209162d606SStefano Zampini PetscScalar *vals; 25219162d606SStefano Zampini 25229162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 25239162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 25249162d606SStefano Zampini } 25259162d606SStefano Zampini total_counts += constraints_n[i]; 2526674ae819SStefano Zampini } 2527674ae819SStefano Zampini } 2528674ae819SStefano Zampini /* assembling */ 2529674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2530674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2531088faed8SStefano Zampini 2532984c4197SStefano Zampini /* 253345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2534984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 2535984c4197SStefano Zampini */ 2536674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 2537674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 2538026de310SStefano Zampini /* dual and primal dofs on a single cc */ 2539984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 2540984c4197SStefano Zampini /* working stuff for GEQRF */ 254181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 2542984c4197SStefano Zampini PetscBLASInt lqr_work; 2543984c4197SStefano Zampini /* working stuff for UNGQR */ 2544984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 2545984c4197SStefano Zampini PetscBLASInt lgqr_work; 2546984c4197SStefano Zampini /* working stuff for TRTRS */ 2547984c4197SStefano Zampini PetscScalar *trs_rhs; 25483f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 2549984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 2550984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 2551984c4197SStefano Zampini PetscScalar *start_vals; 2552984c4197SStefano Zampini /* working stuff for values insertion */ 25534641a718SStefano Zampini PetscBT is_primal; 255464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 2555906d46d4SStefano Zampini /* matrix sizes */ 2556906d46d4SStefano Zampini PetscInt global_size,local_size; 2557906d46d4SStefano Zampini /* temporary change of basis */ 2558906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 2559cf5a6209SStefano Zampini /* extra space for debugging */ 2560cf5a6209SStefano Zampini PetscScalar *dbg_work; 2561984c4197SStefano Zampini 2562906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 2563906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 256416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 2565bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 2566906d46d4SStefano Zampini /* nonzeros for local mat */ 2567bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 2568bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 25699162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 2570a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 25719162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 2572a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 25739162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 2574a717540cSStefano Zampini } else { 25759162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 25769162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 2577a717540cSStefano Zampini } 2578a717540cSStefano Zampini } 2579a717540cSStefano Zampini } 2580906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 2581bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 2582a717540cSStefano Zampini /* Set initial identity in the matrix */ 2583bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 2584906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 2585a717540cSStefano Zampini } 2586a717540cSStefano Zampini 2587a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2588a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 2589a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 2590a717540cSStefano Zampini } 2591a717540cSStefano Zampini 2592a717540cSStefano Zampini 2593a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 2594a717540cSStefano Zampini /* 2595a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 2596a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 2597a717540cSStefano Zampini 2598a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 2599a717540cSStefano Zampini 2600a6b551f4SStefano 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) 2601a6b551f4SStefano Zampini 2602a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 2603a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 2604a717540cSStefano Zampini | ... | 2605a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 2606a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 2607a717540cSStefano Zampini 2608a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 2609a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 2610a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 2611a6b551f4SStefano Zampini 2612a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 2613a717540cSStefano Zampini */ 2614a717540cSStefano Zampini if (qr_needed) { 2615984c4197SStefano Zampini /* space to store Q */ 2616854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 2617984c4197SStefano Zampini /* first we issue queries for optimal work */ 26183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26193f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 26203f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2621984c4197SStefano Zampini lqr_work = -1; 26223f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 2623984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 2624984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 2625785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 2626984c4197SStefano Zampini lgqr_work = -1; 26273f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 26283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 26293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 26303f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26313f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 26323f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 2633984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 2634984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 2635785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 2636984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 2637785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 2638984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 2639785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 2640a717540cSStefano Zampini /* allocating workspace for check */ 2641a717540cSStefano Zampini if (pcbddc->dbg_flag) { 2642cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 2643a717540cSStefano Zampini } 2644a717540cSStefano Zampini } 2645984c4197SStefano Zampini /* array to store whether a node is primal or not */ 26464641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 2647473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 26480e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 264939e2fb2aSStefano Zampini if (i != total_primal_vertices) { 265039e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 26514641a718SStefano Zampini } 265239e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 265339e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 265439e2fb2aSStefano Zampini } 265539e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 2656984c4197SStefano Zampini 2657a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 26589162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 26599162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 26604641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 2661984c4197SStefano Zampini /* get constraint info */ 26629162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 2663984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 2664984c4197SStefano Zampini 2665984c4197SStefano Zampini if (pcbddc->dbg_flag) { 26669162d606SStefano 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); 2667674ae819SStefano Zampini } 2668984c4197SStefano Zampini 2669fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 2670a717540cSStefano Zampini 2671a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 2672a717540cSStefano Zampini if (pcbddc->dbg_flag) { 26739162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2674a717540cSStefano Zampini } 2675984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 26769162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2677984c4197SStefano Zampini 2678984c4197SStefano Zampini /* compute QR decomposition of constraints */ 26793f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 26803f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2682674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26833f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 2684984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 2685674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2686984c4197SStefano Zampini 2687984c4197SStefano Zampini /* explictly compute R^-T */ 2688984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 2689984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 26903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 26913f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 26923f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 26933f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 2694984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 26953f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 2696984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 2697984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2698984c4197SStefano Zampini 2699a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 27003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2704984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27053f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 2706984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 2707984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2708984c4197SStefano Zampini 2709984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 2710984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 2711984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 27123f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 27133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 27143f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 27153f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 27163f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 27173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 2718984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 27199162d606SStefano 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)); 2720984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 27219162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 2722984c4197SStefano Zampini 2723984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 27249162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 2725984c4197SStefano Zampini /* insert cols for primal dofs */ 2726984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 2727984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 27289162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2729906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2730984c4197SStefano Zampini } 2731984c4197SStefano Zampini /* insert cols for dual dofs */ 2732984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 27339162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 2734984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 27359162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2736906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 2737984c4197SStefano Zampini j++; 2738674ae819SStefano Zampini } 2739674ae819SStefano Zampini } 2740984c4197SStefano Zampini 2741984c4197SStefano Zampini /* check change of basis */ 2742984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2743984c4197SStefano Zampini PetscInt ii,jj; 2744984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 2745c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 2746c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 2747c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 2748c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 2749c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 2750c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 2751984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 2752cf5a6209SStefano 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)); 2753984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 2754984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2755984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2756cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 2757cf5a6209SStefano 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; 2758674ae819SStefano Zampini } 2759674ae819SStefano Zampini } 2760984c4197SStefano Zampini if (!valid_qr) { 276122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 2762984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 2763984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 2764cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 2765cf5a6209SStefano 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])); 2766674ae819SStefano Zampini } 2767cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 2768cf5a6209SStefano 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])); 2769984c4197SStefano Zampini } 2770984c4197SStefano Zampini } 2771984c4197SStefano Zampini } 2772674ae819SStefano Zampini } else { 277322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 2774674ae819SStefano Zampini } 2775674ae819SStefano Zampini } 2776a717540cSStefano Zampini } else { /* simple transformation block */ 2777a717540cSStefano Zampini PetscInt row,col; 2778a6b551f4SStefano Zampini PetscScalar val,norm; 2779a6b551f4SStefano Zampini 2780a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 27819162d606SStefano 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)); 2782a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 27839162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 27849162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 2785bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 27869162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 2787906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 27889162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 2789a717540cSStefano Zampini } else { 2790a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 27919162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 2792a717540cSStefano Zampini if (row != col) { 27939162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 2794a717540cSStefano Zampini } else { 27959162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 2796a717540cSStefano Zampini } 2797906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 2798a717540cSStefano Zampini } 2799a717540cSStefano Zampini } 2800a717540cSStefano Zampini } 280198a51de6SStefano Zampini if (pcbddc->dbg_flag) { 280222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 2803a717540cSStefano Zampini } 2804674ae819SStefano Zampini } 2805984c4197SStefano Zampini } else { 2806984c4197SStefano Zampini if (pcbddc->dbg_flag) { 28079162d606SStefano 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); 2808674ae819SStefano Zampini } 2809674ae819SStefano Zampini } 2810674ae819SStefano Zampini } 2811a717540cSStefano Zampini 2812a717540cSStefano Zampini /* free workspace */ 2813a717540cSStefano Zampini if (qr_needed) { 2814984c4197SStefano Zampini if (pcbddc->dbg_flag) { 2815cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 2816984c4197SStefano Zampini } 2817984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 2818984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 2819984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 2820984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 2821984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 2822674ae819SStefano Zampini } 2823a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 2824906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2825906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2826906d46d4SStefano Zampini 2827906d46d4SStefano Zampini /* assembling of global change of variable */ 2828bbb9e6c6SStefano Zampini { 2829bbb9e6c6SStefano Zampini Mat tmat; 283016f15bc4SStefano Zampini PetscInt bs; 283116f15bc4SStefano Zampini 2832906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2833906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2834bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 2835bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 2836bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2837bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 283816f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 283916f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 2840906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2841bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 2842bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2843bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2844bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 2845bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 2846bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2847bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2848bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 2849bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 2850906d46d4SStefano Zampini } 2851906d46d4SStefano Zampini /* check */ 2852906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2853906d46d4SStefano Zampini PetscReal error; 2854906d46d4SStefano Zampini Vec x,x_change; 2855906d46d4SStefano Zampini 2856906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 2857906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 2858906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 2859906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 2860bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2861bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2862bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 2863bbb9e6c6SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2864bbb9e6c6SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2865906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 2866906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2867906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2868906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2869bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 2870906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2871906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2872906d46d4SStefano Zampini } 2873b96c3477SStefano Zampini 2874b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 2875b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 2876b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 2877b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 2878ac632422SStefano Zampini Mat S_new,tmat; 2879bbb9e6c6SStefano Zampini IS is_all_N; 2880bbb9e6c6SStefano Zampini 2881bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 28826816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 2883bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 2884ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2885b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 2886ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2887ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 2888ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2889ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 2890ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 2891b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 2892ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 2893ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 2894ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 2895ac632422SStefano Zampini } 2896b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2897b96c3477SStefano Zampini } 2898b96c3477SStefano Zampini } 2899906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 2900906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 2901b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2902b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 2903b9b85e73SStefano Zampini } 2904906d46d4SStefano Zampini 2905906d46d4SStefano Zampini /* set up change of basis context */ 2906906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 2907906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 2908906d46d4SStefano Zampini 2909906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 2910906d46d4SStefano Zampini PetscInt global_size,local_size; 2911906d46d4SStefano Zampini 2912906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 2913906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 2914906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 2915906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 2916906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 2917906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 2918906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 2919906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 2920906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 2921906d46d4SStefano Zampini } else { 2922906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 2923906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 2924906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 2925906d46d4SStefano Zampini } 2926906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 2927906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 2928906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 2929906d46d4SStefano Zampini } else { 2930906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 2931906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 2932906d46d4SStefano Zampini } 2933906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 2934906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 2935906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2936906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2937b9b85e73SStefano Zampini } 2938a717540cSStefano Zampini 2939727cdba6SStefano Zampini /* check if a new primal space has been introduced */ 2940727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 2941727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 29420e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 29430e6343abSStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 29440e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 29450e6343abSStefano Zampini ierr = PetscMemcmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);CHKERRQ(ierr); 2946c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 2947727cdba6SStefano Zampini } 29480e6343abSStefano Zampini } 29490e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 2950727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 2951727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2952727cdba6SStefano Zampini 2953a717540cSStefano Zampini /* flush dbg viewer */ 2954b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 2955b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2956b8ffe317SStefano Zampini } 2957a717540cSStefano Zampini 2958e310c8b4SStefano Zampini /* free workspace */ 2959a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 29604641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 296108122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 29629162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 29639162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 296408122e43SStefano Zampini } else { 29659162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 29669162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 29679162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 296808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 296908122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 29709162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 29719162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 297208122e43SStefano Zampini } 2973674ae819SStefano Zampini PetscFunctionReturn(0); 2974674ae819SStefano Zampini } 2975674ae819SStefano Zampini 2976674ae819SStefano Zampini #undef __FUNCT__ 2977674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 2978674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 2979674ae819SStefano Zampini { 2980674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2981674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2982674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 29837fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 2984674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 2985674ae819SStefano Zampini 2986674ae819SStefano Zampini PetscFunctionBegin; 29878e61c736SStefano Zampini /* Reset previously computed graph */ 29888e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 2989674ae819SStefano Zampini /* Init local Graph struct */ 29907fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 29917fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,matis->mapping,N);CHKERRQ(ierr); 2992674ae819SStefano Zampini 2993575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 2994575ad6abSStefano Zampini if (pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 2995575ad6abSStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 2996575ad6abSStefano Zampini } 29979577ea80SStefano Zampini 2998674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 29994d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 30004d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 30014d379d7bSStefano Zampini PetscInt nvtxs; 3002e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 3003674ae819SStefano Zampini 30044d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 30052fffb893SStefano Zampini 30062fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30072fffb893SStefano Zampini if (flg_row) { 30084d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 3009b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 30102fffb893SStefano Zampini } 30112fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30129b28b941SStefano 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 */ 30134d379d7bSStefano Zampini IS is_dummy; 30144d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 30154d379d7bSStefano Zampini PetscInt j,sum; 30164d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 30174d379d7bSStefano Zampini const PetscInt *idxs; 30184d379d7bSStefano Zampini PCBDDCGraph graph; 30194d379d7bSStefano Zampini PetscBT is_on_boundary; 30204d379d7bSStefano Zampini 30214d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 30224d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 30234d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 30244d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 30257fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 30264d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3027e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3028e496cd5dSStefano Zampini if (flg_row) { 30294d379d7bSStefano Zampini graph->xadj = xadj; 30304d379d7bSStefano Zampini graph->adjncy = adjncy; 3031e496cd5dSStefano Zampini } 30324d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 30334d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3034e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 30354d379d7bSStefano Zampini 30364d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 30379b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 30384d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30394d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 30404d379d7bSStefano Zampini } 30414d379d7bSStefano Zampini } 30424d379d7bSStefano Zampini 3043e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 30444d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30454d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 30464d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 30474d379d7bSStefano Zampini } 30484d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 30494d379d7bSStefano Zampini 3050e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 30514d379d7bSStefano Zampini sum = 0; 30524d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30534d379d7bSStefano Zampini PetscInt sizecc = 0; 30544d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30554d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30564d379d7bSStefano Zampini sizecc++; 30574d379d7bSStefano Zampini } 30584d379d7bSStefano Zampini } 30594d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30604d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30614d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 30624d379d7bSStefano Zampini } 30634d379d7bSStefano Zampini } 30644d379d7bSStefano Zampini sum += sizecc*sizecc; 30654d379d7bSStefano Zampini } 30664d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 30674d379d7bSStefano Zampini sum = 0; 3068e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 30694d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 30704d379d7bSStefano Zampini cxadj[i] = sum; 30714d379d7bSStefano Zampini sum += temp; 30724d379d7bSStefano Zampini } 3073e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 30744d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 30754d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 30764d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 30774d379d7bSStefano Zampini PetscInt k,sizecc = 0; 30784d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 30794d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 30804d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 30814d379d7bSStefano Zampini sizecc++; 30824d379d7bSStefano Zampini } 30834d379d7bSStefano Zampini } 30844d379d7bSStefano Zampini } 30854d379d7bSStefano Zampini } 30864d379d7bSStefano Zampini } 30879b28b941SStefano Zampini if (sum) { 3088e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 30894d379d7bSStefano Zampini } else { 30904d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 30914d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 30924d379d7bSStefano Zampini } 30934d379d7bSStefano Zampini graph->xadj = 0; 30944d379d7bSStefano Zampini graph->adjncy = 0; 30954d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 30964d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 30974d379d7bSStefano Zampini } 3098674ae819SStefano Zampini } 30999b28b941SStefano Zampini if (pcbddc->dbg_flag) { 31009b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31019b28b941SStefano Zampini } 3102674ae819SStefano Zampini 310363602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 3104674ae819SStefano Zampini vertex_size = 1; 310563602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 310663602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 310795ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 310863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 3109a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 311063602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 3111674ae819SStefano Zampini } 311263602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 311363602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 311463602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 3115674ae819SStefano Zampini } 311663602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 3117674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 311863602bcaSStefano Zampini } else { 311963602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 312063602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 3121854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 312263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 312363602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 312463602bcaSStefano Zampini } 312563602bcaSStefano Zampini } 3126674ae819SStefano Zampini } 3127674ae819SStefano Zampini 3128674ae819SStefano Zampini /* Setup of Graph */ 3129785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 3130a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 3131785d1243SStefano Zampini } 3132785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 3133a7dc3881SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->ctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 3134785d1243SStefano Zampini } 3135302440fdSBarry Smith ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices);CHKERRQ(ierr); 3136674ae819SStefano Zampini 3137674ae819SStefano Zampini /* Graph's connected components analysis */ 3138674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 3139674ae819SStefano Zampini 3140674ae819SStefano Zampini /* print some info to stdout */ 3141674ae819SStefano Zampini if (pcbddc->dbg_flag) { 3142302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 3143674ae819SStefano Zampini } 3144fb180af4SStefano Zampini 3145fb180af4SStefano Zampini /* mark topography has done */ 3146fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 3147674ae819SStefano Zampini PetscFunctionReturn(0); 3148674ae819SStefano Zampini } 3149674ae819SStefano Zampini 3150674ae819SStefano Zampini #undef __FUNCT__ 3151674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 3152674ae819SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(MPI_Comm comm,ISLocalToGlobalMapping l2gmap, PetscInt n_local_dofs, PetscInt local_dofs[], PetscInt local_dofs_mult[], PetscInt* n_global_subset, PetscInt* global_numbering_subset[]) 3153674ae819SStefano Zampini { 3154674ae819SStefano Zampini Vec local_vec,global_vec; 3155674ae819SStefano Zampini IS seqis,paris; 3156674ae819SStefano Zampini VecScatter scatter_ctx; 3157674ae819SStefano Zampini PetscScalar *array; 3158674ae819SStefano Zampini PetscInt *temp_global_dofs; 3159674ae819SStefano Zampini PetscScalar globalsum; 3160674ae819SStefano Zampini PetscInt i,j,s; 316117ccb9bdSStefano Zampini PetscInt nlocals,first_index,old_index,max_local,max_global; 316217ccb9bdSStefano Zampini PetscMPIInt rank_prec_comm,size_prec_comm; 316317ccb9bdSStefano Zampini PetscInt *dof_sizes,*dof_displs; 3164674ae819SStefano Zampini PetscBool first_found; 3165674ae819SStefano Zampini PetscErrorCode ierr; 3166674ae819SStefano Zampini 3167674ae819SStefano Zampini PetscFunctionBegin; 3168674ae819SStefano Zampini /* mpi buffers */ 3169b9b85e73SStefano Zampini ierr = MPI_Comm_size(comm,&size_prec_comm);CHKERRQ(ierr); 3170b9b85e73SStefano Zampini ierr = MPI_Comm_rank(comm,&rank_prec_comm);CHKERRQ(ierr); 3171674ae819SStefano Zampini j = ( !rank_prec_comm ? size_prec_comm : 0); 317217ccb9bdSStefano Zampini ierr = PetscMalloc2(j,&dof_sizes,j,&dof_displs);CHKERRQ(ierr); 3173674ae819SStefano Zampini /* get maximum size of subset */ 3174785e854fSJed Brown ierr = PetscMalloc1(n_local_dofs,&temp_global_dofs);CHKERRQ(ierr); 3175674ae819SStefano Zampini ierr = ISLocalToGlobalMappingApply(l2gmap,n_local_dofs,local_dofs,temp_global_dofs);CHKERRQ(ierr); 3176674ae819SStefano Zampini max_local = 0; 3177b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3178674ae819SStefano Zampini if (max_local < temp_global_dofs[i] ) { 3179674ae819SStefano Zampini max_local = temp_global_dofs[i]; 3180674ae819SStefano Zampini } 3181674ae819SStefano Zampini } 3182b9b85e73SStefano Zampini ierr = MPI_Allreduce(&max_local,&max_global,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr); 3183674ae819SStefano Zampini max_global++; 3184674ae819SStefano Zampini max_local = 0; 3185b9b85e73SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3186674ae819SStefano Zampini if (max_local < local_dofs[i] ) { 3187674ae819SStefano Zampini max_local = local_dofs[i]; 3188674ae819SStefano Zampini } 3189674ae819SStefano Zampini } 3190674ae819SStefano Zampini max_local++; 3191674ae819SStefano Zampini /* allocate workspace */ 3192674ae819SStefano Zampini ierr = VecCreate(PETSC_COMM_SELF,&local_vec);CHKERRQ(ierr); 3193674ae819SStefano Zampini ierr = VecSetSizes(local_vec,PETSC_DECIDE,max_local);CHKERRQ(ierr); 3194674ae819SStefano Zampini ierr = VecSetType(local_vec,VECSEQ);CHKERRQ(ierr); 3195674ae819SStefano Zampini ierr = VecCreate(comm,&global_vec);CHKERRQ(ierr); 3196674ae819SStefano Zampini ierr = VecSetSizes(global_vec,PETSC_DECIDE,max_global);CHKERRQ(ierr); 3197674ae819SStefano Zampini ierr = VecSetType(global_vec,VECMPI);CHKERRQ(ierr); 3198674ae819SStefano Zampini /* create scatter */ 3199674ae819SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n_local_dofs,local_dofs,PETSC_COPY_VALUES,&seqis);CHKERRQ(ierr); 3200674ae819SStefano Zampini ierr = ISCreateGeneral(comm,n_local_dofs,temp_global_dofs,PETSC_COPY_VALUES,&paris);CHKERRQ(ierr); 3201674ae819SStefano Zampini ierr = VecScatterCreate(local_vec,seqis,global_vec,paris,&scatter_ctx);CHKERRQ(ierr); 3202674ae819SStefano Zampini ierr = ISDestroy(&seqis);CHKERRQ(ierr); 3203674ae819SStefano Zampini ierr = ISDestroy(&paris);CHKERRQ(ierr); 3204674ae819SStefano Zampini /* init array */ 3205674ae819SStefano Zampini ierr = VecSet(global_vec,0.0);CHKERRQ(ierr); 3206674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3207674ae819SStefano Zampini ierr = VecGetArray(local_vec,&array);CHKERRQ(ierr); 3208674ae819SStefano Zampini if (local_dofs_mult) { 3209674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3210674ae819SStefano Zampini array[local_dofs[i]]=(PetscScalar)local_dofs_mult[i]; 3211674ae819SStefano Zampini } 3212674ae819SStefano Zampini } else { 3213674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 3214674ae819SStefano Zampini array[local_dofs[i]]=1.0; 3215674ae819SStefano Zampini } 3216674ae819SStefano Zampini } 3217674ae819SStefano Zampini ierr = VecRestoreArray(local_vec,&array);CHKERRQ(ierr); 3218674ae819SStefano Zampini /* scatter into global vec and get total number of global dofs */ 3219674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3220674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,local_vec,global_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3221674ae819SStefano Zampini ierr = VecSum(global_vec,&globalsum);CHKERRQ(ierr); 32225b08dc53SStefano Zampini *n_global_subset = (PetscInt)PetscRealPart(globalsum); 3223674ae819SStefano Zampini /* Fill global_vec with cumulative function for global numbering */ 3224674ae819SStefano Zampini ierr = VecGetArray(global_vec,&array);CHKERRQ(ierr); 3225674ae819SStefano Zampini ierr = VecGetLocalSize(global_vec,&s);CHKERRQ(ierr); 3226674ae819SStefano Zampini nlocals = 0; 3227674ae819SStefano Zampini first_index = -1; 3228674ae819SStefano Zampini first_found = PETSC_FALSE; 3229674ae819SStefano Zampini for (i=0;i<s;i++) { 3230b9b85e73SStefano Zampini if (!first_found && PetscRealPart(array[i]) > 0.1) { 3231674ae819SStefano Zampini first_found = PETSC_TRUE; 3232674ae819SStefano Zampini first_index = i; 3233674ae819SStefano Zampini } 32345b08dc53SStefano Zampini nlocals += (PetscInt)PetscRealPart(array[i]); 3235674ae819SStefano Zampini } 3236674ae819SStefano Zampini ierr = MPI_Gather(&nlocals,1,MPIU_INT,dof_sizes,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3237674ae819SStefano Zampini if (!rank_prec_comm) { 3238674ae819SStefano Zampini dof_displs[0]=0; 3239674ae819SStefano Zampini for (i=1;i<size_prec_comm;i++) { 3240674ae819SStefano Zampini dof_displs[i] = dof_displs[i-1]+dof_sizes[i-1]; 3241674ae819SStefano Zampini } 3242674ae819SStefano Zampini } 3243674ae819SStefano Zampini ierr = MPI_Scatter(dof_displs,1,MPIU_INT,&nlocals,1,MPIU_INT,0,comm);CHKERRQ(ierr); 3244674ae819SStefano Zampini if (first_found) { 3245674ae819SStefano Zampini array[first_index] += (PetscScalar)nlocals; 3246674ae819SStefano Zampini old_index = first_index; 3247674ae819SStefano Zampini for (i=first_index+1;i<s;i++) { 3248b9b85e73SStefano Zampini if (PetscRealPart(array[i]) > 0.1) { 3249674ae819SStefano Zampini array[i] += array[old_index]; 3250674ae819SStefano Zampini old_index = i; 3251674ae819SStefano Zampini } 3252674ae819SStefano Zampini } 3253674ae819SStefano Zampini } 3254674ae819SStefano Zampini ierr = VecRestoreArray(global_vec,&array);CHKERRQ(ierr); 3255674ae819SStefano Zampini ierr = VecSet(local_vec,0.0);CHKERRQ(ierr); 3256674ae819SStefano Zampini ierr = VecScatterBegin(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3257674ae819SStefano Zampini ierr = VecScatterEnd(scatter_ctx,global_vec,local_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3258674ae819SStefano Zampini /* get global ordering of local dofs */ 32592b095fd8SStefano Zampini ierr = VecGetArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3260674ae819SStefano Zampini if (local_dofs_mult) { 3261674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32625b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-local_dofs_mult[i]; 3263674ae819SStefano Zampini } 3264674ae819SStefano Zampini } else { 3265674ae819SStefano Zampini for (i=0;i<n_local_dofs;i++) { 32665b08dc53SStefano Zampini temp_global_dofs[i] = (PetscInt)PetscRealPart(array[local_dofs[i]])-1; 3267674ae819SStefano Zampini } 3268674ae819SStefano Zampini } 32692b095fd8SStefano Zampini ierr = VecRestoreArrayRead(local_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 3270674ae819SStefano Zampini /* free workspace */ 3271674ae819SStefano Zampini ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr); 3272674ae819SStefano Zampini ierr = VecDestroy(&local_vec);CHKERRQ(ierr); 3273674ae819SStefano Zampini ierr = VecDestroy(&global_vec);CHKERRQ(ierr); 327417ccb9bdSStefano Zampini ierr = PetscFree2(dof_sizes,dof_displs);CHKERRQ(ierr); 3275674ae819SStefano Zampini /* return pointer to global ordering of local dofs */ 3276674ae819SStefano Zampini *global_numbering_subset = temp_global_dofs; 3277674ae819SStefano Zampini PetscFunctionReturn(0); 3278674ae819SStefano Zampini } 32799a7d3425SStefano Zampini 32809a7d3425SStefano Zampini #undef __FUNCT__ 32819a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 32829a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 32839a7d3425SStefano Zampini { 32849a7d3425SStefano Zampini PetscInt i,j; 32859a7d3425SStefano Zampini PetscScalar *alphas; 32869a7d3425SStefano Zampini PetscErrorCode ierr; 32879a7d3425SStefano Zampini 32889a7d3425SStefano Zampini PetscFunctionBegin; 32899a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 3290785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 32919a7d3425SStefano Zampini for (i=0;i<n;i++) { 32929a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 32939a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 32949a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 32959a7d3425SStefano Zampini } 32969a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 32979a7d3425SStefano Zampini PetscFunctionReturn(0); 32989a7d3425SStefano Zampini } 32999a7d3425SStefano Zampini 3300e7931f94SStefano Zampini #undef __FUNCT__ 330170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 330228143c3dSStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscBool contiguous, IS* is_sends) 3303e7931f94SStefano Zampini { 3304e7931f94SStefano Zampini Mat subdomain_adj; 3305e7931f94SStefano Zampini IS new_ranks,ranks_send_to; 3306e7931f94SStefano Zampini MatPartitioning partitioner; 3307e7931f94SStefano Zampini Mat_IS *matis; 3308e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 3309e7931f94SStefano Zampini PetscInt prank; 3310e7931f94SStefano Zampini PetscMPIInt size,rank,color; 3311e7931f94SStefano Zampini PetscInt *xadj,*adjncy,*oldranks; 3312e7931f94SStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*is_indices,*ranks_send_to_idx; 33133837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 3314e7931f94SStefano Zampini PetscErrorCode ierr; 33152b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 3316e7931f94SStefano Zampini PetscSubcomm subcomm; 3317a57a6d2fSStefano Zampini 3318e7931f94SStefano Zampini PetscFunctionBegin; 33192b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 33202b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 33212b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 3322e7931f94SStefano Zampini 3323e7931f94SStefano Zampini /* Get info on mapping */ 3324e7931f94SStefano Zampini matis = (Mat_IS*)(mat->data); 3325e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&local_size);CHKERRQ(ierr); 3326e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3327e7931f94SStefano Zampini 3328e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 3329785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 3330e7931f94SStefano Zampini xadj[0] = 0; 3331e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 3332785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 3333785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 3334e7931f94SStefano Zampini 33352b510759SStefano Zampini if (threshold) { 3336d023bfaeSStefano Zampini PetscInt xadj_count = 0; 33372b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 3338d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 3339d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 3340d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 3341d023bfaeSStefano Zampini xadj_count++; 3342e7931f94SStefano Zampini } 3343e7931f94SStefano Zampini } 3344d023bfaeSStefano Zampini xadj[1] = xadj_count; 3345c8587f34SStefano Zampini } else { 3346e7931f94SStefano Zampini if (xadj[1]) { 3347e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 3348e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 3349c8587f34SStefano Zampini } 3350e7931f94SStefano Zampini } 3351e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(matis->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 3352e7931f94SStefano Zampini if (use_square) { 3353e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3354e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 3355e7931f94SStefano Zampini } 3356e7931f94SStefano Zampini } 3357e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3358e7931f94SStefano Zampini 33593837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 3360e7931f94SStefano Zampini 3361e7931f94SStefano Zampini /* 3362e7931f94SStefano Zampini Restrict work on active processes only. 3363e7931f94SStefano Zampini */ 3364e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 3365e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 3366e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 33672b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 3368d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3369e7931f94SStefano Zampini if (color) { 3370e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 3371e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 3372e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 3373c8587f34SStefano Zampini } else { 337428143c3dSStefano Zampini PetscInt coarsening_ratio; 3375306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 3376785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 3377e7931f94SStefano Zampini prank = rank; 3378306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 33798002ef2cSStefano Zampini /* 3380e7931f94SStefano Zampini for (i=0;i<size;i++) { 3381e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 3382c8587f34SStefano Zampini } 33838002ef2cSStefano Zampini */ 3384e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 3385e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 3386c8587f34SStefano Zampini } 3387e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 3388306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 338922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 3390e7931f94SStefano Zampini 3391e7931f94SStefano Zampini /* Partition */ 3392306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 3393e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 3394e7931f94SStefano Zampini if (use_vwgt) { 33953837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 3396e7931f94SStefano Zampini v_wgt[0] = local_size; 3397e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 3398c8587f34SStefano Zampini } 339928143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 340028143c3dSStefano Zampini coarsening_ratio = size/n_subdomains; 340128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 3402e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 3403e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 340422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 3405e7931f94SStefano Zampini 3406e7931f94SStefano Zampini ierr = ISGetIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 340728143c3dSStefano Zampini if (contiguous) { 340828143c3dSStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; /* contiguos set of processes */ 340928143c3dSStefano Zampini } else { 341084ba6060SStefano Zampini ranks_send_to_idx[0] = coarsening_ratio*oldranks[is_indices[0]]; /* scattered set of processes */ 341128143c3dSStefano Zampini } 3412e7931f94SStefano Zampini ierr = ISRestoreIndices(new_ranks,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3413e7931f94SStefano Zampini /* clean up */ 3414e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 3415e7931f94SStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 3416e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 3417e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 3418e7931f94SStefano Zampini } 3419e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 3420e7931f94SStefano Zampini 3421e7931f94SStefano Zampini /* assemble parallel IS for sends */ 3422e7931f94SStefano Zampini i = 1; 3423e7931f94SStefano Zampini if (color) i=0; 3424e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 3425e7931f94SStefano Zampini 3426e7931f94SStefano Zampini /* get back IS */ 3427e7931f94SStefano Zampini *is_sends = ranks_send_to; 3428e7931f94SStefano Zampini PetscFunctionReturn(0); 3429e7931f94SStefano Zampini } 3430e7931f94SStefano Zampini 3431e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 3432e7931f94SStefano Zampini 3433e7931f94SStefano Zampini #undef __FUNCT__ 3434e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 343528143c3dSStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 3436e7931f94SStefano Zampini { 343770cf5478SStefano Zampini Mat local_mat; 3438e7931f94SStefano Zampini Mat_IS *matis; 3439e7931f94SStefano Zampini IS is_sends_internal; 34409d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 344128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 34429d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 3443e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 3444e7931f94SStefano Zampini PetscInt* l2gmap_indices; 3445e7931f94SStefano Zampini const PetscInt* is_indices; 3446e7931f94SStefano Zampini MatType new_local_type; 3447e7931f94SStefano Zampini /* buffers */ 3448e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 344928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 34509d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 3451e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 3452e7931f94SStefano Zampini /* MPI */ 345328143c3dSStefano Zampini MPI_Comm comm,comm_n; 345428143c3dSStefano Zampini PetscSubcomm subcomm; 3455e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 345628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 345728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 345828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 345928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 346028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 3461e7931f94SStefano Zampini PetscErrorCode ierr; 3462e7931f94SStefano Zampini 3463e7931f94SStefano Zampini PetscFunctionBegin; 346428143c3dSStefano Zampini /* TODO: add missing checks */ 346528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 346628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 346728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 346828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 3469e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 347028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 3471e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3472e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 3473e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 3474e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 3475e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 347628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 347770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 347870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 347928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 348070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 348170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 348270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 348370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 348470cf5478SStefano Zampini } 3485e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 3486e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 3487e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 3488e7931f94SStefano Zampini if (!is_sends) { 348928143c3dSStefano Zampini PetscBool pcontig = PETSC_TRUE; 349028143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 349128143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,pcontig,&is_sends_internal);CHKERRQ(ierr); 3492c8587f34SStefano Zampini } else { 3493e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 3494e7931f94SStefano Zampini is_sends_internal = is_sends; 3495c8587f34SStefano Zampini } 3496e7931f94SStefano Zampini 3497e7931f94SStefano Zampini /* get pointer of MATIS data */ 3498e7931f94SStefano Zampini matis = (Mat_IS*)mat->data; 3499e7931f94SStefano Zampini 3500e7931f94SStefano Zampini /* get comm */ 3501a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3502e7931f94SStefano Zampini 3503e7931f94SStefano Zampini /* compute number of sends */ 3504e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 3505e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 3506e7931f94SStefano Zampini 3507e7931f94SStefano Zampini /* compute number of receives */ 3508e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 3509785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 3510e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 3511e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3512e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 3513e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 3514e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 3515e7931f94SStefano Zampini 351628143c3dSStefano Zampini /* restrict comm if requested */ 351728143c3dSStefano Zampini subcomm = 0; 351828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 351928143c3dSStefano Zampini if (restrict_comm) { 3520779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 3521779c1cceSStefano Zampini 352228143c3dSStefano Zampini color = 0; 35236366ad86SStefano Zampini if (!n_recvs && n_sends) color = 1; /* processes sending and not receiving anything will not partecipate in new comm */ 352428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 352528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 352628143c3dSStefano Zampini /* check if reuse has been requested */ 352728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 352828143c3dSStefano Zampini if (*mat_n) { 352928143c3dSStefano Zampini PetscMPIInt subcommsize2; 353028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 353128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 353228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 353328143c3dSStefano Zampini } else { 353428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 353528143c3dSStefano Zampini } 353628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 3537779c1cceSStefano Zampini PetscMPIInt rank; 3538779c1cceSStefano Zampini 3539779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 354028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 354128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 354228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 3543306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 354428143c3dSStefano Zampini } 354528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 354628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 354728143c3dSStefano Zampini } else { 354828143c3dSStefano Zampini comm_n = comm; 354928143c3dSStefano Zampini } 355028143c3dSStefano Zampini 3551e7931f94SStefano Zampini /* prepare send/receive buffers */ 3552785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 3553e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 3554785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 3555e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 355628143c3dSStefano Zampini if (nis) { 3557854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 355828143c3dSStefano Zampini } 3559e7931f94SStefano Zampini 356028143c3dSStefano Zampini /* Get data from local matrices */ 3561e7931f94SStefano Zampini if (!isdense) { 3562a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 3563e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 3564e7931f94SStefano Zampini /* 3565e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 3566e7931f94SStefano Zampini send_buffer_idxs should contain: 3567e7931f94SStefano Zampini - MatType_PRIVATE type 3568e7931f94SStefano Zampini - PetscInt size_of_l2gmap 3569e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 3570e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 3571e7931f94SStefano Zampini */ 3572e7931f94SStefano Zampini } else { 3573e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3574e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetSize(matis->mapping,&i);CHKERRQ(ierr); 3575854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 3576e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 3577e7931f94SStefano Zampini send_buffer_idxs[1] = i; 3578e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3579e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 3580e7931f94SStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 3581e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 3582e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3583e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 3584e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 3585c8587f34SStefano Zampini } 3586c8587f34SStefano Zampini } 3587e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 358828143c3dSStefano Zampini /* additional is (if any) */ 358928143c3dSStefano Zampini if (nis) { 359028143c3dSStefano Zampini PetscMPIInt psum; 359128143c3dSStefano Zampini PetscInt j; 359228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 359328143c3dSStefano Zampini PetscInt plen; 359428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 359528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 359628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 359728143c3dSStefano Zampini } 3598854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 359928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 360028143c3dSStefano Zampini PetscInt plen; 360128143c3dSStefano Zampini const PetscInt *is_array_idxs; 360228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 360328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 360428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 360528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 360628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 360728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 360828143c3dSStefano Zampini } 360928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 361028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 361128143c3dSStefano Zampini } 361228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 361328143c3dSStefano Zampini } 361428143c3dSStefano Zampini 3615e7931f94SStefano Zampini buf_size_idxs = 0; 3616e7931f94SStefano Zampini buf_size_vals = 0; 361728143c3dSStefano Zampini buf_size_idxs_is = 0; 3618e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3619e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 3620e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 362128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 3622e7931f94SStefano Zampini } 3623785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 3624785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 362595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 3626e7931f94SStefano Zampini 3627e7931f94SStefano Zampini /* get new tags for clean communications */ 3628e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 3629e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 363028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 3631e7931f94SStefano Zampini 3632e7931f94SStefano Zampini /* allocate for requests */ 3633785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 3634785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 363595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 3636785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 3637785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 363895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 3639e7931f94SStefano Zampini 3640e7931f94SStefano Zampini /* communications */ 3641e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3642e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 364328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 3644e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3645e7931f94SStefano Zampini source_dest = onodes[i]; 3646e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 3647e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 3648e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3649e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 365028143c3dSStefano Zampini if (nis) { 365128143c3dSStefano 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); 365228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 365328143c3dSStefano Zampini } 3654e7931f94SStefano Zampini } 3655e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 3656e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 3657e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 3658e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 365928143c3dSStefano Zampini if (nis) { 366028143c3dSStefano 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); 366128143c3dSStefano Zampini } 3662e7931f94SStefano Zampini } 3663e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 3664e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 3665e7931f94SStefano Zampini 3666e7931f94SStefano Zampini /* assemble new l2g map */ 3667e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3668e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36699d30be91SStefano Zampini new_local_rows = 0; 3670e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36719d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3672e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3673e7931f94SStefano Zampini } 36749d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 3675e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 36769d30be91SStefano Zampini new_local_rows = 0; 3677e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 36789d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 36799d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 3680e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3681e7931f94SStefano Zampini } 36829d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 36839d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 3684e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 3685e7931f94SStefano Zampini 3686e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 3687e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 3688e7931f94SStefano 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) */ 3689e7931f94SStefano Zampini if (n_recvs) { 369028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 3691e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 3692e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3693e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 3694e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 3695e7931f94SStefano Zampini break; 3696e7931f94SStefano Zampini } 3697e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3698e7931f94SStefano Zampini } 3699e7931f94SStefano Zampini switch (new_local_type_private) { 370028143c3dSStefano Zampini case MATDENSE_PRIVATE: 370128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 3702e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3703e7931f94SStefano Zampini bs = 1; 370428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 370528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 370628143c3dSStefano Zampini bs = 1; 370728143c3dSStefano Zampini } 3708e7931f94SStefano Zampini break; 3709e7931f94SStefano Zampini case MATAIJ_PRIVATE: 3710e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 3711e7931f94SStefano Zampini bs = 1; 3712e7931f94SStefano Zampini break; 3713e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 3714e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 3715e7931f94SStefano Zampini break; 3716e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 3717e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 3718e7931f94SStefano Zampini break; 3719e7931f94SStefano Zampini default: 37209d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 3721e7931f94SStefano Zampini break; 3722e7931f94SStefano Zampini } 372328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 372428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 372528143c3dSStefano Zampini bs = 1; 3726e7931f94SStefano Zampini } 3727e7931f94SStefano Zampini 372870cf5478SStefano Zampini /* create MATIS object if needed */ 372970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 3730e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 373128143c3dSStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,mat_n);CHKERRQ(ierr); 373270cf5478SStefano Zampini } else { 373370cf5478SStefano Zampini /* it also destroys the local matrices */ 373470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 373570cf5478SStefano Zampini } 373670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 3737e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 37389d30be91SStefano Zampini 37399d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 37409d30be91SStefano Zampini 37419d30be91SStefano Zampini /* Global to local map of received indices */ 37429d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 37439d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 37449d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 37459d30be91SStefano Zampini 37469d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 37479d30be91SStefano Zampini buf_size_idxs = 0; 37489d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37499d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 37509d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 37519d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 37529d30be91SStefano Zampini } 37539d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 37549d30be91SStefano Zampini 37559d30be91SStefano Zampini /* set preallocation */ 37569d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 37579d30be91SStefano Zampini if (!newisdense) { 37589d30be91SStefano Zampini PetscInt *new_local_nnz=0; 37599d30be91SStefano Zampini 37609d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 37619d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 37629d30be91SStefano Zampini if (n_recvs) { 37639d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 37649d30be91SStefano Zampini } 37659d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 37669d30be91SStefano Zampini PetscInt j; 37679d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 37689d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 37699d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 37709d30be91SStefano Zampini } 37719d30be91SStefano Zampini } else { 37729d30be91SStefano Zampini /* TODO */ 37739d30be91SStefano Zampini } 37749d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 37759d30be91SStefano Zampini } 37769d30be91SStefano Zampini if (new_local_nnz) { 37779d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 37789d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 37799d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 37809d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37819d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 37829d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 37839d30be91SStefano Zampini } else { 37849d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37859d30be91SStefano Zampini } 37869d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 37879d30be91SStefano Zampini } else { 37889d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 37899d30be91SStefano Zampini } 3790e7931f94SStefano Zampini 3791e7931f94SStefano Zampini /* set values */ 3792e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 37939d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 3794e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 3795e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 3796e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 37979d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 3798e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3799e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 3800e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 380128143c3dSStefano Zampini } else { 380228143c3dSStefano Zampini /* TODO */ 3803e7931f94SStefano Zampini } 3804e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 3805e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 3806e7931f94SStefano Zampini } 3807e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3808e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 380970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 381070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 38119d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 38129d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 3813e7931f94SStefano Zampini 3814dfd14d43SStefano Zampini #if 0 381528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 3816e7931f94SStefano Zampini Vec lvec,rvec; 3817e7931f94SStefano Zampini PetscReal infty_error; 3818e7931f94SStefano Zampini 38192a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 3820e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 3821e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 3822e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 382370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 3824e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 3825e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 3826e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 3827e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 3828e7931f94SStefano Zampini } 382928143c3dSStefano Zampini #endif 3830e7931f94SStefano Zampini 383128143c3dSStefano Zampini /* assemble new additional is (if any) */ 383228143c3dSStefano Zampini if (nis) { 383328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 383428143c3dSStefano Zampini 383528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3836854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 383728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 383828143c3dSStefano Zampini psum = 0; 383928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 384028143c3dSStefano Zampini for (j=0;j<nis;j++) { 384128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 384228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 384328143c3dSStefano Zampini psum += plen; 384428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 384528143c3dSStefano Zampini } 384628143c3dSStefano Zampini } 3847854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 3848854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 384928143c3dSStefano Zampini for (i=1;i<nis;i++) { 385028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 385128143c3dSStefano Zampini } 385228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 385328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 385428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 385528143c3dSStefano Zampini for (j=0;j<nis;j++) { 385628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 385728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 385828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 385928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 386028143c3dSStefano Zampini } 386128143c3dSStefano Zampini } 386228143c3dSStefano Zampini for (i=0;i<nis;i++) { 386328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 386428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 386528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 386628143c3dSStefano Zampini } 386728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 386828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 386928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 387028143c3dSStefano Zampini } 3871e7931f94SStefano Zampini /* free workspace */ 387228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 3873e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3874e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 3875e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 3876e7931f94SStefano Zampini if (isdense) { 3877e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 3878e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 3879e7931f94SStefano Zampini } else { 3880e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 3881e7931f94SStefano Zampini } 388228143c3dSStefano Zampini if (nis) { 388328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 388428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 388528143c3dSStefano Zampini } 3886e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 3887e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 388828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 3889e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 3890e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 389128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 3892e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 3893e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 3894e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 3895e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 3896e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 389728143c3dSStefano Zampini if (nis) { 389828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 389928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 390028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 390128143c3dSStefano Zampini } 390228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 390328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 390428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 390528143c3dSStefano Zampini for (i=0;i<nis;i++) { 390628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 390728143c3dSStefano Zampini } 390828143c3dSStefano Zampini } 3909e7931f94SStefano Zampini PetscFunctionReturn(0); 3910e7931f94SStefano Zampini } 3911a57a6d2fSStefano Zampini 391212edc857SStefano Zampini /* temporary hack into ksp private data structure */ 3913af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 391412edc857SStefano Zampini 3915c8587f34SStefano Zampini #undef __FUNCT__ 3916c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 3917c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 3918c8587f34SStefano Zampini { 3919c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 3920c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 392120a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 39229881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 392320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 39246e683305SStefano Zampini IS coarse_is,*isarray; 39256e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 39266e683305SStefano Zampini PetscInt nis,nisdofs,nisneu; 3927f9eb5b7dSStefano Zampini PC pc_temp; 3928c8587f34SStefano Zampini PCType coarse_pc_type; 3929c8587f34SStefano Zampini KSPType coarse_ksp_type; 3930f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 39314f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 39326e683305SStefano Zampini Mat t_coarse_mat_is; 39336e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 39346e683305SStefano Zampini PetscMPIInt all_procs; 393574e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 393668457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 393722bc73bbSStefano Zampini PetscScalar *array; 39389881197aSStefano Zampini PetscErrorCode ierr; 3939fdc09c96SStefano Zampini 3940c8587f34SStefano Zampini PetscFunctionBegin; 3941c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 394268457ee5SStefano 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 */ 3943fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 39445a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 3945fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 3946f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 3947f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 3948f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 3949fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 3950fa7f1dd8SStefano Zampini if (ocoarse_size != pcbddc->coarse_size) { /* ...but with different size, so reset it and set reuse flag to false */ 3951727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 3952fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3953fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 3954fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3955f4ddd8eeSStefano Zampini } 3956fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 3957fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 3958f4ddd8eeSStefano Zampini } 395970cf5478SStefano Zampini /* reset any subassembling information */ 396070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 39616e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 39626e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 3963fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 3964f4ddd8eeSStefano Zampini } 3965c8587f34SStefano Zampini 39666e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 39672b510759SStefano Zampini im_active = !!(pcis->n); 39682b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 39696e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 39706e683305SStefano Zampini void_procs = all_procs-active_procs; 39716e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 397274e2c79eSStefano Zampini redist = PETSC_FALSE; 397322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 39746e683305SStefano Zampini csin_ml = PETSC_TRUE; 39756e683305SStefano Zampini ncoarse_ml = void_procs; 3976779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 3977779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 397818a45a71SStefano Zampini csin_ds = PETSC_TRUE; 397918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 398018a45a71SStefano Zampini redist = PETSC_TRUE; 398118a45a71SStefano Zampini } else { 39826e683305SStefano Zampini csin_ds = PETSC_TRUE; 3983779c1cceSStefano Zampini ncoarse_ds = active_procs; 3984779c1cceSStefano Zampini redist = PETSC_TRUE; 398518a45a71SStefano Zampini } 39866e683305SStefano Zampini } else { 39876e683305SStefano Zampini csin_ml = PETSC_FALSE; 39886e683305SStefano Zampini ncoarse_ml = all_procs; 39896e683305SStefano Zampini if (void_procs) { 39906e683305SStefano Zampini csin_ds = PETSC_TRUE; 39916e683305SStefano Zampini ncoarse_ds = void_procs; 39926e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 39936e683305SStefano Zampini } else { 3994779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 399574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 399674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 399774e2c79eSStefano Zampini redist = PETSC_TRUE; 399874e2c79eSStefano Zampini } else { 39996e683305SStefano Zampini csin_ds = PETSC_FALSE; 40006e683305SStefano Zampini ncoarse_ds = all_procs; 40016e683305SStefano Zampini } 40026e683305SStefano Zampini } 400374e2c79eSStefano Zampini } 40046e683305SStefano Zampini 40056e683305SStefano Zampini /* 40066e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 40076e683305SStefano Zampini - we have not exceeded the number of levels requested 40086e683305SStefano Zampini - we can actually subassemble the active processes 40096e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 40106e683305SStefano Zampini */ 40116e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 40126e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 40136e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 40146e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 40156e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 4016f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 40172b510759SStefano Zampini } else { 4018f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 4019c8587f34SStefano Zampini } 4020c8587f34SStefano Zampini } 40216e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 40226e683305SStefano Zampini if (multilevel_allowed) { 40236e683305SStefano Zampini ncoarse = ncoarse_ml; 40246e683305SStefano Zampini csin = csin_ml; 402558da7f69SStefano Zampini redist = PETSC_FALSE; 40266e683305SStefano Zampini } else { 40276e683305SStefano Zampini ncoarse = ncoarse_ds; 40286e683305SStefano Zampini csin = csin_ds; 40296e683305SStefano Zampini } 4030e7931f94SStefano Zampini 4031abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 4032abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 4033abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 4034abbbba34SStefano Zampini 4035abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 403622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 403722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 403822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 403922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 4040b9b85e73SStefano Zampini #if 0 4041b9b85e73SStefano Zampini { 4042b9b85e73SStefano Zampini PetscViewer viewer; 4043b9b85e73SStefano Zampini char filename[256]; 4044b9b85e73SStefano Zampini sprintf(filename,"local_coarse_mat%d.m",PetscGlobalRank); 4045b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 4046b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4047b9b85e73SStefano Zampini ierr = MatView(coarse_submat_dense,viewer);CHKERRQ(ierr); 4048b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4049b9b85e73SStefano Zampini } 4050b9b85e73SStefano Zampini #endif 40516e683305SStefano 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); 40526e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 40536e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 40546e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4055abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 4056abbbba34SStefano Zampini 40576e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 40586e683305SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal) ) { /* protects from unneded computations */ 40596e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 40606e683305SStefano Zampini const PetscInt *idxs; 40616e683305SStefano Zampini ISLocalToGlobalMapping tmap; 40626e683305SStefano Zampini 40636e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 40640be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 40656e683305SStefano Zampini /* allocate space for temporary storage */ 4066854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 4067854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 40686e683305SStefano Zampini /* allocate for IS array */ 40696e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 40706e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 40716e683305SStefano Zampini nis = nisdofs + nisneu; 4072854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 40736e683305SStefano Zampini /* dofs splitting */ 40746e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 40756e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 40766e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 40776e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40786e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40796e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 40806e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40816e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->ISForDofsLocal[i]),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 40826e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 40836e683305SStefano Zampini } 40846e683305SStefano Zampini /* neumann boundaries */ 40856e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 40866e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 40876e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 40886e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 40896e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 40906e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 40916e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 40926e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pcbddc->NeumannBoundariesLocal),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 40936e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 40946e683305SStefano Zampini } 40956e683305SStefano Zampini /* free memory */ 40966e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 40976e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 40986e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 40996e683305SStefano Zampini } else { 41006e683305SStefano Zampini nis = 0; 41016e683305SStefano Zampini nisdofs = 0; 41026e683305SStefano Zampini nisneu = 0; 41036e683305SStefano Zampini isarray = NULL; 41046e683305SStefano Zampini } 41056e683305SStefano Zampini /* destroy no longer needed map */ 41066e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 41076e683305SStefano Zampini 41086e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 41096e683305SStefano Zampini coarse_mat_is = NULL; 41106e683305SStefano Zampini if (csin) { 41116e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 411274e2c79eSStefano Zampini if (redist) { 411374e2c79eSStefano Zampini PetscMPIInt rank; 4114779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 411574e2c79eSStefano Zampini 411674e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 411758da7f69SStefano Zampini spc = active_procs/ncoarse; 411858da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 4119779c1cceSStefano Zampini if (im_active) { 4120779c1cceSStefano Zampini destsize = 1; 412174e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 412274e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 412374e2c79eSStefano Zampini } else { 412474e2c79eSStefano Zampini dest[0] = rank/(spc+1); 412574e2c79eSStefano Zampini } 412674e2c79eSStefano Zampini } else { 4127779c1cceSStefano Zampini destsize = 0; 41286e683305SStefano Zampini } 4129779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4130779c1cceSStefano Zampini } else if (csin_type_simple) { 41316e683305SStefano Zampini PetscMPIInt rank; 41326e683305SStefano Zampini PetscInt issize,isidx; 4133779c1cceSStefano Zampini 41346e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 41356e683305SStefano Zampini if (im_active) { 41366e683305SStefano Zampini issize = 1; 41376e683305SStefano Zampini isidx = (PetscInt)rank; 41386e683305SStefano Zampini } else { 41396e683305SStefano Zampini issize = 0; 41406e683305SStefano Zampini isidx = -1; 41416e683305SStefano Zampini } 41426e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 4143779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 41446e683305SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,PETSC_TRUE,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 41456e683305SStefano Zampini } 4146779c1cceSStefano Zampini 4147779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 4148779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 4149779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 4150779c1cceSStefano Zampini PetscInt *coarse_candidates; 4151779c1cceSStefano Zampini const PetscInt* tisindices; 4152779c1cceSStefano Zampini 4153779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 4154779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 4155779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4156779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 4157779c1cceSStefano Zampini if (!coarse_candidates[i]) { 4158779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 4159779c1cceSStefano Zampini } 4160779c1cceSStefano Zampini } 4161779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 4162779c1cceSStefano Zampini 4163779c1cceSStefano Zampini 41646e683305SStefano Zampini if (pcbddc->dbg_flag) { 41656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41666e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 41676e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 4169779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 41706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 41716e683305SStefano Zampini } 41726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 41736e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41746e683305SStefano Zampini } 41756e683305SStefano Zampini /* shift the pattern on coarse candidates */ 41766e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 41776e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 4178854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 41796e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 41806e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 41816e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 41826e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 41836e683305SStefano Zampini } 41846e683305SStefano Zampini if (pcbddc->dbg_flag) { 41856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 41876e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 41886e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41896e683305SStefano Zampini } 4190779c1cceSStefano Zampini } 41916e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 41926e683305SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 41936e683305SStefano Zampini } else { 41946e683305SStefano Zampini if (pcbddc->dbg_flag) { 41956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 41966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 41976e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 41986e683305SStefano Zampini } 41996e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 42006e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 42016e683305SStefano Zampini } 42026e683305SStefano Zampini 42036e683305SStefano Zampini /* create local to global scatters for coarse problem */ 420468457ee5SStefano Zampini if (compute_vecs) { 42056e683305SStefano Zampini PetscInt lrows; 42066e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 42076e683305SStefano Zampini if (coarse_mat_is) { 42086e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 42096e683305SStefano Zampini } else { 42106e683305SStefano Zampini lrows = 0; 42116e683305SStefano Zampini } 42126e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 42136e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 42146e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 42156e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42166e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 42176e683305SStefano Zampini } 42186e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 42196e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 4220c8587f34SStefano Zampini 4221f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 4222f9eb5b7dSStefano Zampini if (multilevel_allowed) { 4223f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 4224f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 4225f9eb5b7dSStefano Zampini } else { 4226f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 4227f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 4228c8587f34SStefano Zampini } 4229c8587f34SStefano Zampini 42306e683305SStefano Zampini /* print some info if requested */ 42316e683305SStefano Zampini if (pcbddc->dbg_flag) { 42326e683305SStefano Zampini if (!multilevel_allowed) { 42336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 42346e683305SStefano Zampini if (multilevel_requested) { 42356e683305SStefano 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); 42366e683305SStefano Zampini } else if (pcbddc->max_levels) { 42376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 42386e683305SStefano Zampini } 42396e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 42406e683305SStefano Zampini } 42416e683305SStefano Zampini } 42426e683305SStefano Zampini 4243f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 42446e683305SStefano Zampini if (coarse_mat_is) { 42456e683305SStefano Zampini MatReuse coarse_mat_reuse; 42466a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 42476e683305SStefano Zampini if (pcbddc->dbg_flag) { 42486e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 42496e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 42506e683305SStefano Zampini } 4251f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 4252312be037SStefano Zampini char prefix[256],str_level[16]; 4253e604994aSStefano Zampini size_t len; 42546e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 4255c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 4256f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 42575f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 4258c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 42596e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 4260c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4261c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 42623e3c6dadSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 4263e604994aSStefano Zampini /* prefix */ 4264e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 4265e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 4266e604994aSStefano Zampini if (!pcbddc->current_level) { 4267e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 4268e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 4269c8587f34SStefano Zampini } else { 4270e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 4271312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 4272312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 427334d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 4274312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 4275e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 4276e604994aSStefano Zampini } 4277e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 42783e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42793e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 42803e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 42813e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 4282f9eb5b7dSStefano Zampini /* allow user customization */ 4283f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 42843e3c6dadSStefano Zampini } 42853e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 42863e3c6dadSStefano Zampini if (nisdofs) { 42873e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 42883e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 42893e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 42903e3c6dadSStefano Zampini } 42913e3c6dadSStefano Zampini } 42923e3c6dadSStefano Zampini if (nisneu) { 42933e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 42943e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 4295312be037SStefano Zampini } 4296f9eb5b7dSStefano Zampini 4297f9eb5b7dSStefano Zampini /* get some info after set from options */ 4298f9eb5b7dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 4299f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 4300f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 43014f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 43026e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 4303f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 4304f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 4305f9eb5b7dSStefano Zampini } 43064f3a063dSStefano Zampini if (isredundant) { 43074f3a063dSStefano Zampini KSP inner_ksp; 43084f3a063dSStefano Zampini PC inner_pc; 43094f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 43104f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 43114f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 43124f3a063dSStefano Zampini } 4313f9eb5b7dSStefano Zampini 4314f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 4315fa7f1dd8SStefano Zampini if (coarse_reuse) { 431681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 4317fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 43186e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 4319fa7f1dd8SStefano Zampini } else { 43206e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 4321fa7f1dd8SStefano Zampini } 4322c8587f34SStefano Zampini if (isbddc || isnn) { 432322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 432470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 432528143c3dSStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,PETSC_TRUE,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 432622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 43276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 43286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 43296e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 43306e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 433122b6e8a2SStefano Zampini } 433270cf5478SStefano Zampini } 43336e683305SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 433470cf5478SStefano Zampini } else { 433522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 433622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 433722bc73bbSStefano Zampini } 433822bc73bbSStefano Zampini } else { 43392e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 4340c8587f34SStefano Zampini } 4341c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 4342c8587f34SStefano Zampini 4343c8587f34SStefano Zampini /* propagate symmetry info to coarse matrix */ 4344b9d89cd5SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pcbddc->issym);CHKERRQ(ierr); 43455a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 4346c8587f34SStefano Zampini 43476e683305SStefano Zampini /* set operators */ 43485f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 43496e683305SStefano Zampini if (pcbddc->dbg_flag) { 43506e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 43516e683305SStefano Zampini } 43526e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 43536e683305SStefano Zampini coarse_mat = 0; 43546e683305SStefano Zampini } 43556e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 4356b9b85e73SStefano Zampini #if 0 4357b9b85e73SStefano Zampini { 4358b9b85e73SStefano Zampini PetscViewer viewer; 4359b9b85e73SStefano Zampini char filename[256]; 4360b9b85e73SStefano Zampini sprintf(filename,"coarse_mat.m"); 4361b9b85e73SStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,filename,&viewer);CHKERRQ(ierr); 4362b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4363b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 4364b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 4365b9b85e73SStefano Zampini } 4366b9b85e73SStefano Zampini #endif 4367c8587f34SStefano Zampini 4368c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 4369c8587f34SStefano Zampini if (pcbddc->NullSpace) { 4370c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 437198a51de6SStefano Zampini } 437298a51de6SStefano Zampini 437398a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 437498a51de6SStefano Zampini Vec crhs,csol; 437598a51de6SStefano Zampini PetscBool ispreonly; 437698a51de6SStefano Zampini if (CoarseNullSpace) { 4377c8587f34SStefano Zampini if (isbddc) { 4378c8587f34SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 4379c8587f34SStefano Zampini } else { 4380c8587f34SStefano Zampini ierr = KSPSetNullSpace(pcbddc->coarse_ksp,CoarseNullSpace);CHKERRQ(ierr); 4381c8587f34SStefano Zampini } 4382c8587f34SStefano Zampini } 4383f9eb5b7dSStefano Zampini /* setup coarse ksp */ 4384f9eb5b7dSStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 4385f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 4386f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 43876e683305SStefano Zampini /* hack */ 4388f347579bSStefano Zampini if (!csol) { 43892a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 4390f9eb5b7dSStefano Zampini } 4391f347579bSStefano Zampini if (!crhs) { 43922a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 4393f347579bSStefano Zampini } 4394cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 4395cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 43966e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 4397c8587f34SStefano Zampini KSP check_ksp; 43982b510759SStefano Zampini KSPType check_ksp_type; 4399c8587f34SStefano Zampini PC check_pc; 44006e683305SStefano Zampini Vec check_vec,coarse_vec; 44016a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 44022b510759SStefano Zampini PetscInt its; 44036e683305SStefano Zampini PetscBool compute_eigs; 44046e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 44056e683305SStefano Zampini PetscInt neigs; 44068e185a42SStefano Zampini const char *prefix; 4407c8587f34SStefano Zampini 44082b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 44096e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 441023ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 4411f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 44122b510759SStefano Zampini if (ispreonly) { 44132b510759SStefano Zampini check_ksp_type = KSPPREONLY; 44146e683305SStefano Zampini compute_eigs = PETSC_FALSE; 44152b510759SStefano Zampini } else { 4416cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 44176e683305SStefano Zampini compute_eigs = PETSC_TRUE; 4418c8587f34SStefano Zampini } 4419c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 44206e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 44216e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 44226e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 4423a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 4424a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 4425a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 4426a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 4427c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 4428c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 4429c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 4430c8587f34SStefano Zampini /* create random vec */ 44316e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 44326e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 4433c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 4434c8587f34SStefano Zampini if (CoarseNullSpace) { 4435c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 4436c8587f34SStefano Zampini } 44376e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 4438c8587f34SStefano Zampini /* solve coarse problem */ 44396e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 4440c8587f34SStefano Zampini if (CoarseNullSpace) { 44416e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 4442c8587f34SStefano Zampini } 4443cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 44446e683305SStefano Zampini if (compute_eigs) { 4445854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 4446854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 44476e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 44486e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 44496e683305SStefano Zampini lambda_min = eigs_r[0]; 44506e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 44516e683305SStefano Zampini if (lambda_max>lambda_min) { 4452cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 4453cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 4454cbcc2c2aSStefano Zampini } 4455c8587f34SStefano Zampini } 4456c8587f34SStefano Zampini } 4457cbcc2c2aSStefano Zampini 4458c8587f34SStefano Zampini /* check coarse problem residual error */ 44596e683305SStefano Zampini if (pcbddc->dbg_flag) { 44606e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 44616e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44626e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 4463c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 44646e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 44656e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 4466c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 4467779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 44686e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 44696e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 44706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 44716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 44726e683305SStefano Zampini if (compute_eigs) { 44736e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 4474deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 4475c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 44766e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 44776e683305SStefano 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); 44786e683305SStefano Zampini for (i=0;i<neigs;i++) { 44796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 4480c8587f34SStefano Zampini } 44816e683305SStefano Zampini } 44826e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 44836e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 44846e683305SStefano Zampini } 4485c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 44866e683305SStefano Zampini if (compute_eigs) { 44876e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 44886e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 4489c8587f34SStefano Zampini } 44906e683305SStefano Zampini } 44916e683305SStefano Zampini } 4492cbcc2c2aSStefano Zampini /* print additional info */ 4493cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 44946e683305SStefano Zampini /* waits until all processes reaches this point */ 44956e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 4496cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 4497cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4498cbcc2c2aSStefano Zampini } 4499cbcc2c2aSStefano Zampini 45002b510759SStefano Zampini /* free memory */ 4501c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 4502fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 4503c8587f34SStefano Zampini PetscFunctionReturn(0); 4504c8587f34SStefano Zampini } 4505674ae819SStefano Zampini 4506f34684f1SStefano Zampini #undef __FUNCT__ 4507f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 4508f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 4509f34684f1SStefano Zampini { 4510f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4511f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 4512f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 45130e6343abSStefano Zampini PetscInt i,local_size,coarse_size=0; 451473be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 45150e6343abSStefano Zampini PetscInt *t_local_primal_indices=NULL; 4516f34684f1SStefano Zampini PetscErrorCode ierr; 4517f34684f1SStefano Zampini 4518f34684f1SStefano Zampini PetscFunctionBegin; 4519f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 45200e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 45210e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 4522727cdba6SStefano Zampini } 45230e6343abSStefano Zampini ierr = PCBDDCSubsetNumbering(PetscObjectComm((PetscObject)(pc->pmat)),matis->mapping,pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult,&coarse_size,&t_local_primal_indices);CHKERRQ(ierr); 45240e6343abSStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&local_primal_indices);CHKERRQ(ierr); 45250e6343abSStefano Zampini local_size = 0; 45260e6343abSStefano Zampini for (i=0;i<pcbddc->local_primal_size_cc;i++) { 45270e6343abSStefano Zampini PetscInt j; 45280e6343abSStefano Zampini for (j=0;j<pcbddc->local_primal_ref_mult[i];j++) local_primal_indices[local_size++] = t_local_primal_indices[i] + j; 45290e6343abSStefano Zampini } 45300e6343abSStefano Zampini ierr = PetscFree(t_local_primal_indices);CHKERRQ(ierr); 45310e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 45320e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 45330e6343abSStefano Zampini } 4534f34684f1SStefano Zampini 4535f34684f1SStefano Zampini /* check numbering */ 4536f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 4537f34684f1SStefano Zampini PetscScalar coarsesum,*array; 4538b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 4539f34684f1SStefano Zampini 4540f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4541f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 4542f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 45430fccc4e9SStefano Zampini ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr); 4544f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 4545f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 4546727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 4547f34684f1SStefano Zampini } 4548f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4549f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4550f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4551f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4552f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4553f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4554f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4555f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4556f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4557f34684f1SStefano Zampini if (array[i] == 1.0) { 4558b9b85e73SStefano Zampini set_error = PETSC_TRUE; 4559f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a single process!\n",PetscGlobalRank,i);CHKERRQ(ierr); 4560f34684f1SStefano Zampini } 4561f34684f1SStefano Zampini } 4562b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4563f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4564f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 4565f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 4566f34684f1SStefano Zampini } 4567f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 4568f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4569f34684f1SStefano Zampini ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4570f34684f1SStefano Zampini ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4571f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 4572f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 4573b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 4574ca8b9ea9SStefano Zampini PetscInt *gidxs; 4575ca8b9ea9SStefano Zampini 4576ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 4577ca8b9ea9SStefano Zampini ierr = ISLocalToGlobalMappingApply(matis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 4578f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 4579f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4580f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4581f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 45824bc2dc4bSStefano 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); 4583f34684f1SStefano Zampini } 4584f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4585ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 4586f34684f1SStefano Zampini } 4587f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4588302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 4589f34684f1SStefano Zampini } 45908bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 4591f34684f1SStefano Zampini /* get back data */ 4592f34684f1SStefano Zampini *coarse_size_n = coarse_size; 4593f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 4594674ae819SStefano Zampini PetscFunctionReturn(0); 4595674ae819SStefano Zampini } 4596674ae819SStefano Zampini 4597e456f2a8SStefano Zampini #undef __FUNCT__ 4598e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 4599a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 4600e456f2a8SStefano Zampini { 4601e456f2a8SStefano Zampini IS localis_t; 4602a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 4603e456f2a8SStefano Zampini PetscScalar *vals; 4604e456f2a8SStefano Zampini PetscErrorCode ierr; 4605e456f2a8SStefano Zampini 4606e456f2a8SStefano Zampini PetscFunctionBegin; 4607a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 4608e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 4609854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 4610e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 4611e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4612a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 4613a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 46141035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 4615a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 46161035eff8SStefano Zampini } 4617a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 4618e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 4619e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4620a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 4621a7dc3881SStefano Zampini /* now compute set in local ordering */ 4622a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4623a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4624a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4625a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 4626a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4627ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4628e456f2a8SStefano Zampini lsize++; 4629e456f2a8SStefano Zampini } 4630e456f2a8SStefano Zampini } 4631854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 4632a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 4633ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 4634e456f2a8SStefano Zampini idxs[lsize++] = i; 4635e456f2a8SStefano Zampini } 4636e456f2a8SStefano Zampini } 4637a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 4638a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 4639e456f2a8SStefano Zampini *localis = localis_t; 4640e456f2a8SStefano Zampini PetscFunctionReturn(0); 4641e456f2a8SStefano Zampini } 4642906d46d4SStefano Zampini 4643906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 4644906d46d4SStefano Zampini #undef __FUNCT__ 4645906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 4646906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 4647906d46d4SStefano Zampini { 4648906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4649906d46d4SStefano Zampini PetscErrorCode ierr; 4650906d46d4SStefano Zampini 4651906d46d4SStefano Zampini PetscFunctionBegin; 4652906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4653906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4654906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4655906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4656906d46d4SStefano Zampini PetscFunctionReturn(0); 4657906d46d4SStefano Zampini } 4658906d46d4SStefano Zampini 4659906d46d4SStefano Zampini #undef __FUNCT__ 4660906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 4661906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 4662906d46d4SStefano Zampini { 4663906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4664906d46d4SStefano Zampini PetscErrorCode ierr; 4665906d46d4SStefano Zampini 4666906d46d4SStefano Zampini PetscFunctionBegin; 4667906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 4668906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 4669906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 4670906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 4671906d46d4SStefano Zampini PetscFunctionReturn(0); 4672906d46d4SStefano Zampini } 4673b96c3477SStefano Zampini 4674b96c3477SStefano Zampini #undef __FUNCT__ 4675b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 467608122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 4677b96c3477SStefano Zampini { 4678a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4679b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4680b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4681a64f4aa4SStefano Zampini Mat S_j; 4682b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 4683b96c3477SStefano Zampini PetscBool free_used_adj; 4684b96c3477SStefano Zampini PetscErrorCode ierr; 4685b96c3477SStefano Zampini 4686b96c3477SStefano Zampini PetscFunctionBegin; 4687b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 4688b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 468908122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 4690b96c3477SStefano Zampini used_xadj = NULL; 4691b96c3477SStefano Zampini used_adjncy = NULL; 4692b96c3477SStefano Zampini } else { 469308122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 469408122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 469508122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 469608122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 4697b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 4698b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 4699b96c3477SStefano Zampini } else { 47002fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 4701b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 4702b96c3477SStefano Zampini PetscInt nvtxs; 4703b96c3477SStefano Zampini 47042fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 47052fffb893SStefano Zampini if (flg_row) { 4706b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 4707b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 4708b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 4709b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 47102fffb893SStefano Zampini } else { 47112fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 47122fffb893SStefano Zampini used_xadj = NULL; 47132fffb893SStefano Zampini used_adjncy = NULL; 47142fffb893SStefano Zampini } 47152fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 4716b96c3477SStefano Zampini } 4717b96c3477SStefano Zampini } 4718d5574798SStefano Zampini 4719d5574798SStefano Zampini /* setup sub_schurs data */ 4720a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 4721a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { 4722a64f4aa4SStefano Zampini /* pcbddc->ksp_D up to date only if not using MUMPS */ 4723a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 47246816873aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4725a64f4aa4SStefano Zampini } else { 47266816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 47276816873aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->use_edges,pcbddc->use_faces);CHKERRQ(ierr); 4728a64f4aa4SStefano Zampini } 4729a64f4aa4SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 4730b96c3477SStefano Zampini 4731b96c3477SStefano Zampini /* free adjacency */ 4732b96c3477SStefano Zampini if (free_used_adj) { 4733b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 4734b96c3477SStefano Zampini } 4735b96c3477SStefano Zampini PetscFunctionReturn(0); 4736b96c3477SStefano Zampini } 4737b96c3477SStefano Zampini 4738b96c3477SStefano Zampini #undef __FUNCT__ 4739b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 474008122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 4741b96c3477SStefano Zampini { 4742b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 4743b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 4744b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4745b96c3477SStefano Zampini PCBDDCGraph graph; 4746b96c3477SStefano Zampini PetscErrorCode ierr; 4747b96c3477SStefano Zampini 4748b96c3477SStefano Zampini PetscFunctionBegin; 4749b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 475008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 4751b96c3477SStefano Zampini IS verticesIS; 4752b96c3477SStefano Zampini 4753b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 4754b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 47557fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 4756b96c3477SStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticesIS);CHKERRQ(ierr); 4757b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4758b96c3477SStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 4759b96c3477SStefano Zampini /* 4760b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 4761b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 4762b96c3477SStefano Zampini } 4763b96c3477SStefano Zampini */ 4764b96c3477SStefano Zampini } else { 4765b96c3477SStefano Zampini graph = pcbddc->mat_graph; 4766b96c3477SStefano Zampini } 4767b96c3477SStefano Zampini 4768b96c3477SStefano Zampini /* sub_schurs init */ 4769a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 4770a64f4aa4SStefano Zampini 4771b96c3477SStefano Zampini /* free graph struct */ 477208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 4773b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4774b96c3477SStefano Zampini } 4775b96c3477SStefano Zampini PetscFunctionReturn(0); 4776b96c3477SStefano Zampini } 4777